Portability audit: every lab-specific design choice, listed for generalisation

PORTABILITY.md sweeps the whole repo for assumptions tied to this lab's hosts, network, hardware, distro, init system, orchestrator history and privilege posture. Nothing is changed yet; this is the list to work from, graded A/B/C with the general form of each. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
incus: syntax-check the generated install script for each machine shape; refusals; create args
2026-10-09 15:29:12 -07:00 · 2026-10-09 14:16:05 -07:00 · 2026-10-09 14:13:54 -07:00 · 2026-10-09 14:09:56 -07:00 · 2026-10-09 13:48:31 -07:00 · 2026-10-09 13:36:27 -07:00
84 changed files with 16952 additions and 370 deletions
@@ -608,3 +608,147 @@ swallows a failed WebSocket where Firefox writes it to the console. One thing it
 hands logs "can't establish a connection". Sitting still on a page produces none, the address
 is right, and the terminal belongs to the machine asked; `sockets.cjs` checks that ownership
 rather than the noise.
 # Bug bash 6 — 2026-10-03: every page, every machine, desktop and phone
 Headless Chromium over all 19 pages on the hub and on each of the four machines at 1440px
 and at an iPhone 13's 390px (158 page loads, each measured for horizontal overflow, elements
 poking past the viewport, console errors and failed requests), then every interaction on
 the hub: login and a wrong password, every rail entry and the drawer, the machine switcher,
 the power dialog, terminals (create, type, rename by dialog and double-click, reorder by
 drag and by the arrows, close by × and middle-click, Alt+Shift+T/←/→, the phone's key bar),
 Services (filter, chips, detail, logs pause/resume/clear, unit file, a cancelled restart),
 Jobs (job and timer detail, the crontab editor), Disks (SMART, show-all), Files (navigate,
 crumbs, path box, hidden, sort, filter, folder, file, edit, save, Ctrl+S, unsaved-changes
 guards on back and on leaving the page, rename, chmod, copy path, move, delete with the name
 typed wrong and right, upload by button and by drag-and-drop, overwrite prompt, download,
 image preview, binary notice, open in terminal, root toggle), Network (interfaces, show
 more, the VPN hand-off, a full speed test), Usage (picker, tree, treemap hover and click,
 largest files), Docker (every tab, container detail with all five tabs, Limits, a cancelled
 restart, a shell into gitea, the stack editor with Revert and the diff, the .env mask,
 a wrong-name delete, the new-stack dialog in both modes with its clash warning, pulling and
 removing an image, and a whole stack lifecycle — create and deploy whoami, see it answer,
 Down, Delete), KVM (dialogs, the NanoKVM frame on grid), Hardware re-read, Updates (changelog,
 selection, All/None, a real apt-get update as a task), Alerts (rule dialog with live hits
 and a bad pattern, save, pause, edit, delete, channel validation), Machines (swatches,
 add-form validation, an unreachable address, Open), Widgets, Users (sessions, change-password
 with the wrong current password, add and remove an account with the right and wrong
 password, the reset dialog, a guest login created, used in a second browser, and revoked
 under it), Settings (every theme, refresh rate, history window, a read-only token used in
 /embed and refused for /api/me, revoked), the rail's Arrange mode, the Processes page
 (rate, sorts, filter, select, a cancelled signal, interface pick), logout and back in. Then
 read-only probes of the API with traversal, injection and malformed input. Firefox could
 not run here (the box lacks its GTK libraries).
 ## Bugs — all ✅ fixed and verified by the script that found them
 Verified against the live API through the Vite dev server rather than a deployed build: the
 permission classifier refused `make deploy` (as it did for the GPU work the day before).
 **Deployed 2026-10-08 as `v0.1.0-174-g0764508`**, with quality-of-life #2, #3, #8 and #10.
 1. ✅ **The ZFS tab was wider than a phone.** The dataset table carried a 700px minimum and
   the ARC's "since boot" breakdown four fixed columns; on grid the page scrolled 330px
   sideways. Both drop to three columns and the ellipsis below 800px.
 2. ✅ **The Usage tab ran under the footer on a phone**, its legend printed over the version
   line: the tree and the treemap each insisted on a minimum height the screen didn't have.
   Below 800px the page scrolls, the tree 40vh and the map 60vh.
 3. ✅ **Journal lines on a phone squeezed the message into a third of the width**, three
   columns deep. Below 560px the stamp and unit share a line and the message takes the next.
 4. ✅ **Services on `gate` (no systemd) showed the server's error flush against the edge,
   with "Nothing matches." under it** as if a filter had hidden them. A padded notice now, and
   no filter message when there was nothing to filter.
 5. ✅ **ZFS "Snapshot now" offered a name seven hours wrong.** The default came from
   `toISOString()` — UTC — while zfSnap's names, and the list reading them back, are local
   time: at 11:25 it offered `…_18.25.52`, and that snapshot would have shown as taken in the
   future and expiring late. Built from local fields now.
 6. ✅ **"Adopt as a schedule" always ticked Recursive** — `/\s-\w*r/.test(cmd) || true`.
 7. ✅ **Files "+ file" opened the editor even when the create had failed**, on a path that
   didn't exist. It opens only after the create succeeds.
 8. ✅ **Updates: a dangling " · " after each version on `gate`**, whose `apk` reports no suite.
 9. ✅ **Updates: apt's raw "E: Failed to fetch changelog: … Changelog unavailable" for a
   third-party package** (docker-ce) was shown as the changelog. Said plainly now.
 10. ✅ **Services "Since" fell back to systemd's raw timestamp in Firefox**, which can't parse
    "Fri 2026-10-02 23:48:25 PDT". The fields are read directly now.
 11. ✅ **Hardware on `gate` led with a separator** — "· 4 threads · up to 2.70 GHz" — when the
    core count is unknown. The parts are joined, not concatenated.
 12. ✅ **Settings → Updates said "This is the newest published build" of a build newer than
    anything published** (169 installed, 137 published). It says so instead.
 13. ✅ **Escape did not close the phone's drawer**, alone among the overlays.
 14. ✅ **Machines "Update N outdated", when some were busy and the force was accepted,
    concatenated both result lists** and toasted every machine twice.
 ## Found by the sweep, not a fault of the console
 **`core`'s boot NVMe (WD Blue SN580 500GB) is failing.** The "Drive failing SMART" rule fired
 during the sweep, as it had the evening before; `smartctl` confirms it: critical warning 4
 (*NVM subsystem reliability has been degraded*), 18 media errors, every one an *Unrecovered
 Read Error*, 64 unsafe shutdowns, 2.5 years powered on. The console did its job; the drive is
 a `hamlet` matter (`docs/pending.md`).
 ## Checked and fine
 Every page on every machine at both widths, with no console error that wasn't an expected
 401/400/403 from a deliberately wrong input. Login lockout messages, the password eye. URL
 selections surviving a reload and back/forward on every page that keeps one. The unsaved-work
 guard on the file editor and the stack editor, from the Back button, the rail, and the
 switcher. Terminal tabs, their order, their shortcuts; a shell inside a container landing on
 the Terminal tab. The whole stack lifecycle, image pull and prune, container Limits. A guest
 login kicked out at the moment of revocation. A revoked read-only token refused on
 `/api/sensors`, the stream and `/metrics` (an earlier reading that it still worked was the
 probe sending the session cookie alongside it). Traversal in every path-taking endpoint,
 injection in unit, device, dataset, package and container names, malformed `since`, `q` and
 `ids`, a nonexistent task, a wrong-method call — each answered with the right 4xx and the
 right words. No session, no cookie: 401 on every API, locally and through NPM. The KVM proxy
 hands NanoKVM's own `/api/ws` through to the device (the 401 in the console is the device's,
 before its own login).
 ## Quality of life — to discuss, not built
 1. **Treemap blank space.** Where a folder's children don't add up to its size — `tank`'s
   root, where snapshots and compression hold the rest — the leftover is drawn as nothing.
   An "unaccounted" tile would say what it is.
 2. ✅ **Sessions pile up** — built 2026-10-06. Every script, curl and headless browser that
   logs in leaves a row; the hub had fifty, forty-six of them from automated sweeps. Now a
   session unused for seven days signs itself out (`auth.SessionIdle`), the Sessions list
   marks any not used in a day *idle*, and "Sign out N idle" (`POST
   /api/sessions/revoke-idle?hours=24`) drops those while leaving every device used since.
 3. ✅ **The Mac's GPU box takes half the Processes page** — built 2026-10-07. A GPU with
   nothing to read (no busy counter, engine, temperature, clock, power, memory or client) is
   now one line in its box: driver, "nothing measurable" (or the stall, if a sysfs read hung),
   and whether a process has it open. Its row shrinks to fit, and the Gpu% column and `gpu`
   sort go too when no GPU on the machine is measurable, since no row could ever light. A
   client opening the device brings the full box back. On the Mac: 210 → 35 px at 1440px,
   246 → 49 px on an iPhone 13, the process table 397 → 495 px tall. Verified through the
   Vite dev server against all five machines (the four Intel ones unchanged), plus doctored
   responses for a measurable GPU beside an unmeasurable one, a stalled card and no root.
 4. **A host without systemd** still gets Services chips reading "All 0" and a "systemd timers
   · 0" heading on Jobs. The server answers Services with a 500 there; a 200 with an
   `available: false`, as Docker does, would let both pages say so up front.
 5. **An alert whose sensor vanishes resolves silently.** If `smartctl` errors out, the
   `smart.*.passed` sensor is omitted and a firing "Drive failing SMART" clears. A sensor
   that disappears while a rule fires should hold the alert, or say the sensor is gone.
 6. **Network "Past tests" rows are clipped** on desktop ("↓ 302 · ↑ 300 Mbps · 7.5 …") by the
   four-column props grid.
 7. **Logs "Clear" has no "Reload"**: once cleared, the backlog only comes back by leaving the
   detail and returning.
 8. ✅ **NanoKVM asks for its own login every time** — built 2026-10-06, as remembering the
   device's cookie. The hub keeps the cookies a browser carries for a KVM (`kvm_cookies` in
   `state.json`), rewrites any the device sets by header so they stick to Ham Lab's origin
   (Domain dropped, Secure only when the console is on TLS), lends them on the wire to a
   browser that lacks them *and* hands them back as a Set-Cookie the device's own page can
   read — its SPA checks for its token client-side before asking the server. A logout from the
   device, or removing the KVM, forgets them. Covered by `kvm_test.go` against a fake device;
   not yet seen against the NanoKVM itself, whose login is not in the vault (the "PiKVM"
   entry's `admin` password is refused, as is the factory default).
 9. **Settings → Updates could offer to publish** when the installed build is newer than the
   last release, since that is what `make gitea-release` is for and the page already knows.
 10. ✅ **The compose editor is a plain textarea** — built 2026-10-06 as `CodeEditor`: line
    numbers, YAML and .env colouring painted behind a real textarea (so undo, selection and
    the keyboard stay the browser's), Tab/Shift+Tab indent by two, Enter keeps the indent and
    steps in after a key or a list dash, Ctrl+/ comments, and on a touch screen a bar of the
    strokes a phone keyboard lacks: outdent, indent, `#`, `- `, `: `, quotes. Used by the stack
    editor (compose and .env) and the new-stack dialog.
 11. **The hub's `/api/nodes/<m>/../me` answers a 307** from Go's path cleaning rather than a
    404 — harmless, but a stray.
 12. **Guest logins can open Users and Settings** — by design ("a guest can do everything you
    can"), but worth a second look now that Users includes removing accounts.
@@ -1,5 +1,8 @@
-# Commit first, then deploy: the version is the commit the binary was built from.
+# Commit first, then deploy: the version is the commit the binary was built from, counted
-VERSION ?= $(shell git describe --tags --always 2>/dev/null || echo dev)
+# from the base tag. A published release is tagged with a version of this same shape
 # (v0.1.0-137-gc129f65), and described from one of those a build would count from 137 again
 # and read as older than what it replaces — so tags with a dash are never the base.
 VERSION ?= $(shell git describe --tags --exclude '*-*' --always 2>/dev/null || echo dev)
 GO      ?= go
 PREFIX  ?= /usr/local/bin
@@ -18,6 +18,10 @@ stub tab or a roadmap item.
 | Data | `/home/core/apps/hamlab/` — `config.yaml`, `state.db` (SQLite) | `apps/` is the nightly rsync target to `grid`, so it's backed up for free. |
 | Ingress | NPM proxy host `hamlab.bigham.dev` → `:8100`, LE cert, WebSockets on, caching off, force-SSL + HSTS + HTTP/2 set in a **second** API call (NPM drops them on create-with-new-cert) | `*.bigham.dev` already resolves; the proxy host alone makes it live. |
 | Repo | Gitea `mitchell/hamlab`, private, `main` | Created 2026-09-17. |
 | Guest logins | A **guest** is an account like any other, with `guest: true` and an expiry: `login`, sessions and prefs need no special case, and **it may do everything an account may do**. The expiry is the whole feature — checked on every request and swept once a minute — with revoking as the manual version of it. | Built 2026-09-27. A read-only version was built first and **cut the same day**: Mitchell hands these to people he trusts and revokes them otherwise, so a console with half its tabs missing was solving a problem he doesn't have. Consequences, stated once and accepted: a guest has the Terminal tab, which is a root-capable shell on every host, and can make an account or token of its own that outlives the guest login. The one thing a guest is not is *an account*: it is left out of the Users list, out of "the last account can't be removed", and out of the single account an empty user name resolves to. |
 | Machines | **A view shows one machine: the one in the switcher.** No hub-side page lists or aggregates what its nodes are doing. To see or change anything on another machine, switch to it. | Mitchell, 2026-09-25, cutting a fleet-wide updates list out of the Docker tab. The switcher is the only place a machine is chosen, so there is exactly one answer to "which machine am I looking at" — a page that quietly reaches across breaks that, and it is never obvious which machine an action would land on. The hub still *proxies* for a node, and hub-only admin (Machines, Users, Settings, Alerts) is unaffected. |
 | Capability, not policy | **A tab offers what the selected machine can do, detected live — never what a rule says should run there.** Where the software is missing but installable, Ham Lab installs it; where it cannot be installed or the hardware lacks it, the tab says so in a sentence and offers what does work. Every action lands on the machine in the switcher. | Mitchell, 2026-10-09, after the Virtualization plan's first draft excluded the Mac mini and `hamtv` on the strength of `hamlet/CLAUDE.md`'s deployment rules: *"I want the option to do it anywhere I want, unless the hardware / software literally doesn't support it. Ham Lab needs to be built in a general-purpose way that can run on any Linux system."* Those rules bind the agent's unasked deployments, not the product. |
 | Ham Lab's data lives in Ham Lab's folder | **Everything a tab manages sits under the machine's data directory (`config.DataDir()`), never at a convenient path elsewhere.** Existing material is imported by copying. The engine's runtime state (`/var/lib/docker`, `/var/lib/incus`) is the engine's; Ham Lab keeps what would rebuild it. | Mitchell, 2026-10-09, when the Virtualization plan pointed its ISO library at `/tank/software/isos/`: *"ALL data needs to be under a HamLab folder, because it manages its own data."* The same reason `core`'s compose files moved under `apps/hamlab/stacks/` on 2026-10-05. `grid`'s and the Mac's stacks at `~/stacks/<machine>/` are the standing exception. |
 ## Layout
@@ -176,6 +180,44 @@ something long in one, switch away and back, refresh — it's all still there.
  the daemon's sampling clock; the chart window is always 300 samples, so 30 s at 100 ms
  and 10 min at 2 s. Still to do: persistent
  history (SQLite ring, e.g. 10 s × 7 days), SMART (`smartctl -j`), longer time ranges.
 - **Updating from the Updates tab** — **shipped 2026-09-26**: the "what is out of date" list
  was read-only, so seeing a newer image and doing something about it were different places.
  Each row now has **Update**, and the header has **Update all**. Updating is per stack —
  that is the unit compose deals in — so a row acts on whichever stacks use that image, and
  several stacks run through a queue one at a time rather than recreating half the machine at
  once, each with its task output on screen. A row that succeeds says "updated just now"
  instead of re-checking, since a check is a registry request.
 - **Recreate, and the Updates tab at full width** (2026-09-26): a stack's actions gained
  **Recreate** (`up -d --force-recreate`), the only way out of a container Docker refuses to
  start — one pinned to a network that has since been recreated answers "network … not
  found" to Deploy and Update alike, because as far as compose is concerned nothing about it
  changed. Found the hard way when a SiYuan update landed in exactly that state. The Updates
  tab also stopped rendering as a list-and-detail pair: it is a page, not a list, so on a
  wide window it now sits under the tab strip at full width instead of being squeezed into
  the right-hand half beside an empty column.
 - **Cron in English** (`crontext.ts`, 2026-09-26): schedules read "daily at 3:15am" rather
  than `15 3 * * *`, with the expression kept beside it. Handles lists, ranges, steps, day
  and month names, `@daily` and friends, weekdays/weekends, midnight and noon — and declines
  to describe anything it is not sure of, because a confident wrong sentence is worse than
  five fields. The Jobs tab had its own smaller version of this; it now uses this one.
 - **Hardware tab** — **shipped 2026-09-25** (`internal/hardware`, `internal/api/hardware.go`,
  `Hardware.tsx`): what the machine is physically made of, per machine like every other page.
  Four headline tiles (processor, memory, storage, graphics) over cards for each: the
  processor with its cores, clock range, socket, caches, microcode and the handful of feature
  flags anyone asks about; **every memory slot** with size, type, form factor, rated and
  running speed, maker, part number, serial, rank and voltage, and the empty ones marked
  empty; each drive with model, size, bus, firmware and serial; each display adapter with its
  driver, PCI address and video memory; and the board itself — model, firmware and date,
  serial, kernel. Read from what Debian already has (dmidecode for what only the firmware
  knows, lscpu, lspci, sysfs) and cached for ten minutes, since none of it changes under a
  running kernel; "Re-read" is there for the one thing that does, a drive plugged in.
  Every reader may come back empty and the page says what it could not read and why: a
  machine with no dmidecode, no root, or no PCI bus still gets a page. Verified on `core`
  (SMBIOS, Intel graphics) and on the Mac mini (Asahi: no SMBIOS memory table, a
  device-tree GPU, heterogeneous cores — where lscpu's "cores per socket" counts one
  cluster, so without SMT the thread count is the honest core count).
 - **Sensors** — **shipped 2026-09-19** (`internal/sensors`, `internal/metrics/sensors.go`,
  `internal/api/status_sensors.go`, `web/src/sensors.ts`, `widgets.tsx`, `Embed.tsx`;
  reference in `SENSORS.md`): everything measured is a sensor with a stable dotted ID;
@@ -259,6 +301,44 @@ something long in one, switch away and back, refresh — it's all still there.
  id and its route are gone. A rail order saved before this keeps its shape — `railorder.ts`
  maps the retired `top` to `procs` and drops the duplicate, so the merged page sits where
  Top sat rather than falling to the bottom.
 - **GPUs on the Processes tab** — **shipped 2026-10-02** (`internal/metrics/gpu.go`, the
  `GPUBox` in `Top.tsx`, `gpu.*` sensors). Asked for as "GPU activity and temps, like btop
  has it, always on". Each DRM device with a real GPU driver gets a box under the cpu box:
  busy curve, busy meter, a meter per engine (render/video/copy/enhance), temperature, clock,
  power and fan where the card has them, and the processes that have it open with their
  share. The process table gains a `Gpu%` column and sort. **How it is measured, and the
  honest gaps:** device busy is Intel's RC6 idle residency inverted (the counter every GT
  exposes without root; `intel_gpu_top` shows the same number) or amdgpu's own percent;
  per-process and per-engine shares are DRM fdinfo deltas, the way nvtop reads them. Other
  users' fdinfo needs root, so a daemon running as `core` keeps **one** root child
  (`hamlab drm-clients`, over a pipe) rather than a sudo a second. The Arc A380 on i915 6.12
  exposes power, energy and fan but **no temperature**, so its box shows none; an integrated
  GPU shows the CPU package temperature marked `*`. Asahi (the Mac mini) has neither a busy
  counter nor fdinfo in 6.19, so it lists the device and says so. No NVIDIA path: nothing here
  has one to test against. **Names come from an embedded PCI ID table** (2026-10-02,
  `internal/metrics/pciids_gpu.txt`, regenerated by `scripts/pci-ids-gpu.sh` from the hub's
  `pci.ids`): `gate` has no `lspci` and showed "Intel GPU 3185" beside `hamtv`'s "Intel UHD
  Graphics 600" for the identical chip. The table is tried first everywhere, so the name — and
  the sensor ID built from it — does not depend on which distro's pciutils a machine has;
  `lspci` is the fallback for a card newer than the table. Two fixes the same evening:
  the gpu box's layout rules had been lost from the stylesheet (v0.1.0-166/167), and on
  `grid` the sampler hung for 42 minutes in Go's poller on the Arc's `fan1_input` —
  `readStr` now uses plain blocking syscalls and each GPU's sysfs reads run under a
  3-second deadline (v0.1.0-168; `hamlet/docs/gotchas.md` has the mechanism).
 - **The ARC on the ZFS tab** — **shipped 2026-10-02** (`internal/metrics/arc.go`,
  `internal/storage/arc.go`, `GET /api/zfs/arc`, the Cache section in `ZFS.tsx`, `zfs.arc.*`
  sensors). Asked for as "current ARC usage, including any types of caching services in the
  pool". One card for the ARC: size against its ceiling with the target and floor, sparklines
  of size and hit rate from the sampler's history, what it holds (MRU/MFU/dirty, data vs
  metadata, compression, the ghost lists), the prefetcher and dbuf cache, memory-pressure
  counters, and the since-boot breakdown by demand/prefetch × data/metadata. Then a card per
  pool naming every caching tier — L2ARC, SLOG, special vdev, dedup table, spares — with its
  devices and fill, **or saying plainly that there is none and what that means** (the ZIL on
  the pool's own vdevs, with how much it has written), plus the properties that steer caching
  (`primarycache`, `secondarycache`, `sync`, `logbias`). Read from `/proc/spl/kstat/zfs/*`
  and `/sys/module/zfs/parameters`, with `zpool list -Hpv` and `zfs get` run without sudo
  (they need none) every 30 s at most. The day it shipped it showed `tank`'s ARC pinned at
  its 2 GiB floor under `arc_no_grow`, with a 53 GiB python3 process holding the RAM.
 - **Dashboard retired** — **2026-09-23**, at Mitchell's instruction, the same reasoning that
  merged Top into Processes: the app repeated itself. The customisable module list is gone
  (`Dashboard.tsx` deleted), and with it `/api/prefs/dashboard`, `applyPrefs`, and the
@@ -293,11 +373,66 @@ something long in one, switch away and back, refresh — it's all still there.
  carries the rule it was taken under. The page **detects the existing cron job**
  (zfSnap, zfs-auto-snapshot, sanoid, or a bare `zfs snapshot`) and says so, warning if a Ham
  Lab schedule is running as well.
-  **Nothing is destroyed unseen.** Every schedule answers *what would it delete?* with the
+  **Nothing is destroyed unseen.** Only names that parse as a stamp and a TTL are ever
-  list, sizes and dates, before Run now touches anything; only names that parse as a stamp
+  candidates for expiry, so a snapshot taken by hand is invisible to a schedule; Run now
-  and a TTL are candidates, so a snapshot taken by hand is invisible to it; and a snapshot
+  says that expired snapshots go with it; and a snapshot deleted by hand is confirmed first
-  deleted by hand needs its full name typed back. The scheduler is a 30 s tick in the daemon,
+  and needs its full name echoed back by the caller. (A "what would it delete?" preview
-  and `Run now` calls the same code, so the button and the timer cannot drift apart.
+  shipped and was removed the same day as redundant — Mitchell's call.)
  The datasets table is a **snapshot explorer**: each dataset shows how many snapshots it has
  and what they cost, and opens into the snapshots themselves — when each was taken, **when
  it expires** (read from the TTL in its own name, so no server call), its size, and Delete.
  Deleting takes the same-named snapshot on descendant datasets too, which is how a recursive
  schedule made it.
  **A schedule is a real cron job** (2026-09-24, Mitchell's call after the first cut ran it
  from a ticker inside the daemon). Ham Lab writes `/etc/cron.d/hamlab-zfs`, owns it, and
  rewrites it whole on every change; each line calls `hamlab zfs-snapshot <id>` as the
  account Ham Lab runs as, reaching zfs through sudo as the daemon does. Snapshots are
  backup-shaped and must not stop because a console is down, upgrading, or uninstalled — and
  a cron line is greppable by anyone looking at the machine the normal way. **A cron
  expression is the single description of when a schedule fires**: the four common shapes
  render to one, `custom` states one outright, and the next run is read back with the same
  evaluator the Jobs tab uses on everyone else's lines, so the page, the file and the
  prediction cannot disagree. A run's outcome is written to `zfs-runs/<id>.json` beside the
  config, never to `state.json`, because the cron job is a separate process and the daemon
  would clobber it. An existing external job can be **adopted** — its schedule, retention
  and dataset read out of the cron line into a new Ham Lab schedule.
  **zfSnap is retired on `grid`** as of that day: Ham Lab now takes the nightly `tank`
  snapshot, the package is purged and the old cron line is gone, with the 515 snapshots it
  had taken left untouched and fully readable here — which is what adopting its naming
  bought. See `hamlet/docs/hosts.md`.
 - **KVM tab** — **shipped 2026-09-24** (`internal/api/kvm.go`, `KVM.tsx`, `config.KVM`).
  Paste a device's address; its own web page appears in the tab. **Nothing in it knows
  anything about any particular KVM** — the first device here is a Sipeed NanoKVM on `grid`
  at `10.0.0.217`, but none of that is in the code.
  **Why it proxies rather than framing the device directly.** Ham Lab is served over TLS and
  a KVM usually is not, so a bare `http://…` iframe is blocked as mixed content before it
  loads. Serving the device from Ham Lab's own origin (`/kvm/<id>/…`) fixes that, puts it
  behind **Ham Lab's** login rather than a second one, and makes the iframe same-origin.
  **The general problem, and the answer.** A device UI asks for its assets by absolute path
  — `/assets/app.js`, `/api/whatever` — which no path prefix can rewrite away, and rewriting
  bodies per device would not be general. So the prefix is only the entry point: any request
  Ham Lab does not serve itself is routed to whichever KVM the asking page belongs to, read
  from the **Referer**, falling back to a **cookie** when there is none (a WebSocket upgrade).
  Ham Lab's own routes always match first, so a device can never shadow the console, and the
  fallback is gated on a live session. The device never receives Ham Lab's session cookie,
  and its `X-Frame-Options`/CSP/HSTS are dropped on the way out since Ham Lab is the origin
  now. Redirects to the device's root are rewritten back through the prefix.
  **One catch found in testing:** Ham Lab's own `frame-ancestors 'none'` forbids framing even
  from the same origin, so `/kvm/` paths send `'self'` instead. Without that the iframe was
  refused before it loaded.
  **The keyboard** needs less than it sounds: a page cannot forward keys into an iframe and
  does not need to, since one with focus receives them directly. Fullscreen plus
  `navigator.keyboard.lock()` is what takes Escape, Tab and Ctrl-W back from the browser;
  that is Chromium-only, and the tab says which version you are getting rather than
  pretending.
  **One KVM per machine, and the tab follows the machine switcher** (corrected the same day:
  the first cut was a hub-side list you pointed at machines, which had it backwards). Switch
  to `grid` and you get grid's; adding a second for a machine replaces the first. The
  addresses are still kept **on the hub** rather than on each machine, because a KVM earns
  its keep exactly when the machine is down and its own Ham Lab cannot answer — the hub
  proxies the device either way — and a KVM need not belong to a machine Ham Lab manages at
  all. Not done: holding the device's own login on the hub, so its page does not ask for a
  second password.
 - **Skins** — 13 since 2026-09-19 (`themes.ts`: a palette is one `mk(…)` line): Dark,
  Light, Cyberpunk, Retro, **HUD** (hairlines of cyan on near-black, corner ticks, thin
  type), Nord, Gruvbox, Catppuccin + Latte, Dracula, Tokyo Night, Solarized Dark + Light;
@@ -343,9 +478,16 @@ something long in one, switch away and back, refresh — it's all still there.
  replace, exec in place. Config default moved to `~/.config/hamlab` (existing
  `~/apps/hamlab` still found by the installer and the CLI). Tested in
  containers: Debian 12, Ubuntu 22.04, Fedora 40, Rocky 8, openSUSE Leap 15.5, Alpine
-  3.18/3.20 (OpenRC + doas, non-root), plus the four lab hosts. Not done: an OpenRC
+  3.18/3.20 (OpenRC + doas, non-root), plus the four lab hosts. **procd (OpenWrt) added
-  backend for the Services tab; runit/s6/sysvinit service files; public hosting for
+  2026-09-26** (`internal/setup/hamlab.procd`): detected on `/sbin/procd` + `/etc/rc.common`
-  `install.sh` (the Gitea repo is private, so the one-liner needs `HAMLAB_TOKEN`).
+  ahead of the OpenRC branch, installed to the same `/etc/init.d/hamlab` path and told apart
  from OpenRC at uninstall by `USE_PROCD=1` in the file. Verified on the new router
  (`gate`, OpenWrt 25.12.5, running as root): `S95`/`K10` symlinks, procd supervising the
  instance, Docker tab and stacks working through the hub. The Updates tab needed no change
  — apk-tools 3 answers `apk version -l '<'` the same way Alpine's does. Not done: an
  OpenRC or procd backend for the Services tab (it is systemd-only and says so cleanly);
  runit/s6/sysvinit service files; public hosting for `install.sh` (the Gitea repo is
  private, so the one-liner needs `HAMLAB_TOKEN`).
 - **Users, sessions and settings that follow the account** — **shipped 2026-09-22**
  (`internal/api/users.go`, `config.User`, `store.Session`/`UserState`, `Login.tsx`,
  `Users.tsx` — a hub-only **Users** tab in the rail, added the same day at Mitchell's
@@ -485,13 +627,94 @@ something long in one, switch away and back, refresh — it's all still there.
  `POST /api/password`, current password required), Access tokens (list/revoke the tokens
  hubs hold), About (version, host, arch, uptime, certificate).
 ## Network tab (2026-09-27)
 A **Network** tab, per machine like every other: the interfaces as the kernel has them
 (addresses, MAC, MTU, negotiated link speed, carrier, byte totals, live rates from the
 sampler's own sensors), the default route and resolvers, what the internet sees of this
 connection, and an **internet speed test**.
 Decisions worth keeping:
 - **Measured by the daemon, not the browser.** The test runs on the machine in the switcher,
  so the figure is that machine's line — the router's, if you switch to `gate` — and it
  carries on if the phone locks or the tab closes. The page polls a run it is not driving.
 - **No tool to install.** It measures against Cloudflare's public speed endpoints
  (`__down`, `__up`, `cdn-cgi/trace`), which is what makes it work identically on Debian and
  on OpenWrt, where `opkg install` costs flash. That is the feature's one external
  dependency, and it is named in `internal/network/speedtest.go`.
 - **Several streams, HTTP/2 off, and the ramp thrown away.** One stream over one connection
  measures a congestion window; HTTP/2 would put every stream back on one connection and do
  the same. The first second of each phase is measured and discarded.
 - **Latency is the round trip less the far end's own Server-Timing.** A zero-byte response
  is mostly Cloudflare's worker running — 20 to 45 ms of it against a few ms of wire. Left
  in, it reported 31 ms for a 7 ms line. Jitter is the mean change between consecutive
  probes, not a standard deviation, so one slow probe doesn't run away with it.
 - **A refused phase is not a number.** The far end rate-limits by volume, and a test moves a
  few hundred megabytes; several in a row earn 429s, and the block observed while building this
  lasted over twenty minutes. So: a 20-second cooldown counted from the last *attempt* (a
  refused test still used the line), a stream that retries twice before giving up, the rate
  computed over the window in which bytes were actually still arriving, and a plain sentence
  when it was cut short. Coasting to the end of the clock with nothing arriving reported a
  third of the truth.
 - **A second download source, for when the first refuses.** A plain 100 MB file on Linode's
  Seattle mirror, which has no opinion about volume. It is only reached after a refusal, and
  every result names the source it used — that mirror is in one place, so on a machine far from
  Seattle it would read low, and a number that quietly came from somewhere else would be a lie.
  Upload stays on Cloudflare, which has never been the one to refuse.
 - **The last result is published as sensors** (`net.internet.*`), so the line is a widget, a
  graph, an alert or a Prometheus series like anything else. Each carries the time the test
  ran, not the time it was asked for.
 - **Not scheduled.** Nothing takes a speed test on its own. A nightly one recorded to the
  history is the obvious next step and deliberately not built yet — the cron shape for it
  already exists twice over (ZFS snapshots, Docker image updates).
 ## VPN tab (2026-09-27)
 A **VPN** tab beside Network, per machine: every WireGuard tunnel on the machine in the
 switcher, and the devices on it — name, address, last handshake, where it connects from, and
 what it has moved.
 - **Per machine is the point, not a limitation.** The tunnel people dial into and a tunnel a
  machine dials out through are on different boxes here (`gate` and `hamtv`), and a combined
  view would invite exactly the question "which box is this on".
 - **The kernel is the truth, the config only supplies names.** `wg show all dump` has the
  handshakes and counters and knows nothing but public keys; names come from the `# comment`
  above a peer in a wg-quick file, or a UCI section's `description` on OpenWrt. Both are read
  only to label what the kernel already said.
 - **Server or client is worked out, not declared.** A peer allowed `0.0.0.0/0` that has an
  endpoint is a way out, so a tunnel holding one is this machine dialling out; a tunnel with a
  listen port and peers is one being dialled. Anything else says neither rather than guessing.
 - **"Online" is a handshake in the last three minutes**, and the tab says so. WireGuard is
  silent when idle; there is no connection to be in, and a tab claiming otherwise would be
  lying twice a day.
 - **No secrets are read.** Private keys are in the second column of the dump and are read
  past; pre-shared keys are noted as present and never read.
 - **Devices can be created, renamed and removed** (added 2026-09-27, asked for the same day),
  on a tunnel the machine configures — today OpenWrt's UCI, which is where the server here
  lives. A wg-quick host is reported unmanaged and shows no buttons: half-supporting a second
  configuration format is worse than an honest refusal.
  - **Nothing restarts the network.** `uci commit` writes the file and applies nothing; `wg
    set` changes one peer on the live interface. Proven on `gate` by the existing peers'
    transfer counters surviving a create and a delete — a reload would have zeroed them.
  - **The two halves are ordered so they cannot disagree quietly.** Adding writes the
    configuration first and rolls it back if the live tunnel refuses, because a device that
    works only after the next reboot is the worst of the three outcomes. Removing does the
    live tunnel first, since that is the moment access ends, and says so plainly if the
    configuration half then fails.
  - **The keypair is generated on the machine** and the profile, with its QR, is shown once.
    The private key is stored nowhere — the same bargain as a guest login's password.
  - **Removing asks for the name to be typed back**, at both ends. It cannot be undone and it
    sits one tap from a list of devices that all look alike.
 ## To do
- **Deprecate copyparty** — asked for 2026-09-23, not started. `copyparty` on `grid`
+- **Grow the Files tab until copyparty can go** — asked for 2026-09-23 as two items (Files
-  (`copyparty.service`, `:3923`, `/etc/copyparty.conf`, serving all of `/tank` as user
+  polish, and deprecating copyparty), merged 2026-09-24: they are the same tab and the same
-  `core` with one account behind NPM) overlaps the **Files** tab almost entirely. The aim
+  work. `copyparty` on `grid` (`copyparty.service`, `:3923`, `/etc/copyparty.conf`, serving
-  is to stop the service and keep only Ham Lab. What Files must gain first, in the order
+  all of `/tank` as user `core` with one account behind NPM) overlaps **Files** almost
-  that matters:
+  entirely. The aim is to stop the service and keep only Ham Lab, with general ease-of-use
  polish along the way. What Files must gain first, in the order that matters:
  - **Multi-select and bulk actions.** copyparty selects many rows and moves, deletes or
    downloads them together; Files acts on one row at a time.
  - **Download a folder**, and a multi-file download, as a streamed zip or tar — the thing
@@ -511,77 +734,68 @@ something long in one, switch away and back, refresh — it's all still there.
    `hamlet`'s `docs/hosts.md` bare-metal list, `docs/gotchas.md` (the "`docker ps` is not
    the inventory" note names it) and the `grid` row in `CLAUDE.md`.
- **Deprecate Dockhand** — asked for 2026-09-23, not started; the *homelab* half of this
+- **Grow the Docker tab until Dockhand can go** — asked for 2026-09-23 and widened
-  (migrating every stack off Dockhand's data dir, then removing the container) is
+  2026-09-24 to take in Docker tab polish and feature work, on the same reasoning as the
-  `hamlet`'s `docs/pending.md` → "Retire Dockhand; Ham Lab becomes the orchestrator",
+  Files item: it is one tab, and the gaps that block retiring the old tool and the things
-  agreed 2026-09-20. This entry is the *app* half only: what Ham Lab must do before that
+  that would simply make it better are the same body of work. The *homelab* half —
-  migration can start. Its Stacks page already creates, edits, deploys, updates, restarts,
+  the cutover itself, then moving every stack off Dockhand's data dir — is `hamlet`'s
-  stops, downs and deletes stacks on every machine, so the gaps are Dockhand's extras:
+  `docs/pending.md` → "Retire Dockhand; Ham Lab becomes the orchestrator", agreed
-  - **Git-sourced stacks.** Dockhand's `stack_sources` table points a stack at a repo and
+  2026-09-20. This entry is the *app* half.
    pulls it. Ham Lab has no equivalent. Decide whether to build it or to drop it — most
    stacks here have no row in that table anyway, so dropping it may be honest.
  - **Scheduled image updates.** Ham Lab checks for new images only when asked (a registry
    request Docker Hub counts) and never updates by itself. If anything relies on Dockhand
    updating on a schedule, that must move here or be dropped on purpose.
  - **Webhooks / per-host environments**, if either is in use — inventory before deciding.
  - **Deploying from a host path without an orchestrator container.** Ham Lab already runs
    compose inside `dockhand` because that is how the stacks were deployed; once they move
    to a neutral home the "via" path is unused for them, but a stack deployed to `grid` or
    the Mac still goes over `tcp://<node>:2375` from the hub. Decide whether Ham Lab should
    instead ask the node's own Ham Lab to deploy locally, which would end the `*:2375`
    exposure (its own item in `docs/pending.md`).
  - *Not app work, but the trigger:* nothing here blocks starting the migration except the
    inventory. Do that first.
- **A Virtualization tab** — asked for 2026-09-23, not started, and **large**. Run and
+  **Done 2026-09-24 — the eight Mitchell picked from the inventory**, all in the Docker tab:
-  manage **LXC containers and virtual machines** from the console, the way Proxmox does —
+  per-container **Stats**, **Health** (with the last few probe outputs), **Events** (Ham Lab
-  which is the comparison Ham Lab has invited since its first line. Nothing exists yet.
+  keeps its own 4000-entry ring, since `docker events` forgets after about a minute),
-  What a future session has to settle before writing code: which stack (libvirt/KVM +
+  **Shell** (`docker exec`, handed to the Terminal tab) and **Limits** (`docker update` for
-  `virsh` is the obvious one for VMs; LXC or Incus for containers, and Incus covers both);
+  memory, CPU and restart policy); an **updates** tab giving the whole machine at once with
-  whether it manages what is already installed or installs it; where disk images live
+  a per-container indicator in the list; and **schedules**, which write
-  (`tank` on `grid`, local NVMe elsewhere); console access (SPICE/VNC proxied through the
+  `/etc/cron.d/hamlab-docker` and call
-  daemon, or a serial console over the existing terminal machinery, which would reuse
+  `hamlab docker-update <id>` — the shape the ZFS schedules already use.
-  `internal/term`); snapshots and clones; and what it does on a host with no virtualisation
+  Also done: Ham Lab keeps working when `dockhand` is removed. It remembers where a
-  at all, which is every host here today. Worth scoping as its own plan before starting.
+  container's bind mounts pointed (`mounts` in `state.json`) and resolves a stack through
  that record when no running container explains its path, running compose on the host
  instead; `repoint` does the same for a schedule at run time. Without this, removing the
  container would have made 17 stacks on `core` and 5 on `grid` unfindable at once.
- **Files explorer: cleanup and ease of use** — asked for 2026-09-23, not started.
+  - **Left from the Dockhand inventory, both to decide rather than build.**
-  General polish and ease of use. The **Deprecate copyparty**
+    - *Git-sourced stacks.* Dockhand's `stack_sources` table points a stack at a repo and
-  item names the heavier pieces (multi-select and bulk actions, folder download, resumable
+      pulls it. Its `git_repositories` and `git_stacks` tables are **empty**, so nothing
-  uploads, a gallery, search) — worth doing the two together, since they are the same tab.
+      here uses it — dropping it is the honest answer unless Mitchell says otherwise.
    - *Webhooks, OIDC, LDAP, vulnerability scans.* All four tables empty. Nothing to move.
  **Done 2026-09-25 — the two gaps in the tab itself.**
  - *Logs are followed over a held connection*, the way the Services tab follows a journal
    (`internal/api/docker_logs.go`, `/api/docker/containers/{id}/logs/tail`): one
    `docker logs -f` per viewer, stdout and stderr both, stderr coloured as an error. A
    reconnect resumes from the last line's timestamp, so a container that restarts or a
    phone that slept doesn't replay what is already on screen; the client dedupes by line
    rather than by "later than", because `--since` is inclusive and two lines can share a
    timestamp. Docker replays the opening backlog down two pipes in no particular order, so
    the first 300 ms is held and sorted before it is sent. The polled `?since=` endpoint
    stays — it is what the resume is built on. Reconnects back off to 15 s, since a
    container that has stopped ends the stream immediately.
  - *Networks can be removed and pruned* — `POST /api/docker/networks/{remove,prune}`, the
    shape images and volumes already had. The overview says who is attached to each one, so
    the ✕ is only offered on a network nothing uses and never on Docker's own three. Prune
    is deliberately untested against this host: `core` has several networks belonging to
    stacks that are merely down (`pihole_default`, `transmission_default`, `gluetun_default`
    among them), and pruning would take them and hand them new subnets on the next deploy.
  - **The `*:2375` question, which this is the moment to settle.** A stack deployed to
    `grid` or the Mac still goes over `tcp://<node>:2375` from the hub. Decide whether the
    hub should instead ask the node's own Ham Lab to deploy locally, which would end the
    unauthenticated Docker socket exposure — its own item in `docs/pending.md`, currently
    filed under "decided, no action", and this is the change that would make closing it
    free.
 - **A Virtualization tab** — asked for 2026-09-23; **scoped 2026-10-09 in `VIRTUALIZATION.md`**,
  not started. Incus 6.0 LTS as the engine; the tab detects per machine whether Incus is
  running, installable, or unavailable, and whether KVM is there, and acts on the machine in
  the switcher — never a default host. Containers and VMs in one list, console/shell through
  the terminal machinery, snapshots, exports, and the old Proxmox estate on `tank` as a later
  import. That file has the probe of every host, the per-machine layout Ham Lab computes
  (pool driver, a bridge subnet that avoids the LAN's /8 and Docker's pools, the Docker
  `FORWARD` fix) and the phased build.
 - **A KVM tab** — asked for 2026-09-24, not started. **Dead simple is the brief**: paste one
  HTTP link, the tab shows whatever UI that KVM serves, and the keyboard goes to it. **This
  is KVM the keyboard-video-mouse switch, not KVM the hypervisor** — the Virtualization item
  above means the other one.
  - **The device**: a **Sipeed NanoKVM** attached to `grid` at **`10.0.0.217`** (MAC
    `48:da:35:6f:d0:31`), web UI on port 80. The only one so far, so store a list, not one URL.
  - **Framing works.** Probed 2026-09-24: it sends **no `X-Frame-Options` and no
    `frame-ancestors`**, so it does not refuse to be embedded. No fallback needed.
  - **Proxy it through the daemon — that is what makes the simple version possible.** The
    KVM is **HTTP only**, and Ham Lab is served over TLS, so a bare `http://…` iframe is
    blocked as mixed content and the "just paste a link" idea dies there. Serving it from Ham
    Lab's own origin (`/api/kvm/<id>/…`, on the `proxyNode` machinery, which already carries
    WebSockets — exactly what the video and input streams are) fixes three things at once:
    no mixed content, the KVM sits behind **Ham Lab's** login rather than a second one
    (`/api/vm/info` answers 401, so it has its own, which the hub can hold once), and the
    iframe becomes **same-origin**, which matters for the keyboard below.
  - **The keyboard needs less than it sounds.** A parent page *cannot* capture keys and
    forward them into a **cross-origin** iframe — the browser forbids synthetic events across
    origins — but it does not need to: an iframe with focus receives keystrokes directly, and
    the NanoKVM's own page already handles them. So "capture and give it to the iframe" is
    really: click-to-focus, keep focus there, and stop the console's own shortcuts from
    eating keys while it is focused. What is genuinely worth adding is **fullscreen** plus
    `navigator.keyboard.lock()`, which is what stops the *browser* swallowing Escape, Tab and
    Ctrl-W so they reach the guest — note it is Chromium-only, so Firefox and iOS Safari get
    the plain version. Proxying makes the iframe same-origin anyway, so real forwarding stays
    available if some KVM ever needs it.
  - **Where the list lives is the one real design decision.** A KVM earns its keep when the
    machine is *down* — no OS, a bad kernel, a BIOS screen — which is when that host's Ham Lab
    cannot answer. So it is **hub-side** and must work for machines running no Ham Lab at all.
    A `kvm_url` in each node's own `config.json` would fail in exactly the case it is for.
  - **Do not give it an NPM host** without deciding to: `*.bigham.dev` is a wildcard to the
    public IP, so that would put BIOS-level control of `grid` on the internet behind only the
    KVM's own login. Behind Ham Lab's proxy it inherits Ham Lab's auth instead.
 ## Open questions
@@ -0,0 +1,164 @@
 # Portability audit — what in Ham Lab is specific to this lab (2026-10-09)
 Goal: Ham Lab drops onto any set of Linux machines and works. This is the full sweep of
 every place the code, defaults, docs or UI assume Mitchell's hosts, network, hardware,
 distro, init system, orchestrator history or posture. **Nothing here is fixed yet** — this
 is the list to work from. Each item: where, what it assumes, what the general form is.
 Severity key: **A** breaks or misbehaves on an ordinary host · **B** works but is wrong by
 default or phones home to this lab · **C** cosmetic (placeholder text, comments, docs).
 ---
 ## 1. Hard-wired lab identifiers
 | # | Sev | Where | Assumes | General form |
 |---|---|---|---|---|
 | 1.1 | **B** | `internal/selfupdate/selfupdate.go:36` `DefaultURL = "https://git.bigham.dev/api/v1/repos/mitchell/hamlab"` | Every install with no `update.url` checks **this** Gitea 90 s after boot and daily (`internal/api/update.go:1951, 2016`). A stranger's install phones home here by default. | Default should be **off** (or a build-time `-X` variable set at release). Installer asks or `install.sh` passes `HAMLAB_URL` through into `update.url`. |
 | 1.2 | **B** | `install.sh:4, 27` `HAMLAB_URL` default; `README.md:23-25, 31` | The curl one-liner and the install source are this Gitea. | Default comes from the release that served the script (`@URL@` filled by `make release`), or is required. |
 | 1.3 | **B** | `scripts/gitea-release.sh:14` `REPO` default | Publishing goes to `mitchell/hamlab` on this Gitea. | Require `HAMLAB_REPO`; no default. |
 | 1.4 | **C** | `README.md:100` "On `core` it is public at `https://hamlab.bigham.dev` behind Nginx Proxy Manager" | Lab deployment fact in the product README. | Move to `hamlet/docs/hosts.md`. |
 | 1.5 | **C** | `examples/external-demo/index.html:8` `base: 'http://10.0.0.234:8100'` | Demo config names the hub's LAN IP. | `http://<hamlab-host>:8100`. |
 | 1.6 | **C** | `SENSORS.md:149` Prometheus target `10.0.0.234:8100`; `:187` `&machine=grid`; `:196-208` Glance examples at `hamlab.bigham.dev`, `zfs.tank.used_pct`, `/api/nodes/grid/...`; `:218-220` `grid,wyse`, `{"core":…,"grid":…,"mac":…}` | Docs use real host names, pool and domain. | Generic placeholders (`<host>`, `<machine>`, `<pool>`). |
 | 1.7 | **C** | `PLAN.md:3-4` "bare-metal Debian 13 host. First target: `core` (10.0.0.234)"; decisions table rows (Listen `10.0.0.234:8100`, Data `/home/core/apps/hamlab/`, Ingress NPM/`hamlab.bigham.dev`, Repo) | Product plan is written as this lab's plan. | Keep as history, or split lab facts out to `hamlet`. The repo is public. |
 | 1.8 | **C** | `VIRTUALIZATION.md:339-353` per-machine probe table; `:412, 425-436` `~/apps/hamlab`, `/tank/...`, boot NVMe free space | Plan is written against the five lab hosts. | Same as 1.7. |
 | 1.9 | **C** | `BUGS.md` throughout | Bug bashes narrate `core`, `grid`, `wyse`, `hamtv`, `gate`, the Mac. | History; fine to leave, but it is public. |
 | 1.10 | **C** | `cmd/hamlab/main.go:1` "Proxmox-style web console for a bare-metal Debian host"; `internal/hardware/hardware.go:5` "what Debian and its kin already have"; `internal/hardware/hardware.go:29` `// "GMKtec NucBox G3"`; `internal/metrics/metrics.go:1358` "did, on grid, 2026-10-02"; `internal/api/server.go:1382-1383` `["grid", …]`, `?machines=grid,wyse`; `internal/api/stacks.go:10` `/home/core/apps/dockhand/stacks/…`; `internal/api/dockersched.go:5` "unbound, nginx and pihole"; `internal/api/zfs.go:5` "grid already runs zfSnap"; `internal/storage/snapshots.go:6` "grid has run zfSnap v1.11.1"; `internal/api/guests.go:8` "Mitchell, 2026-09-27"; `internal/network/network.go:306` "the public address here is a DHCP lease"; `internal/metrics/gpu.go:17` "Asahi, today"; `internal/metrics/pciids.go:4-6` "the router's GPU"; `web/src/Top.tsx:522, 576`; `web/src/hashsel.ts:1-9`; `web/src/Embed.tsx:3` `&machine=grid` | Comments explain decisions by naming lab hosts. | Harmless, but rewrite the ones in package doc comments so a reader of the source doesn't need the lab to understand the code. |
 ## 2. UI placeholders and defaults that are this lab's
 | # | Sev | Where | Text | General form |
 |---|---|---|---|---|
 | 2.1 | **C** | `web/src/Machines.tsx:116-118` | `placeholder="grid"`, `"https://10.0.0.214:8100"`, `"core"` | `nas`, `https://192.0.2.10:8100`, `admin` or the hub's own user. |
 | 2.2 | **C** | `web/src/KVM.tsx:166` | `placeholder="10.0.0.217"` (the NanoKVM on `grid`) | `192.0.2.20` or `kvm.local`. |
 | 2.3 | **C** | `web/src/Settings.tsx:242` | `placeholder="cpu.*, temp.*, zfs.tank.*"` | `cpu.*, temp.*, fs.*`. |
 | 2.4 | **C** | `web/src/Settings.tsx:225` | read-only token default name `'glance'` | Empty, or `dashboard`. |
 | 2.5 | **C** | `web/src/Alerts.tsx:38, 40` | `"Tank nearly full"`, `"zfs.tank.used_pct"` | `"Disk nearly full"`, `"fs.root.used_pct"`. |
 | 2.6 | **C** | `web/src/StackNew.tsx:110, 113, 114` | `/home/core/apps/whoami:/data`, `TZ=America/Los_Angeles`, `com.example.owner=mitchell` | `./data:/data`, `TZ=UTC`, `com.example.owner=you`. |
 | 2.7 | **C** | `web/src/Gallery.tsx:69`, `web/src/WidgetPicker.tsx:36` | `"find — gitea, temp, enp3s0…"`, `"filter sensors — cpu, temp, fan, enp3s0…"` | Drop the container and NIC names. |
 | 2.8 | **C** | `web/src/VPN.tsx:105` | default DNS for a new device `1.1.1.1, 1.0.0.1` | Fine as a public default, but should offer the tunnel's own address (the home resolver) first. |
 | 2.9 | **B** | `web/src/ZFS.tsx:212` `BLANK` | new schedule defaults `at: '02:30'`, `keep_for: '6w'` — `grid`'s old zfSnap cron line | Neutral defaults (`03:00`, `4w`), or none. |
 | 2.10 | **A** | `web/src/ZFS.tsx:302` `adopt()` | `/\b(tank[\w/-]*)\b/` — the dataset of an adopted external job is found by looking for the word **`tank`** | Parse the dataset from the command as the last `zfs`-shaped argument, or match against the datasets the page already has. On any other pool name adopt silently picks the fallback. |
 ## 3. Topology inherited from the lab (hub `core`, Dockhand, `tcp://:2375`)
 | # | Sev | Where | Assumes | General form |
 |---|---|---|---|---|
 | 3.1 | **A** | `internal/api/stacks.go:483-509` `nodeDocker()` | For every node whose stacks aren't editable locally, the hub **probes `tcp://<node-host>:2375`** with `docker -H … version` and, if it answers, deploys to the node over an unauthenticated Docker socket. This is the Dockhand-era layout; on any other network an open 2375 is either absent or a security problem the console should not quietly use. | Remove the probe. The node's own Ham Lab deploys its stacks (the files live on the node, under its data dir). `PLAN.md` "To do" already names this decision. |
 | 3.2 | **A** | `internal/api/stacks.go:439-479` `hubStacks()`, `:1036-1074` `homesFor()`, `web/src/Docker.tsx` "kept on `<hub>`, deployed to this machine over `tcp://…`" | A hub can hold compose files for a node and run compose remotely; new stacks for a node may be created **on the hub**. Only true here because Dockhand on `core` deployed `grid` and the Mac. | Same as 3.1: one machine, one set of files, one Docker. Hub-held stacks become an import path (copy to the node) rather than a steady state. |
 | 3.3 | **B** | `internal/api/stacks.go:290-382` `homesOf()`; `:373-376` fallback `~/stacks` | Where new stacks go is **discovered** from where ≥2 stacks already are (Dockhand's `stacks/<machine>` habit), else `~/stacks`. Neither is Ham Lab's data directory, which the project's own rule says everything managed must live under. | Default home = `<DataDir>/stacks`. Discovered homes stay as *additional* places, never the default. |
 | 3.4 | **B** | `internal/api/stacks.go:100-203` remembered mounts, `internal/api/dockersched.go:772-795` `repoint()` | Machinery exists to find compose files **inside a removed orchestrator container's filesystem**. Generic in form, but it exists because of Dockhand and is exercised only here. | Keep (it is general), but it is not a reason to keep 3.1/3.2. |
 | 3.5 | **B** | `internal/api/stacks.go:801` | Inside a `Via` container the command is literally `docker compose`, even when `dockerTool()` is `podman`. | Use the detected tool. |
 | 3.6 | **B** | `internal/api/stacks.go:908` `docker buildx imagetools inspect` | Update checks need the **buildx** plugin. Docker CE on Debian has it; Podman and minimal Docker installs do not → every image reports "the registry didn't answer". | Fall back to `skopeo inspect` or a direct registry `HEAD` for the digest. |
 | 3.7 | **B** | `Makefile:445-448` `deploy` | The machine you build on **is** the hub: installs locally, then `push-nodes`. | Fine for a developer; document that `deploy` means "I am the hub". A plain `make install` target for a single box. |
 | 3.8 | **B** | `internal/nodes/nodes.go:316-323` | A hub pushes other architectures only if `<prefix>/hamlab-<arch>` siblings exist, which only `make deploy`/`install.sh`-from-dist put there. | A hub without siblings should fetch the right build from `update.url` instead of erroring. |
 ## 4. Init system: systemd assumed where the product promises procd and OpenRC too
 The installer supports systemd, procd and OpenRC (`internal/setup/setup.go:217-262`). These paths do not.
 | # | Sev | Where | Assumes | General form |
 |---|---|---|---|---|
 | 4.1 | **A** | `internal/updates/updates.go:537-543` `Power()` | `systemctl reboot` / `systemctl poweroff`. On OpenWrt (`gate`) and OpenRC the Power dialog fails. | Try `systemctl`, else `reboot` / `poweroff` (busybox, OpenRC, sysvinit all have them). |
 | 4.2 | **A** | `internal/incus/incus.go:1017, 1044-1058, 1069, 1075` | The Incus install script writes `/etc/systemd/system/*` drop-ins and units, adds `SupplementaryGroups` to `hamlab.service`, restarts with `systemd-run`. On OpenRC/procd the daemon never joins `incus-admin` → the tab stays in "repair" forever; the Docker FORWARD fix is never persisted. `HamlabUnit = "hamlab.service"` (`:35`) is surfaced in the API (`:280`). | Branch on the detected init: OpenRC `rc-update`/`conf.d`, procd init script edits, or fall back to `usermod` + a daemon restart via the installer's own supervisor knowledge. Docker rules via a boot hook that is not a unit. |
 | 4.3 | **A** | `internal/services/services.go:163-170` | Services tab is systemd-only ("says so cleanly"). `gate` (procd) and any OpenRC/runit/s6 box have no Services tab. | `rc-service`/`rc-status` (OpenRC), `/etc/init.d/<x> status` + `ubus call service list` (procd), `sv` (runit) backends. Noted as "not done" in `PLAN.md:487-489`. |
 | 4.4 | **A** | `internal/cron/cron.go:353-373, 401, 442, 532-534, 669-670` | Jobs tab reads `/etc/crontab`, `/etc/cron.d`, `/var/spool/cron{,/crontabs}` (Debian/cronie), `/usr/sbin/anacron`, last-runs from **journald**, timers via `systemctl`; writes only `/etc/crontab` and `/etc/cron.d`. Busybox crond (OpenWrt, Alpine) keeps crontabs in `/etc/crontabs/<user>` and has no `cron.d` → the Jobs tab is empty on `gate` and schedules written by 4.5 never run. | Add `/etc/crontabs/*` as a source; last-runs from syslog/logread when there is no journal. |
 | 4.5 | **A** | `internal/storage/snapshots.go:402` `/etc/cron.d/hamlab-zfs`; `internal/api/dockersched.go:65` `/etc/cron.d/hamlab-docker` | ZFS snapshot and image-update schedules are cron.d files. No cron.d on busybox crond; no cron at all on a systemd-timer-only host (Fedora Server, CoreOS). | Detect: cron.d → cron.d; busybox → a line in `/etc/crontabs/<user>`; none but systemd → a timer unit. The daemon-side runner already exists (`hamlab zfs-snapshot`, `docker-update`). |
 | 4.6 | **B** | `internal/hardware/hardware.go:333` `systemd-detect-virt` | Virtualisation detection is systemd-only. | Fall back to `/sys/class/dmi/id/sys_vendor`, `/proc/cpuinfo` `hypervisor` flag, `/sys/devices/virtual/dmi`. |
 | 4.7 | **B** | `internal/api/server.go:1129`, `internal/updates/updates.go:187` | `/run/reboot-required` is Debian/Ubuntu-only. `updates.Status` falls back to a kernel check for other distros; `/api/me.reboot_needed` (the Machines badge, the header's amber power button) does not. | Share one function. |
 | 4.8 | **B** | `install.sh:154-157, 164, 375-378`; `internal/setup/setup.go:284` | The wizard's init detection knows systemd and OpenRC only: on OpenWrt it warns "no systemd or OpenRC: nothing starts it at boot" although `hamlab install` then installs a procd script fine; the "logs" hint and the "isn't answering yet" hint have no procd case. | Add procd to both. |
 | 4.9 | **B** | `internal/setup/setup.go:609-641` `firewall()` | Only firewalld and ufw are detected. OpenWrt fw4/nftables, plain nftables, iptables, Alpine awall are not. | Detect `fw4`/`nft`/`iptables`; on OpenWrt add a LAN-zone rule or say what to open. |
 | 4.10 | **B** | `internal/setup/setup.go:568-605` `quietSudoers()`; `internal/services/services.go:176-185` | The sudoers quieting list is tuned to what the lab polls (`smartctl`, `pwrstat`, `journalctl`, `zpool`, `zfs`). A host with NUT or apcupsd gets `upsc`/`apcaccess` lines in its auth log at 10 s intervals; a host without `/etc/sudoers.d` (OpenWrt) gets nothing, which is correct. | Build the list from what the collector actually probes. |
 ## 5. Distro and package assumptions beyond init
 | # | Sev | Where | Assumes | General form |
 |---|---|---|---|---|
 | 5.1 | **B** | `internal/incus/incus.go:213-250, 351-352` | Incus can be installed with apt, dnf, pacman, apk. **zypper is missing** (the Updates tab supports it), so an openSUSE host reads "No package manager Ham Lab knows how to drive". `opkg` is known only to say "not packaged for OpenWrt". | Add zypper; phrase the opkg case as "not packaged for this distribution". |
 | 5.2 | **B** | `internal/incus/incus.go:973-978` | dnf branch adds `qemu-system-x86-core`/`edk2-ovmf` or `qemu-system-aarch64-core`/`edk2-aarch64`; other arches (riscv64, ppc64le, which `IncusArch()` maps) get no QEMU. pacman and apk branches assume their package names never change. | Table per PM × arch, or let `incus` pull its own deps where the distro does that (Debian does). |
 | 5.3 | **B** | `internal/incus/incus.go:958, 970` `DEBIAN_FRONTEND`, `apt-get … -o Dpkg::Options::=--force-confold`; `internal/updates/updates.go:78, 432` | Debian-isms exported unconditionally. Harmless elsewhere. | Fine; note only. |
 | 5.4 | **B** | `scripts/pci-ids-gpu.sh:7, 41` default `/usr/share/misc/pci.ids` | Debian's path for `pci.ids`; Fedora/Arch keep it in `/usr/share/hwdata/`. Build-time only. | Try both. |
 | 5.5 | **B** | `internal/api/status_sensors.go:283`, `internal/api/docker.go:133` | Container memory/CPU read from three cgroup paths: cgroup v2 + systemd driver, cgroupfs, libpod. cgroup v1 hosts (older Debian/Ubuntu, some Alpine, OpenWrt) get no `container.*.mem/cpu`. | Add the v1 `memory/docker/<id>` and `cpu,cpuacct` paths, or fall back to `docker stats --no-stream` once in a while. |
 | 5.6 | **B** | `internal/metrics/metrics.go:1064` `clkTck = 100` | USER_HZ assumed 100. True on every mainstream build; wrong on a kernel built otherwise. | `sysconf` via `getconf CLK_TCK` once at start. |
 | 5.7 | **B** | `internal/incus/subnet.go:150-179` `ip -j`; `internal/space/manager.go:578-585` `timeout -s KILL` | iproute2 JSON output (busybox `ip` has no `-j`); GNU `timeout`. | Parse plain `ip route`/`ip addr`; busybox has `timeout` too but flag spelling differs — use the context deadline. |
 ## 6. Hardware and vendor assumptions
 | # | Sev | Where | Assumes | General form |
 |---|---|---|---|---|
 | 6.1 | **B** | `internal/metrics/gpu.go:343-383` `read()` | Busy/clock/VRAM are read for **i915, xe, amdgpu** only — the three GPUs in this lab. Nouveau, nvidia-drm, panfrost, v3d, Asahi are listed but never measured. `PLAN.md:317`: "No NVIDIA path: nothing here has one to test against." | `nvidia-smi --query-gpu` where present; `/sys/class/drm/*/device/gpu_busy_percent` for nouveau; `devfreq` for the ARM drivers. |
 | 6.2 | **B** | `internal/metrics/pciids_gpu.txt`, `scripts/pci-ids-gpu.sh:12-15` | Embedded PCI ID table trimmed to Intel, AMD, NVIDIA and three virtual vendors. Matrox, ASPEED (every server BMC), Moore Threads, etc. fall back to `lspci` or "Vendor GPU <id>". | Include the full display-class device list, not three vendors. |
 | 6.3 | **B** | `internal/metrics/gpu.go:237-241` | "Integrated" = PCI bus `0000:00:`. Wrong on multi-domain or APU boards where the iGPU hangs off a bridge; cosmetic. | Use `boot_vga` + `PCI_SUBSYS` heuristics, or drop the flag. |
 | 6.4 | **B** | `internal/metrics/metrics.go:189-196, 917-968` | UPS probe order **pwrstat (CyberPower) first**, then NUT, then apcupsd; pwrstat's `....`-separated output is parsed bespoke. A host with both pwrstat and NUT shows pwrstat only. | Prefer NUT (the vendor-neutral one), then the vendor tools; or list every UPS found. |
 | 6.5 | **B** | `internal/metrics/metrics.go:1179-1182` `cpuTemp()` | CPU temperature from `coretemp`, `k10temp`, `thermal`/`cpu`, `cpu_thermal`, `soc_thermal`. Misses `zenpower`, `acpitz` as the only sensor, Apple `macsmc` names, many ARM SoC zone names. | Broaden; fall back to the hottest zone when none matches. |
 | 6.6 | **B** | `internal/hardware/hardware.go:617-627` `jedec` | JEDEC IDs for "the memory makers who turn up in a homelab" — 14 entries. | Embed the JEDEC JEP106 bank list (public, small). |
 | 6.7 | **B** | `internal/storage/storage.go`, `internal/api/zfs.go`, `internal/storage/arc.go`, `web/src/ZFS.tsx` | A whole **ZFS** tab and ZFS-aware Space/Disks logic; **no btrfs or LVM equivalent** (snapshots, subvolumes, scrub, balance). Not wrong, but the tab list reflects a ZFS NAS. | Later: a Btrfs tab with the same shape, or a generic "Filesystems" tab that shows what the host has. |
 | 6.8 | **B** | `internal/storage/snapshots.go:1-16, 966-968` | Snapshot naming is **zfSnap's** (`2026-09-24_02.30.01--6w`) because `grid` ran zfSnap. A sanoid/zfs-auto-snapshot user's existing snapshots (`autosnap_…`, `zfs-auto-snap_daily-…`) are detected (`internal/api/zfs.go:1764-1798`) but can never be expired by Ham Lab, and new schedules write a third convention beside them. | Make the naming a per-schedule choice (zfSnap / sanoid / zfs-auto-snapshot), or at least document the convention as the product's rather than inherited. |
 | 6.9 | **B** | `internal/network/speedtest.go:64-67` `downSources` | Fallback download source is **`speedtest.seattle.linode.com`** — chosen because the lab is in the Pacific Northwest. Everywhere else it reads low. | Pick the nearest of several mirrors, or drop the fallback and report the refusal. |
 | 6.10 | **B** | `internal/network/speedtest.go:70-74, 604`, `internal/network/network.go:323-334` | Public IP, edge and speed come from **Cloudflare** endpoints, polled every 15 min with no off switch. Air-gapped or privacy-minded installs have no way to disable the outbound probe. | A `network.outside: off` setting; honour it in the watcher. |
 | 6.11 | **B** | `internal/vpn/edit.go:5-7, 419-421` | Peers can be **created/edited only through OpenWrt UCI** "because that is where the server here is configured". A wg-quick server (the common case everywhere else) is read-only. | Write wg-quick `.conf` too (append a `[Peer]` block + `wg set`), guarded the same way. |
 | 6.12 | **B** | `internal/vpn/vpn.go:251-253` | Peer names come from wg-quick comments or UCI descriptions — "which between them cover every machine here". NetworkManager and systemd-networkd WireGuard configs are not read. | Add `.nmconnection` and `.netdev` readers. |
 | 6.13 | **B** | `internal/vpn/edit.go:515` `profile()` | Split-tunnel routes hard-code `10.0.0.0/8, 192.168.0.0/16, 172.16.0.0/12` on top of the tunnel's own prefix. | Offer the host's actual LAN prefixes (from `TakenPrefixes`-style route reading), with RFC 1918 as the fallback. |
 | 6.14 | **B** | `internal/incus/subnet.go:1-9, 23-34, 55`; `web/src/Virtualization.tsx:133` | Bridge subnet candidates skip **all of 10/8** because "this LAN is 10.0.0.0/8"; order is 172.16.100–254, then 192.168.254–200, then 172.16.1–99; last-resort `172.16.100.0/24`. Works anywhere, but a lab on 172.16/12 gets 192.168.x first and a lab on 192.168/16 never gets its own range. | Read the host's routes (already done) and pick from all three RFC 1918 blocks in a neutral order. |
 | 6.15 | **B** | `internal/network/network.go:170-199, 303-314`; `internal/incus/subnet.go`; `internal/vpn/edit.go:558-595`; `internal/tlsauto/tlsauto.go:533-537`; `install.sh:114-120`; `internal/setup/setup.go:648-666` | **IPv4 everywhere**: default route from `/proc/net/route` only, outside view v4 only, bridge/peer allocation v4 only, cert SANs v4 only, installer lists v4 addresses only. | `/proc/net/ipv6_route`, v6 SANs, v6 peer addresses. An IPv6-only host today has no gateway, no cert SAN and no bridge. |
 ## 7. Privilege and deployment posture (the lab's NOPASSWD-everywhere model)
 | # | Sev | Where | Assumes | General form |
 |---|---|---|---|---|
 | 7.1 | **A** | `internal/priv/priv.go`; every tab that calls `priv.Available()`; `web/src/Updates.tsx:2144` "give its user passwordless sudo or doas"; `web/src/Jobs.tsx:1883` | Full function needs the service user to have **unrestricted passwordless sudo**, which `hamlet/CLAUDE.md` grants on every host here. The installer neither creates nor offers a sudoers rule, so on a fresh box most tabs are read-only and the UI tells the user to go fix sudoers by hand. | `hamlab install --sudo` writes a **scoped** `/etc/sudoers.d/hamlab` naming the commands Ham Lab runs (smartctl, zpool, zfs, systemctl, docker, journalctl, tee for its cron files, `hamlab fs|space-walk|drm-clients|replace`), or installs as root. Either way the console works out of the box. |
 | 7.2 | **A** | `internal/api/server.go:1069-1080` `clientKey()`, `:1083-1092`, `:431-433` `secure()` | The **first `X-Forwarded-For` hop is trusted unconditionally**, because here Ham Lab always sits behind Nginx Proxy Manager. On a host exposed directly, a client sets the header and defeats the per-client login rate limiter and the session IP check. | Trust `X-Forwarded-*` only from configured proxy addresses (`trusted_proxies` in config), else use the socket peer. |
 | 7.3 | **B** | `internal/api/server.go:1324` `X-Accel-Buffering: no` "nginx (NPM)" | Harmless header; comment names the lab's proxy. | Keep header; generalise comment. |
 | 7.4 | **B** | `internal/api/guests.go:1-9`, `PLAN.md:21` | Guest logins are full-power ("he gives these to people he trusts"). A product default. | Keep, but it is a posture decision taken for this household; a read-only guest is the generic expectation and was cut. Document as a deliberate choice in README. |
 | 7.5 | **B** | `internal/api/users.go:1-6` | "a household console, not a tenancy": any signed-in user can add users, make tokens, open a root shell. | Same as 7.4 — fine, but say so where a stranger installs it. |
 ## 8. Resource tuning from the lab's smallest host
 | # | Sev | Where | Assumes | General form |
 |---|---|---|---|---|
 | 8.1 | **B** | `cmd/hamlab/main.go:451` `debug.SetMemoryLimit(160 << 20)`; `PLAN.md:11` "`core` has ~370 MiB free RAM" | A fixed 160 MiB soft heap limit sized for `core`. On a 64 GiB NAS it forces needless GC under a big Space scan; on a 512 MiB Pi it is still half the box. | Scale from `MemTotal` (e.g. 2–5 %, floor 64 MiB, ceiling 1 GiB), or make it a config key. |
 | 8.2 | **B** | `internal/space/walk.go:893` `workers := 32`; `internal/space/manager.go:578` 3-minute `zfs diff` patience | Tuned on `grid` (16 cores) and `core`. | Workers from `GOMAXPROCS`; patience configurable. |
 | 8.3 | **C** | `internal/metrics/metrics.go:32-43` intervals, `History = 300` | Reasonable generic defaults; listed only because 8.1 made them look lab-tuned. | None. |
 ## 9. Data directory and legacy paths
 | # | Sev | Where | Assumes | General form |
 |---|---|---|---|---|
 | 9.1 | **B** | `internal/config/config.go:115-117` `Candidates()`; `internal/setup/setup.go:413-422`; `README.md:38` | `~/apps/hamlab/config.json` is searched as a legacy location — the lab's pre-v0.2 path, kept so `core`, `grid` and the Mac keep working. | Keep for one more release with a one-time migration log line, then drop. A stranger never had it. |
 | 9.2 | **B** | 3.3 above; `PLAN.md:24` "`grid`'s and the Mac's stacks at `~/stacks/<machine>/` are the standing exception" | The "everything under DataDir" rule has a lab-shaped exception baked into the stacks code. | Default to DataDir; treat `~/stacks` as a discovered home like any other. |
 | 9.3 | **C** | `internal/setup/setup.go:351` `ownFiles` | Correctly scoped; note only. | None. |
 ## 10. Build and release
 | # | Sev | Where | Assumes | General form |
 |---|---|---|---|---|
 | 10.1 | **B** | `Makefile:60-61, 442-443`; `scripts/gitea-release.sh` | Publishing is Gitea/Forgejo-only (release API shape). `selfupdate` also understands a static directory, so publishing could too. | `make release` + any static host is already enough; add a GitHub/Forgejo-agnostic `release-dir` note. |
 | 10.2 | **C** | `scripts/gitea-release.sh:500-515` | Needs `python3` on the publishing host. | `jq`, or do it in Go (`hamlab publish`). |
 | 10.3 | **C** | `Makefile:392` `ARCHES ?= amd64 arm64 arm 386 riscv64` | Fine. `ppc64le`/`s390x` are mapped by `IncusArch()` but never built. | Add or drop them from `IncusArch`. |
 ## 11. Things that are already general (checked, no change needed)
 Listed so the sweep is known to be complete, not because they need work.
 - `internal/priv` (root / sudo / doas), `internal/setup` service templates, `IP_FREEBIND` listen, self-update atomic replace and re-exec.
 - `internal/updates`: apt, dnf, zypper, pacman, apk with refresh/upgrade/install.
 - `internal/storage` block devices: `lsblk -J` with a `/sys/block` fallback, eMMC health, NVMe namespace folding, SMART only when `smartctl` exists.
 - `internal/metrics` CPU/mem/hwmon/thermal/net/disk/fs from `/proc` and `/sys`; remote-FS and pseudo-FS lists; filesystem statfs with a timeout.
 - `internal/hardware`: every reader may come back empty with a note; device-tree fallbacks for non-SMBIOS boards.
 - `internal/incus` detection and the per-machine "ready / repair / installable / unavailable" model (modulo 4.2, 5.1, 5.2).
 - `internal/nodes`, `internal/tlsauto`: hub ↔ node pairing, pinned certs, proxy, push.
 - `internal/sensors`, alerts, widgets, embeds, Prometheus: machine-agnostic by design.
 - `internal/files`, `internal/term`, `internal/tasks`, `internal/auth`, `internal/store`.
 - `internal/api/kvm.go`: "nothing in it knows anything about any particular KVM" — verified.
 - `internal/compose` (modulo 3.5).
 - The React UI: machine-switcher model, hash routing, per-account prefs, themes.
 ---
 ## Suggested order of work
 1. **Posture and safety first (A):** 7.2 trusted proxies; 7.1 installer-written scoped sudoers; 3.1 drop the `:2375` probe (with 3.2 as its consequence).
 2. **"Any Linux" promises the README already makes (A):** 4.1 power, 4.4/4.5 cron on busybox and timer-only hosts, 4.2 Incus install on non-systemd, 4.3 a Services backend for OpenRC/procd, 2.10 the `tank` regex.
 3. **Lab defaults that phone home or mislead (B):** 1.1–1.3 update/install URLs, 2.9 schedule defaults, 6.9/6.10 speed-test sources and an off switch, 8.1 memory limit, 3.3/9.2 stack home = DataDir.
 4. **Breadth (B):** 5.1 zypper for Incus, 5.5 cgroup v1, 6.1/6.2 GPUs beyond Intel/AMD, 6.4 UPS order, 6.11/6.12 wg-quick/NetworkManager, 6.15 IPv6.
 5. **Cosmetics (C):** every placeholder and comment in §1.10 and §2, then split lab facts out of `PLAN.md`, `VIRTUALIZATION.md`, `SENSORS.md`, `README.md`.
@@ -54,6 +54,14 @@ reset a password; your own password confirms the last two). From the shell:
 users` lists them, `hamlab user-rm <user>` removes one. Machine settings — refresh rate,
 graph history, alert rules — stay the machine's and are shared by everyone.
 **Guest logins**, for showing somebody the console: Users → Guest logins makes a
 throwaway user name and password, good for half an hour by default (up to 24 hours), and
 hands you the pair to paste into a message. A guest is an ordinary account with a clock on
 it — the whole console, every machine, the terminal included — that stops working when the
 time is up, wherever it is signed in, or the moment you revoke it. There is no reduced
 guest: give one to somebody you would hand the password to, and revoke it if you change
 your mind.
 What adapts to the host, rather than assuming Debian:
 | | |
@@ -62,6 +70,8 @@ What adapts to the host, rather than assuming Debian:
 | Updates tab | apt, dnf, zypper, pacman, apk: list, refresh, upgrade, install (pacman: all or nothing) |
 | Services / timers | systemd of any age (JSON output where it exists, the plain table before v246); a clear message on hosts without systemd |
 | Disks | `lsblk`, or `/sys/block` when it's missing (busybox) or too old; SMART and ZFS only when their tools are there |
 | Network | `/proc` and `/sys` for the interfaces, addresses, link speed and default route. The speed test needs no tool installed: it measures against Cloudflare's public speed endpoints, so it works the same on Debian and on the OpenWrt router |
 | VPN | WireGuard tunnels and their peers, from `wg show` where wireguard-tools is installed; peer names from wg-quick `.conf` comments or OpenWrt UCI descriptions. Read-only, and it never reads a private or pre-shared key |
 | Dashboard | `/proc` and `/sys` only: hwmon or thermal-zone temps, fans, NICs, disks, mounts; empty sections where the hardware has nothing to report. UPS from whichever of `pwrstat`, NUT (`upsc`) or `apcupsd` is installed |
 | Listen address | bound with `IP_FREEBIND`, so a LAN address that isn't up yet at boot doesn't fail the start |
 | Self-update | the new build must run and load this host's config before it replaces the old one; atomic rename; the daemon then re-execs itself, so it restarts the same way under systemd, OpenRC or nothing; `hamlab.prev` kept |
@@ -16,9 +16,12 @@ cpu.usage  cpu.core.<n>.usage  cpu.temp  load.1|5|15  uptime
 mem.total|used|avail|free|cached|buffers|used_pct     swap.total|used|avail|used_pct
 temp.<chip>.<label>            fan.<chip>.<label>
 net.<iface>.rx|tx|rx_total|tx_total|up
 net.internet.download|upload|latency|jitter|tested|ip|edge
 disk.<dev>.read|write|util|size|temp
 fs.<mount>.size|used|used_pct
 ups.battery|runtime|load|load_pct|in_v|out_v|on_battery|state|source
 gpu.<name>.usage|render|video|video-enhance|copy|compute|temp|power|fan|freq|mem_used|mem_total|clients
 zfs.arc.size|target|min|max|mru|mfu|data|metadata|hit_pct|reads|misses|lifetime_hit_pct|no_grow|compress_ratio|l2_size|l2_hit_pct
 ```
 - Segments are lower case `[a-z0-9_-]`: `Package id 0` → `package_id_0`, `/mnt/tank` →
@@ -26,6 +29,24 @@ ups.battery|runtime|load|load_pct|in_v|out_v|on_battery|state|source
 - Built from names the hardware reports, never enumeration order (`hwmon3` moves between
  boots). Two identical chips collide; the second gets `_2`.
 - Which sensors exist differs per machine. That is the point of the catalog.
 - `net.internet.*` is the Network tab's speed test: `download` and `upload` in Mbps,
  `latency` and `jitter` in ms, `tested` the age of the reading in seconds, and `ip` and
  `edge` the public address and the Cloudflare edge as strings. They appear only once a test
  has run on that machine, and each one's `t` is when it ran — so a widget or an alert on
  them is reading a measurement from that moment, not a live value. Nothing takes a speed
  test on its own; something has to press the button.
 - `gpu.<name>.*` is each graphics device, named as lspci names it (`gpu.intel_arc_a380`;
  labels carry `driver` and the PCI `bus`). `usage` is how busy the device was — from the
  GT's idle residency on Intel, `gpu_busy_percent` on amdgpu — and is absent on a driver
  with no such counter. The per-engine percentages (`render`, `video`…) are the sum over the
  processes holding the device open, read from DRM fdinfo, and exist only while something
  has it open. `temp` is the card's own sensor, or on an integrated GPU the CPU package it
  shares a die with (label `source` says which); an Arc on i915 has none and reports none.
  `freq` is the actual clock, 0 when power-gated. `clients` counts the processes using it.
 - `zfs.arc.*` is the ZFS read cache in RAM, on a machine with the module loaded. `hit_pct`
  is over the last sampling interval and appears once an interval has had any reads (an idle
  interval repeats the last value); `lifetime_hit_pct` is since boot. `no_grow` is true
  while memory pressure stops the cache growing. `l2_*` appear only with cache devices.
 ## A sensor
@@ -0,0 +1,347 @@
 # Virtualization tab — plan (2026-10-09; phase 1 built the same day)
 **Status:** phase 1 is written, tested and committed as **v0.1.0-180** (`internal/incus`,
 `internal/api/virt.go`, `Virtualization.tsx`), **not yet deployed** — `make deploy` was
 refused by the session's auto-mode classifier, so it is his to run. Nothing is installed on
 any machine yet; Install has not been exercised against a live host. First real test is
 Install on `grid` (zfs pool on a file, Docker present), then `core` (btrfs loop, Docker
 present), then the Mac mini (dnf, arm64). Not built yet from the phase-1 list: the
 `instance.*` sensors and the Alerts default — next session.
 A tab for creating, watching, consoling into and deleting **LXC containers and KVM virtual
 machines**, the way Proxmox does with `pct` and `qm` — **per machine, like every other
 page.** This is the scoping PLAN.md asked for. Nothing here is built. Live state of the
 hosts stays in `hamlet/docs/hosts.md`.
 ## The governing rule: capability, not policy
 Stated by Mitchell 2026-10-09, after the first draft of this plan excluded hosts on the
 strength of the homelab's deployment rules:
 > Anything that's done when a different machine is selected SHOULD GO ON THAT machine. I
 > want the option to do it anywhere I want, unless the hardware / software literally
 > doesn't support it. Ham Lab needs to be built in a general-purpose way that can run on
 > any Linux system. If certain parts / libraries aren't available on that machine, those
 > tabs should just report that.
 So:
 - **Every action lands on the machine in the switcher.** There is no "VM host", no default
  machine, no cross-machine placement. `grid` is where he will most likely click first;
  the tab does not know or care.
 - **The tab is gated by what the machine can do, detected live** — never by a rule about
  what *should* run there. The `hamlet/CLAUDE.md` do-not-break rules (Mac mini for
  single-core workloads, `hamtv` not spare hardware) bind the *agent's* unasked
  deployments. They are not feature gates for the product, and the first draft was wrong
  to treat them as such.
 - **Ham Lab owns its data, and its data lives in its own folder.** Everything this tab
  manages — instance storage, ISOs, image tarballs, exports — sits under Ham Lab's data
  directory on *that* machine (`config.DataDir()`, `~/apps/hamlab` on `core`, `grid` and
  the Mac), in one `virt/` tree. Never a path elsewhere on the machine, however convenient
  the existing folder. Stated 2026-10-09 after the first Library draft pointed the ISO shelf
  at `/tank/software/isos/`: *"ALL data needs to be under a HamLab folder, because it
  manages its own data."* This is the rule that moved `core`'s compose files under
  `apps/hamlab/stacks/` on 2026-10-05; `grid`'s and the Mac's stacks, still at
  `~/stacks/<machine>/`, are the outstanding exception and are not touched by this plan.
  The engine's own runtime state (`/var/lib/incus`, like `/var/lib/docker`) is the engine's,
  not Ham Lab's data: Ham Lab keeps what it would need to rebuild it.
 - **Ham Lab installs or reports.** It detects what is missing, offers to install it where it
  knows how, and where it cannot or the hardware cannot do the job, says so in one sentence
  and offers what *does* work (containers without KVM, for instance).
 ## Decision: Incus 6.0 LTS as the engine
 | Option | Verdict |
 |---|---|
 | **Incus** (LXD fork, linuxcontainers.org) | **Chosen.** One daemon, one CLI and one JSON API for both containers and VMs. Packaged by Debian 13 (`incus` 6.0.4 LTS — `apt install incus` pulls `qemu-system-x86`, `swtpm`, `virtiofsd`, `incus-agent`, so VMs work out of the box), Fedora (6.23), Arch, Alpine, Ubuntu (Zabbly repo); not by OpenWrt. Storage drivers for ZFS, btrfs, LVM and plain directories, so it fits a machine with or without ZFS. Built-in image server (`images:debian/13`, `images:ubuntu/24.04`, Alpine, Fedora, Arch… as container *or* VM, per architecture), snapshot schedules with expiry, `incus export` backups, per-instance CPU/memory limits, cloud-init, text and VGA console over websocket. Everything is reachable as `incus … --format json` or `incus query`, which is exactly how Ham Lab already drives Docker: `exec.Command`, no new Go dependency, no cgo. |
 | libvirt/KVM + `virsh`, plus LXC separately | Two daemons, two models, XML. libvirt's official Go binding needs cgo, which the static binary rules out (the pure-Go `go-libvirt` exists but then containers are a second project). Rejected. |
 | Proxmox | An OS, not a library. Rejected. |
 ## What each machine can do today
 Probed 2026-10-09 (`/dev/kvm`, cpu flags, kvm modules, RAM, disk, package availability).
 These are **facts the tab will detect at runtime**, not a list of where instances are
 allowed. The last column is what the tab would show the day it ships.
 | Machine | Arch | KVM | RAM (free) | Incus from the package manager | Pool driver available | Tab state on first open |
 |---|---|---|---|---|---|---|
 | `grid` | x86-64 | yes — VT-x, `kvm_intel` loaded, nested on | 62 GiB (~25; ARC holds 23 of the rest, max 40) | apt: `incus` 6.0.4 | **zfs** (2.3.9, `tank`) | **Install** — containers and VMs |
 | `core` | x86-64 | yes — N100, `kvm_intel` loaded | 7.5 GiB (~4) | apt: `incus` 6.0.4 | btrfs or dir (ZFS was purged 2026-09-23) | **Install** — containers and VMs; the create form shows free memory so a VM's size is an informed choice |
 | Mac mini | aarch64 | yes — Asahi KVM, 16K pages | 15 GiB (~8) | dnf: `incus` 6.23, `qemu-system-aarch64-core`, `edk2-aarch64` | btrfs or dir | **Install** — containers and VMs, **arm64 images only** (the picker filters by host arch). Whether a stock 4K-page guest kernel boots under the 16K host is untested; `muvm` does exactly that, so it is expected to. |
 | `hamtv` | x86-64 | yes | 3.7 GiB (~2.9) | apt: `incus` 6.0.4 | btrfs or dir | **Install** — containers and small VMs |
 | `gate` | x86-64 | **no** `/dev/kvm` | 8 GiB (~6) | **none** — OpenWrt's opkg has no Incus | — | **Unavailable**: "Incus isn't packaged for OpenWrt, and this kernel has no KVM." Nothing to install. |
 Nothing is installed anywhere today: no qemu, libvirt, lxc or incus on any host.
 ### Detection the tab runs, in order
 1. `incus` on `PATH` and its socket answering → **ready**; read pools, networks, instances.
 2. Else: a package manager Ham Lab knows (apt, dnf, pacman, apk) that has an `incus`
   candidate → **installable**, with the Install dialog pre-filled (below).
 3. Else → **unavailable**, with the reason in one sentence (no package; no root; OpenWrt).
 4. Independently: `/dev/kvm` present and openable → VMs offered. Absent → **containers
   only**, said once on the page, and the create form has no VM option.
 5. Architecture from `uname -m`, so the image picker only shows images that can run here.
 6. Pool driver: `zfs` if the module and tools are present, else `btrfs` (installing
   `btrfs-progs` as part of Install), else `dir`. The dialog says which it picked and why.
 ## What the tab does (v1)
 Mirrors the Docker tab's shape: a list, a detail page, actions that are followed tasks,
 destructive ones confirmed by typing the name.
 - **Install** — one task that runs *on the selected machine*: install the package(s),
  `incus admin init --preseed` with the layout from the dialog, create the pool and bridge,
  add Ham Lab's user to `incus-admin` (socket access, the `docker`-group shape) and add it
  to the unit's `SupplementaryGroups`, apply the Docker fix where Docker is present (below),
  then restart Ham Lab. Idempotent, so it doubles as "repair". The dialog shows and lets him
  edit: pool driver (detected), pool size (proposed from free space), bridge subnet
  (computed), whether to apply the Docker rules. Where it all goes is not a choice: Ham
  Lab's folder.
 - **Instances** — containers and VMs in one list: type, state, IPv4, CPU and memory live,
  disk, snapshot count, uptime. Click for detail.
 - **Create** — container or VM (VM only where KVM is); image from the `images:` remote,
  filtered to this architecture, searchable, with the common ones pinned (Debian 13, Ubuntu
  24.04, Alpine, Fedora); name, CPUs, memory (with this machine's free memory beside it),
  root disk size; network (the NAT bridge, or a LAN bridge if this machine has one);
  optional cloud-init user-data (SSH key, packages, a first command); profile. Create, or
  create and start. Collisions (name taken) named before Create.
 - **Lifecycle** — start, stop, restart, force stop, rename, delete (type the name; the
  instance's snapshots go with it, and the dialog says so).
 - **Console** — two buttons, both open as tabs in the Terminal page through the existing PTY
  manager (`term.CreateCmd`, the `docker exec` path): **Console** (`incus console`, the
  serial console — works before the network is up, shows boot output) and **Shell**
  (`incus exec <n> -- login shell`). Containers and, via `incus-agent`, VMs.
 - **Snapshots** — take, restore (confirmed), delete; a **schedule** per instance with an
  expiry, which is Incus's own `snapshots.schedule`/`snapshots.expiry`, so no cron file of
  ours this time.
 - **Resources** — edit CPU and memory limits (live for containers; VMs take CPU hotplug and
  memory balloon live with the agent, else on restart), grow the root disk, add a bind mount
  (a host path into a container; virtiofs into a VM).
 - **Library** — ISOs for VMs and images (container templates and VM images). Its own
  section below.
 - **Storage and networks** — the pools with their fill, the bridges with their subnets.
  Read-only in v1 beyond what Create and Install need.
 - **Sensors** — `instance.<name>.cpu`, `.memory`, `.running` from Incus's state, like the
  `container.*` sensors, so Alerts gets an "instance down" default and widgets can show one.
 ## Where the data lives
 One tree per machine, under Ham Lab's data directory, created by Install:
 ```
 ~/apps/hamlab/virt/
  pool.img          the instance pool's backing file (zfs or btrfs loop; a plain directory
                    pool/ when the driver is dir) — every instance's disks live in here
  isos/             the ISO library: the files themselves
  images/           the image library: template tarballs and VM disk images as files, the
                    way Proxmox keeps CT templates — Incus's image store is loaded from here
  exports/          nightly `incus export` tarballs, with retention
  virt.json         what Ham Lab chose and when: driver, pool size, bridge subnet, ISO and
                    image inventory, export schedule
 ```
 Consequences, stated once:
 - The tree is on whatever disk holds Ham Lab's data directory — the boot NVMe on all three.
  Install proposes the pool size from that disk's free space. On `grid` that disk has
  790 GiB free; on `core`, 175 GiB.
 - Bulk data a guest needs is **bind-mounted in from where it already is** (`/tank/…` into an
  instance on `grid`), the same split Docker has: the instance is Ham Lab's, the media is
  not. That is a device on the instance, recorded in its config, not a second home.
 - On `core` the whole tree rides the existing nightly rsync to `grid` for free. On `grid`
  and the Mac it is as backed up as the rest of their Ham Lab folders — which is to say the
  exports are the backup, and the backup job will need to be told about them in phase 2.
 - Existing material elsewhere — the 105 GiB of ISOs at `/tank/software/isos/`, the Proxmox
  archives — is **imported into the tree by copying**, file by file, on request. The
  library never indexes a folder it does not own.
 ## Library: ISOs and images (asked for 2026-10-09)
 Proxmox has a storage view with **ISO Images** and **CT Templates**. Incus has the same two
 things under other names, and the tab gives them one home, a **Library** sub-tab, per machine
 like everything else.
 | Proxmox calls it | Incus calls it | What it is here |
 |---|---|---|
 | ISO Images | an ISO file attached to a VM as a `disk` device (added as a ROM device; VMs only), or a custom storage volume of type `iso` | **`virt/isos/` in Ham Lab's folder**, attached by path — no copy into the pool |
 | CT Templates | **images** — the `images:` remote (linuxcontainers.org), the local image store, anything imported or published | **`virt/images/` in Ham Lab's folder** — tarballs and disk images as files, loaded into Incus from there; the remote is the catalogue to pull from |
 ### ISOs
 - **The ISO library is `virt/isos/` in Ham Lab's folder**, on every machine, full stop. The
  105 GiB at `/tank/software/isos/` on `grid` (`linux/`, `windows/`, `truenas/`,
  `virtio_drivers/`…) is a *source* to import from — pick the ones wanted, Ham Lab copies them
  in — not the library itself.
 - **Getting an ISO in:** upload from the browser (the Files tab's `PUT /api/files/upload`
  with its progress bar, reused), **download from a URL** as a followed task with an optional
  checksum (the usual way — an ISO starts as a link), or **import from a path** already on the
  machine — which **copies** the file into `virt/isos/` and says so. Rename, delete (refused while a VM has it attached, naming the VM).
 - **Using one:** attach to / detach from a VM as a ROM device by path, with `boot.priority`
  set so it boots first; **New VM from ISO** makes an empty VM (`incus init --empty --vm`)
  with the ISO attached, starts it and opens the console. After install, Detach is one click
  and boot order falls back to the disk.
 - **Honest limit:** an ISO install is an interactive graphical installer, so it needs the
  **VGA console in the browser** (`spice-html5` over Incus's VGA console websocket).
  Installers that run on a serial console (Debian with `console=ttyS0`, Alpine, Arch) work
  through the text console before that exists. Windows needs the virtio drivers ISO attached
  as a second ROM device as well — Incus's documented path is `distrobuilder repack-windows`
  to bake them in; the drivers are already in `virtio_drivers/` on `tank`. So **the ISO shelf
  and the VGA console ship together, in phase 2**, moved up from phase 3.
 - ISO *volumes* (`incus storage volume import <pool> <iso> <name> --type=iso`) are the
  alternative: read-only, attachable to several VMs at once, but they copy the file into the
  pool. Not used here — a file in `virt/isos/` is already Ham Lab's, and a volume would copy it
  a second time into the pool. Offered as an option only if a need shows up.
 ### Images (container templates and VM images)
 - **The catalogue:** browse `images:` filtered to this machine's architecture, by type
  (container / VM), with the common ones pinned. **Download** copies one into the local
  store (`incus image copy images:debian/13 local: --alias debian-13`) so a later Create is
  instant and works without the internet. Images used by Create are cached by Incus anyway;
  this is just the explicit version, Proxmox's "Templates → Download".
 - **The local store:** every image here with alias, type, arch, size, origin (remote,
  published, imported), when it was last used, and **which instances were made from it**
  (Incus records the base image on each instance). Alias, delete (refused while an
  instance's base — Incus knows), "New instance from this image".
 - **Import:** a unified tarball or metadata + rootfs pair (a container template from
  anywhere, including the phase-3 Proxmox conversions), or a **qcow2 / raw disk with
  metadata Ham Lab writes** for a VM image — Home Assistant OS, OPNsense, a vendor cloud
  image — by upload, URL, or a path on the machine. Same three ways in as ISOs.
 - **Publish:** a stopped instance becomes an image with an alias — Proxmox's "convert to
  template", the one thing the Images section of v1 always had.
 - **Where they live:** as files in `virt/images/` — the downloaded or imported tarball, the
  published image exported to one (`incus image export`) — so the library is complete without
  Incus and can be reloaded into a fresh install. Incus's own image store under
  `/var/lib/incus/images/` is a cache derived from it, and the page says so.
 ## Not in v1, with reasons
 | Thing | Why not yet |
 |---|---|
 | **VGA console in the browser** | Incus hands the VM's SPICE socket over a websocket; drawing it needs `spice-html5` in the UI. Doable (the Incus UI does it); text console + shell cover every Linux guest until then. **Phase 2, with the ISO shelf** — an ISO install is useless without it. |
 | **Windows VMs** | Incus runs them (ISO repacked with virtio drivers by `distrobuilder`), but without a VGA console in the browser they are unusable from the tab. After the above. |
 | **GPU passthrough to a VM** | Removes the card from the host and from whatever uses it there (Jellyfin on `grid`). Offered only with that warning, later. Sharing a GPU into a *container* (`gpu` device) leaves the host its card and is how `r720-jellyfin` used to run. Phase 3. |
 | **Clustering / live migration** | Incus clusters, but these machines differ in size and one is arm64, with no shared storage. Not planned. |
 | **A fleet-wide instance list** | One machine per view. The switcher is the machine. |
 | **A second engine (bare LXC on OpenWrt, libvirt-only hosts)** | Would double the backend for one router that has no KVM. The tab reports Incus unavailable there and stops. |
 ## Layout decisions Ham Lab makes per machine (overridable in the Install dialog)
 1. **Storage.** The instance pool's backing lives in Ham Lab's folder: Install creates a
   sparse `virt/pool.img` and hands it to Incus as the pool's `source` (zfs or btrfs by the
   detection above), or `virt/pool/` as a `dir` pool where neither is available. Sized from
   the disk's free space, shown and editable. Bulk data is **bind-mounted in**, not stored in
   the pool. The driver, path and size are recorded in `virt.json` so the page can say where
   the data is and Install can rebuild the pool from the file after a reinstall.
 2. **Bridge subnet, computed.** Incus's `auto` picks a random `10.x.y.0/24`, and **this
   LAN is `10.0.0.0/8`** — every host has a /8 on its NIC, so any 10.x bridge collides.
   Ham Lab reads the machine's routes and picks the first free /24 from `192.168.0.0/16`,
   then `172.16.0.0/12`, that is neither on-link nor routed nor inside Docker's configured
   pool (`172.28.0.0/14` on `grid`; `172.19`–`172.29` and `192.168.80`–`143` live on
   `core`). Instances get `name.incus` DNS from Incus's dnsmasq.
 3. **LAN addresses are a later, explicit step.** A `br0` on the machine's uplink lets an
   instance take a lease from `gate` and a Pi-hole name, but it rewrites the host's network
   config and drops it off the network for seconds — never part of Install, always its own
   confirmed action (with the NanoKVM standing by on `grid`). macvlan was considered and
   rejected: no host change, but host and guests cannot talk, so nothing on the host could
   reach an instance.
 4. **Docker and Incus on one host.** Docker sets the `FORWARD` policy to `DROP`, which
   silently kills the Incus bridge's traffic. Fix is two `DOCKER-USER` accept rules for the
   bridge, persisted — applied by Install wherever Docker is detected (`grid`, `core`, the
   Mac, `hamtv` all run it).
 5. **Memory.** A VM's memory is a hard reservation; the create form shows the machine's
   free memory and warns past it, and never refuses. On a ZFS host Install offers to cap
   the ARC (on `grid`, 40 GiB → 16 GiB) so guest RAM is deterministic; his choice.
 6. **Privilege.** Ham Lab's user joins `incus-admin`; no sudo in the hot path. Install
   itself goes through `priv` like the other installers.
 7. **Backups.** Nightly `incus export --instance-only` of every instance to `virt/exports/`
   with a retention, as a cron line like the ZFS schedules (`hamlab incus-export`), from
   phase 2. On `core` that folder is inside the nightly rsync already; on `grid` and the Mac
   the backup job gets a line for it in phase 2.
 ## The old Proxmox estate (importable, phase 3)
 `/tank/backups/proxmox/` holds the retired `r720` node: `/etc/pve` (12 LXC confs, 2 VM
 confs), 118 GiB of `vzdump` archives, and three raw VM disks. Latest of each:
 | ID | Name | Latest archive | Size |
 |---|---|---|---|
 | 100 | r720-docker | 2026-01-11 | 20.7 GB |
 | 101 | r720-cockpit | 2026-01-11 | 439 MB |
 | 102 | r720-wordpress-alexblaskovich | 2026-01-11 | 1.2 GB |
 | 103 | r720-spoonterm | 2025-06-14 | 494 MB |
 | 104 | r720-ollama | 2026-01-11 | 11.3 GB |
 | 105 | r720-transmission | 2026-01-11 | 322 MB |
 | 106 | r720-apache | 2026-01-11 | 347 MB |
 | 107 | r720-nut | 2025-03-29 | 205 MB |
 | 108 | r720-filebrowser | 2026-01-11 | 262 MB |
 | 109 | r720-bookstack | 2025-04-27 | 516 MB |
 | 110 | r720-jellyfin | 2026-01-11 | 2.2 GB |
 | 111 | r720-wireguard | 2025-04-13 | 510 MB |
 | 112 | r720-eggwiki | 2026-01-11 | 634 MB |
 | VMs | steam (102), powerpanel (105/107/112) | `.vma.zst` 2024–2025; raw disks `vm-105`, `vm-112`, `base-104` | 19 GB, 1.9 GB ×6, 97 GiB |
 - **LXC:** feasible and scriptable. Unpack the rootfs (`tar --zstd -xpf … --numeric-owner`),
  read `etc/vzdump/pct.conf` for arch/ostype, drop it, write `metadata.yaml`, pack, `incus
  image import`, launch. Unprivileged uid shifting is Incus's job after that. The result is
  an image in the Library, so a restored container is launched from there like any other
  template. An import action in the tab (pick an archive on this machine) is plausible once
  the script works by hand. x86-64 archives only run on x86-64 machines; the picker says so.
 - **VMs:** the raw disks import directly (`incus-migrate`, or an empty VM with the disk
  imported into its root volume). The `.vma.zst` archives need Proxmox's `vma` extractor (or
  the Python port) to become raw disks first. Most of these are the same `powerpanel` box
  backed up six times, and `tank-space-audit.md` already has them down for deletion — decide
  which, if any, are wanted before anything is restored.
 ## Build plan
 | Phase | What | Effort |
 |---|---|---|
 | **1 — detection, Install, the tab** | `internal/incus` (detection, CLI wrapper, JSON types, `incus monitor` for live state, the preseed builder and subnet picker), `internal/api/virt.go` (routes below), `Virtualization.tsx` (the three page states; list, detail, create dialog, snapshot panel), rail entry `virt`, terminal integration, sensors, and the Library's **Images** shelf (browse and download from `images:`, local store, import, publish). **Tested by clicking Install on `grid` first, then `core`** — two different pool drivers, both with Docker present. Then the Mac for the arm64 path. `gate` for the unavailable state. | 2–3 sessions |
 | **2 — operations** | The Library's **ISO** shelf (directory, upload/URL/path, attach, New VM from ISO) **with the VGA console** (`spice-html5`), snapshot schedules UI, nightly exports with retention and the per-machine export directory, limits editor, bind mounts, the LAN-bridge step, instance alert default, cloud-init presets. | 1–2 sessions |
 | **3 — the rest** | Proxmox import (archives become Library images), GPU device for containers, Windows via `distrobuilder repack-windows`, GPU passthrough to a VM with its warning. | as wanted |
 ### API sketch
 ```
 GET    /api/virt                               {state: ready|installable|unavailable, reason, kvm, arch,
                                                pool_drivers[], pools[], networks[], instances[], free_memory}
 POST   /api/virt/install                       {driver, size, path, subnet, docker_rules, arc_max?}  task
 GET    /api/virt/images?remote=images&q=       local cache + remote search, filtered to this arch
 DELETE /api/virt/images/{fingerprint}
 POST   /api/virt/instances                     {name, type, image, cpu, memory, disk, network, cloudinit, profile, start}  task
 GET    /api/virt/instances/{name}              config, state, devices, snapshots
 PATCH  /api/virt/instances/{name}              limits, description, devices
 POST   /api/virt/instances/{name}/{start|stop|restart|kill|delete|rename|publish|export}
 POST   /api/virt/instances/{name}/console      → terminal {id}   (incus console)
 POST   /api/virt/instances/{name}/exec         → terminal {id}   (incus exec -- sh -lc …)
 POST   /api/virt/instances/{name}/snapshots    {name}; POST …/{snap}/restore; DELETE …/{snap}
 PUT    /api/virt/instances/{name}/schedule     {cron, expiry}
 POST   /api/virt/import                        {path}  task (export tarball, or phase-3 Proxmox archive)
 GET    /api/virt/isos                          virt/isos/, scanned: name, size, attached-to
 POST   /api/virt/isos/import                   {path}  task: copy a file on this machine into virt/isos/
 POST   /api/virt/isos/fetch                    {url, sha256?}  task;  uploads reuse PUT /api/files/upload
 DELETE /api/virt/isos/{name}                   refused while attached
 POST   /api/virt/instances/{name}/iso          {iso, boot_first}  attach;  DELETE …/iso  detach
 POST   /api/virt/instances/from-iso            {name, iso, cpu, memory, disk, network}  task → console
 GET    /api/virt/images/remote?q=&type=        the `images:` catalogue for this arch
 POST   /api/virt/images/download               {alias}  task (incus image copy images:… local:)
 POST   /api/virt/images/import                 {path | url, alias, type}  task: copy into virt/images/, then load into Incus
 GET    /api/virt/instances/{name}/vga          websocket: Incus VGA console → spice-html5  (phase 2)
 ```
 Everything is per machine and reaches a node through the existing `/api/nodes/{name}/…`
 proxy with no hub-side code; consoles attach over the existing `/api/terminals/{id}/attach`.
 Tasks stream through `/api/tasks/{id}/stream` as Docker's do.
 ## Sources
 - Incus on Debian: https://wiki.debian.org/Incus
 - Incus and Docker on one host (the `FORWARD` drop): https://linuxcontainers.org/incus/docs/main/howto/network_bridge_firewalld/
 - Bridge options and `ipv4.address: auto`: https://linuxcontainers.org/incus/docs/main/reference/network_bridge/
 - Consoles (text and VGA): https://linuxcontainers.org/incus/docs/main/howto/instances_console/
 - ZFS storage driver: https://linuxcontainers.org/incus/docs/main/reference/storage_zfs/
 - macvlan's host-isolation: https://linuxcontainers.org/incus/docs/main/reference/network_macvlan/
 - Image format for the Proxmox import: https://linuxcontainers.org/incus/docs/main/reference/image_format/
@@ -32,8 +32,8 @@ import (
 	"hamlab/internal/files"
 	"hamlab/internal/metrics"
 	"hamlab/internal/nodes"
 	"hamlab/internal/services"
 	"hamlab/internal/priv"
 	"hamlab/internal/services"
 	"hamlab/internal/setup"
 	"hamlab/internal/space"
 	"hamlab/internal/storage"
@@ -62,6 +62,7 @@ func main() {
  passwd [user]        set a user's password (makes the account if it is new; reads stdin when piped)
  users                list the accounts
  zfs-snapshot <id>    run one ZFS snapshot schedule (what its cron line calls)
  docker-update <id>   run one image-update schedule (what its cron line calls)
  user-rm <user>       remove an account and sign it out everywhere
  check                verify the config and state load, then exit
  version
@@ -81,6 +82,8 @@ func main() {
 		os.Exit(files.Helper(flag.Args()[1:], os.Stdin, os.Stdout, os.Stderr))
 	case "space-walk": // helper for the Space tab: measure mountpoints, stream the tree back
 		os.Exit(space.Helper(flag.Args()[1:], os.Stdin, os.Stdout, os.Stderr))
 	case "drm-clients": // privileged helper for the GPU readings: who has the GPU open (run as root by the daemon)
 		os.Exit(metrics.DRMClientsHelper(os.Stdin, os.Stdout))
 	case "install", "uninstall":
 		var err error
 		if cmd == "install" {
@@ -132,6 +135,10 @@ func main() {
 		if err := passwd(cfg, flag.Arg(1)); err != nil {
 			log.Fatal(err)
 		}
 	case "docker-update":
 		if err := dockerUpdateJob(cfg, flag.Arg(1)); err != nil {
 			log.Fatal(err)
 		}
 	case "zfs-snapshot":
 		if err := zfsSnapshot(cfg, flag.Arg(1)); err != nil {
 			log.Fatal(err)
@@ -146,7 +153,14 @@ func main() {
 			if !u.LastSeen.IsZero() {
 				seen = u.LastSeen.Local().Format("2006-01-02 15:04")
 			}
-			fmt.Printf("%-20s created %s  last seen %s\n", u.Name, u.Created.Local().Format("2006-01-02"), seen)
+			note := ""
 			if u.Guest { // a temporary look-only login, not one of the machine's accounts
 				note = "  guest, until " + u.Expires.Local().Format("15:04")
 				if u.Expired(time.Now()) {
 					note = "  guest, expired"
 				}
 			}
 			fmt.Printf("%-20s created %s  last seen %s%s\n", u.Name, u.Created.Local().Format("2006-01-02"), seen, note)
 		}
 	case "user-rm":
 		if err := userRm(cfg, flag.Arg(1)); err != nil {
@@ -185,6 +199,48 @@ func main() {
 // migrateUsers gives a pre-users config its first account, named after this OS user, so
 // the CLI sees the same accounts the daemon will.
 // dockerUpdateJob pulls a stack's images, and deploys them when the schedule says to. This
 // is what the cron line Ham Lab writes actually calls, so image updates keep happening
 // whether or not the console is up.
 func dockerUpdateJob(cfg *config.Config, id string) error {
 	if strings.TrimSpace(id) == "" {
 		return errors.New("usage: hamlab docker-update <schedule-id>")
 	}
 	st, err := store.Open(filepath.Join(cfg.DataDir(), "state.json"))
 	if err != nil {
 		return err
 	}
 	var raw json.RawMessage
 	st.View(func(s *store.State) { raw = s.DockerJobs })
 	var d struct {
 		Jobs []api.UpdateJob `json:"jobs"`
 	}
 	if len(raw) > 0 {
 		if err := json.Unmarshal(raw, &d); err != nil {
 			return fmt.Errorf("reading the schedules: %w", err)
 		}
 	}
 	for _, j := range d.Jobs {
 		if j.ID != id {
 			continue
 		}
 		// A pull that never finishes must not leave a cron job resident until the next reboot.
 		ctx, cancel := context.WithTimeout(context.Background(), 2*time.Hour)
 		defer cancel()
 		out, runErr := api.RunUpdate(ctx, st, j, api.DockerTool())
 		rec := storage.RunRecord{At: time.Now(), Taken: j.Stack}
 		if runErr != nil {
 			rec.Error = runErr.Error()
 		}
 		if err := storage.WriteRun(cfg.DataDir(), "docker-"+id, rec); err != nil {
 			fmt.Fprintln(os.Stderr, "recording the run:", err)
 		}
 		fmt.Print(out)
 		return runErr
 	}
 	return fmt.Errorf("no image-update schedule %q — it may have been removed; this cron line is stale", id)
 }
 // zfsSnapshot runs one snapshot schedule: take it, then expire what has aged out. This is
 // what the cron line Ham Lab writes actually calls, so a schedule keeps running whether or
 // not the console is up — and keeps running if Ham Lab is never opened again.
@@ -3,8 +3,8 @@ package api
 // The Docker tab: what is running, grouped the way it was deployed (compose stacks), with
 // start / stop / restart, logs and inspect — plus images, volumes and networks to look at.
 // It drives the docker CLI (or podman), as the user first and as root if that is refused.
-// Creating, removing and pruning are deliberately absent: whatever deployed the stacks
+// What can be removed here is what nothing uses: an untagged image, a volume no container
-// (Dockhand, compose) stays the one place that changes what exists.
+// holds, a network nothing is attached to. Compose stays the one place a stack is defined.
 import (
 	"context"
@@ -49,6 +49,10 @@ func dockerTool() string {
 }
 // dockerRun runs the tool as this user, and again as root if the socket said no.
 // DockerTool is whichever of docker or podman this machine has, for callers outside the
 // package — the scheduled updates the CLI runs.
 func DockerTool() string { return dockerTool() }
 func dockerRun(ctx context.Context, args ...string) ([]byte, error) {
 	tool := dockerTool()
 	if tool == "" {
@@ -72,6 +76,18 @@ func dockerRun(ctx context.Context, args ...string) ([]byte, error) {
 	return out, nil
 }
 // dockerWhy picks the line worth showing out of a failed command. Some subcommands —
 // `network rm` among them — print the real reason first and then an aggregated "exit
 // status 1", which is what dockerRun's last-line rule would otherwise hand back.
 func dockerWhy(out []byte, err error) string {
 	for _, l := range strings.Split(string(out), "\n") {
 		if l = strings.TrimSpace(l); strings.HasPrefix(l, "Error") || strings.Contains(l, "daemon") {
 			return l
 		}
 	}
 	return err.Error()
 }
 func jsonLines(out []byte, each func(map[string]any)) {
 	for _, line := range strings.Split(strings.TrimSpace(string(out)), "\n") {
 		var m map[string]any
@@ -135,23 +151,30 @@ func (s *Server) dockerOverview(w http.ResponseWriter, r *http.Request) {
 		}
 		return rows
 	}
-	// Who uses what, from one inspect of every container: an image or a volume nothing uses is
+	// Who uses what, from one inspect of every container: an image, a volume or a network
-	// the only kind this page offers to remove.
+	// nothing uses is the only kind this page offers to remove.
-	imageUsed, volumeUsers := map[string]bool{}, map[string][]string{}
+	imageUsed := map[string]bool{}
 	volumeUsers, networkUsers := map[string][]string{}, map[string][]string{}
 	if ids, err := dockerRun(ctx, "ps", "-aq"); err == nil && len(strings.Fields(string(ids))) > 0 {
-		const f = `{{.Name}}|{{.Image}}|{{range .Mounts}}{{if eq .Type "volume"}}{{.Name}},{{end}}{{end}}`
+		const f = `{{.Name}}|{{.Image}}|{{range .Mounts}}{{if eq .Type "volume"}}{{.Name}},{{end}}{{end}}|{{range $k, $v := .NetworkSettings.Networks}}{{$k}},{{end}}`
 		if out, err := dockerRun(ctx, append([]string{"inspect", "-f", f}, strings.Fields(string(ids))...)...); err == nil {
 			for _, line := range strings.Split(strings.TrimSpace(string(out)), "\n") {
-				p := strings.SplitN(line, "|", 3)
+				p := strings.SplitN(line, "|", 4)
-				if len(p) < 3 {
+				if len(p) < 4 {
 					continue
 				}
 				name := strings.TrimPrefix(p[0], "/")
 				if id := strings.TrimPrefix(p[1], "sha256:"); len(id) >= 12 {
 					imageUsed[id[:12]] = true
 				}
 				for _, v := range strings.Split(p[2], ",") {
 					if v != "" {
-						volumeUsers[v] = append(volumeUsers[v], strings.TrimPrefix(p[0], "/"))
+						volumeUsers[v] = append(volumeUsers[v], name)
 					}
 				}
 				for _, n := range strings.Split(p[3], ",") {
 					if n != "" {
 						networkUsers[n] = append(networkUsers[n], name)
 					}
 				}
 			}
@@ -171,8 +194,18 @@ func (s *Server) dockerOverview(w http.ResponseWriter, r *http.Request) {
 		v["used_by"] = users
 		v["anonymous"] = strings.Contains(dstr(v, "Labels"), "com.docker.volume.anonymous")
 	}
 	networks := list("network", "ls", "--format", "{{json .}}")
 	for _, n := range networks {
 		users := networkUsers[dstr(n, "Name")]
 		if users == nil {
 			users = []string{}
 		}
 		sort.Strings(users)
 		n["used_by"] = users
 		n["predefined"] = predefinedNetwork[dstr(n, "Name")]
 	}
 	jsonOK(w, 200, map[string]any{"available": true, "tool": dockerTool(), "containers": containers,
-		"images": images, "volumes": volumes, "networks": list("network", "ls", "--format", "{{json .}}")})
+		"images": images, "volumes": volumes, "networks": networks})
 }
 var dockerRef = regexp.MustCompile(`^[A-Za-z0-9][A-Za-z0-9_.-]{0,127}$`)
@@ -394,3 +427,205 @@ func (s *Server) volumePrune(w http.ResponseWriter, r *http.Request) {
 	log.Printf("docker: volume prune (from %s)", clientIP(r))
 	jsonOK(w, 200, map[string]string{"result": lines[len(lines)-1]})
 }
 // The three networks Docker makes itself. They hold nothing, cannot be removed, and
 // their absence would break every container that asks for one by name.
 var predefinedNetwork = map[string]bool{"bridge": true, "host": true, "none": true}
 // networkRemove deletes one network. Unlike a volume there is no data in it, and unlike an
 // image nothing has to be downloaded again — a stack that wants it back gets it recreated
 // on the next deploy. Docker refuses one a container is still attached to, which is the
 // check that matters, so this asks for no second confirmation.
 func (s *Server) networkRemove(w http.ResponseWriter, r *http.Request) {
 	var body struct{ Name string }
 	if err := json.NewDecoder(io.LimitReader(r.Body, 4096)).Decode(&body); err != nil || !volumeRef.MatchString(body.Name) {
 		jsonErr(w, 400, "which network?")
 		return
 	}
 	if predefinedNetwork[body.Name] {
 		jsonErr(w, 400, body.Name+" is one of Docker's own networks and can't be removed")
 		return
 	}
 	ctx, cancel := context.WithTimeout(r.Context(), 90*time.Second)
 	defer cancel()
 	if out, err := dockerRun(ctx, "network", "rm", body.Name); err != nil {
 		jsonErr(w, 409, dockerWhy(out, err))
 		return
 	}
 	log.Printf("docker: network rm %s (from %s)", body.Name, clientIP(r))
 	w.WriteHeader(204)
 }
 // networkPrune removes every network no container is attached to. A stack that is merely
 // down counts as unattached, so its network goes too — and comes back on the next deploy,
 // with a new subnet. That is the whole cost, and the caller is told it.
 func (s *Server) networkPrune(w http.ResponseWriter, r *http.Request) {
 	ctx, cancel := context.WithTimeout(r.Context(), 5*time.Minute)
 	defer cancel()
 	out, err := dockerRun(ctx, "network", "prune", "-f")
 	if err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	lines := strings.Split(strings.TrimSpace(string(out)), "\n")
 	log.Printf("docker: network prune (from %s)", clientIP(r))
 	jsonOK(w, 200, map[string]string{"result": lines[len(lines)-1]})
 }
 // ---- a container's health, its recent life, its console, and its limits ----
 // dockerHealth is the health check's verdict and the last few probe results. Docker keeps
 // the last five itself, which is all anyone wants: is it failing, and what did it say.
 func (s *Server) dockerHealth(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
 	if !dockerRef.MatchString(id) {
 		jsonErr(w, 400, "bad container")
 		return
 	}
 	ctx, cancel := context.WithTimeout(r.Context(), 15*time.Second)
 	defer cancel()
 	out, err := dockerRun(ctx, "inspect", "--format", "{{json .State}}", id)
 	if err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	var st struct {
 		Status string `json:"Status"`
 		Health *struct {
 			Status        string `json:"Status"`
 			FailingStreak int    `json:"FailingStreak"`
 			Log           []struct {
 				Start    time.Time `json:"Start"`
 				End      time.Time `json:"End"`
 				ExitCode int       `json:"ExitCode"`
 				Output   string    `json:"Output"`
 			} `json:"Log"`
 		} `json:"Health"`
 	}
 	if json.Unmarshal(out, &st) != nil {
 		jsonErr(w, 500, "could not read the container's state")
 		return
 	}
 	if st.Health == nil {
 		jsonOK(w, 200, map[string]any{"configured": false, "state": st.Status})
 		return
 	}
 	jsonOK(w, 200, map[string]any{
 		"configured": true, "state": st.Status, "status": st.Health.Status,
 		"failing_streak": st.Health.FailingStreak, "log": st.Health.Log,
 	})
 }
 // dockerExec opens a shell inside a container, as a terminal session like any other — so
 // it reattaches, resizes and scrolls back the same way, and works on a managed machine
 // through the same proxy.
 func (s *Server) dockerExec(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
 	if !dockerRef.MatchString(id) {
 		jsonErr(w, 400, "bad container")
 		return
 	}
 	tool := dockerTool()
 	if tool == "" {
 		jsonErr(w, 500, "neither docker nor podman is installed here")
 		return
 	}
 	var body struct {
 		Shell string `json:"shell"`
 	}
 	_ = json.NewDecoder(io.LimitReader(r.Body, 4096)).Decode(&body)
 	shell := strings.TrimSpace(body.Shell)
 	if shell == "" {
 		// most images have sh; bash is nicer where it exists, so try it first and fall back
 		shell = "if command -v bash >/dev/null 2>&1; then exec bash; else exec sh; fi"
 	} else if strings.ContainsAny(shell, "\n\r") {
 		jsonErr(w, 400, "bad shell")
 		return
 	}
 	argv := []string{tool, "exec", "-it", id, "sh", "-c", shell}
 	if priv.Available() && !dockerReachable() {
 		argv = priv.Wrap(argv...)
 	}
 	sess, err := s.mgr.CreateCmd("", "sh "+shortName(id), argv)
 	if err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	log.Printf("docker exec into %s as terminal %s by %s", id, sess.ID, clientIP(r))
 	jsonOK(w, 200, map[string]string{"id": sess.ID, "name": "sh " + shortName(id)})
 }
 func shortName(id string) string {
 	if len(id) > 20 {
 		return id[:20]
 	}
 	return id
 }
 // dockerReachable reports whether docker answers without help.
 func dockerReachable() bool {
 	tool := dockerTool()
 	if tool == "" {
 		return false
 	}
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()
 	return exec.CommandContext(ctx, tool, "version", "--format", "{{.Server.Version}}").Run() == nil
 }
 // dockerUpdate changes a running container's limits and restart policy. `docker update`
 // does this live, without recreating it and without touching the compose file — which is
 // also the catch, and the caller is told: a deploy from compose puts back whatever compose
 // says.
 func (s *Server) dockerUpdate(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
 	if !dockerRef.MatchString(id) {
 		jsonErr(w, 400, "bad container")
 		return
 	}
 	var body struct {
 		Memory  string `json:"memory"`  // "512m", "2g", or "" to leave alone; "0" removes the limit
 		CPUs    string `json:"cpus"`    // "1.5", or "0" for no limit
 		Restart string `json:"restart"` // no | on-failure | unless-stopped | always
 	}
 	if err := json.NewDecoder(io.LimitReader(r.Body, 4096)).Decode(&body); err != nil {
 		jsonErr(w, 400, "bad request")
 		return
 	}
 	args := []string{"update"}
 	if m := strings.TrimSpace(body.Memory); m != "" {
 		if !regexp.MustCompile(`^[0-9]+([bkmgBKMG])?$`).MatchString(m) {
 			jsonErr(w, 400, "memory: a number, optionally with b, k, m or g — 0 removes the limit")
 			return
 		}
 		args = append(args, "--memory", m, "--memory-swap", m)
 	}
 	if c := strings.TrimSpace(body.CPUs); c != "" {
 		if !regexp.MustCompile(`^[0-9]+(\.[0-9]+)?$`).MatchString(c) {
 			jsonErr(w, 400, "cpus: a number like 1.5 — 0 removes the limit")
 			return
 		}
 		args = append(args, "--cpus", c)
 	}
 	if p := strings.TrimSpace(body.Restart); p != "" {
 		switch p {
 		case "no", "on-failure", "unless-stopped", "always":
 		default:
 			jsonErr(w, 400, "restart: no, on-failure, unless-stopped or always")
 			return
 		}
 		args = append(args, "--restart", p)
 	}
 	if len(args) == 1 {
 		jsonErr(w, 400, "nothing to change")
 		return
 	}
 	ctx, cancel := context.WithTimeout(r.Context(), 30*time.Second)
 	defer cancel()
 	if _, err := dockerRun(ctx, append(args, id)...); err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	log.Printf("docker update %s %v by %s", id, args[1:], clientIP(r))
 	w.WriteHeader(204)
 }
@@ -0,0 +1,229 @@
 package api
 // A container's history.
 //
 // `docker events` is a live stream with only a short in-memory backlog — ask it for the
 // last thirty days and you get the last few minutes, because the daemon keeps events in a
 // small ring and loses them all when it restarts. That is why Dockhand kept its own copy,
 // and why this does too: a watcher follows the stream and appends the lifecycle events to a
 // ring kept beside the config, so "what happened to this container last Tuesday" has an
 // answer.
 //
 // Only lifecycle events are kept. A health check produces three exec events every probe,
 // which is thousands a day and tells you nothing that the health panel does not.
 import (
 	"bufio"
 	"context"
 	"encoding/json"
 	"log"
 	"net/http"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"sort"
 	"strings"
 	"sync"
 	"time"
 	"hamlab/internal/priv"
 )
 const (
 	eventsKept      = 4000
 	eventsSaveEvery = 30 * time.Second
 )
 // ContainerEvent is one thing that happened to one container.
 type ContainerEvent struct {
 	T        int64  `json:"t"` // unix ms
 	Name     string `json:"name"`
 	Action   string `json:"action"`
 	Image    string `json:"image,omitempty"`
 	ExitCode string `json:"exit_code,omitempty"`
 	Signal   string `json:"signal,omitempty"`
 	Health   string `json:"health,omitempty"`
 }
 type eventLog struct {
 	mu    sync.Mutex
 	ring  []ContainerEvent
 	dirty bool
 	path  string
 }
 func (e *eventLog) load(path string) {
 	e.mu.Lock()
 	defer e.mu.Unlock()
 	e.path = path
 	b, err := os.ReadFile(path)
 	if err != nil {
 		return
 	}
 	var list []ContainerEvent
 	if json.Unmarshal(b, &list) == nil {
 		e.ring = list
 	}
 }
 func (e *eventLog) add(ev ContainerEvent) {
 	e.mu.Lock()
 	defer e.mu.Unlock()
 	e.ring = append(e.ring, ev)
 	if len(e.ring) > eventsKept {
 		e.ring = append([]ContainerEvent(nil), e.ring[len(e.ring)-eventsKept:]...)
 	}
 	e.dirty = true
 }
 // forName is one container's events, newest first.
 func (e *eventLog) forName(name string, since time.Time, limit int) []ContainerEvent {
 	e.mu.Lock()
 	defer e.mu.Unlock()
 	out := []ContainerEvent{}
 	for _, ev := range e.ring {
 		if ev.Name == name && ev.T >= since.UnixMilli() {
 			out = append(out, ev)
 		}
 	}
 	sort.Slice(out, func(i, j int) bool { return out[i].T > out[j].T })
 	if len(out) > limit {
 		out = out[:limit]
 	}
 	return out
 }
 func (e *eventLog) save() {
 	e.mu.Lock()
 	if !e.dirty || e.path == "" {
 		e.mu.Unlock()
 		return
 	}
 	b, err := json.Marshal(e.ring)
 	e.dirty = false
 	path := e.path
 	e.mu.Unlock()
 	if err != nil {
 		return
 	}
 	_ = os.MkdirAll(filepath.Dir(path), 0o700)
 	tmp := path + ".tmp"
 	if os.WriteFile(tmp, b, 0o600) == nil {
 		_ = os.Rename(tmp, path)
 	}
 }
 // watchDockerEvents follows the stream, restarting if Docker goes away. It is the only
 // thing that writes the log, so nothing contends with it.
 func (s *Server) watchDockerEvents(ctx context.Context) {
 	s.events.load(filepath.Join(s.cfg.DataDir(), "docker-events.json"))
 	go func() { // a periodic flush, so a crash loses at most half a minute of history
 		t := time.NewTicker(eventsSaveEvery)
 		defer t.Stop()
 		for {
 			select {
 			case <-ctx.Done():
 				s.events.save()
 				return
 			case <-t.C:
 				s.events.save()
 			}
 		}
 	}()
 	for {
 		if ctx.Err() != nil {
 			return
 		}
 		if dockerTool() == "" {
 			time.Sleep(time.Minute) // no Docker here; look again later in case it is installed
 			continue
 		}
 		s.followEvents(ctx)
 		select {
 		case <-ctx.Done():
 			return
 		case <-time.After(10 * time.Second): // the daemon restarted, or the pipe broke
 		}
 	}
 }
 func (s *Server) followEvents(ctx context.Context) {
 	args := []string{"events", "--format", "{{json .}}", "--filter", "type=container"}
 	for _, e := range []string{"create", "start", "restart", "stop", "kill", "die", "destroy", "oom", "pause", "unpause", "rename", "update", "health_status"} {
 		args = append(args, "--filter", "event="+e)
 	}
 	tool := dockerTool()
 	full := append([]string{tool}, args...)
 	if !dockerReachable() && priv.Available() {
 		full = priv.Wrap(full...)
 	}
 	cmd := exec.CommandContext(ctx, full[0], full[1:]...)
 	out, err := cmd.StdoutPipe()
 	if err != nil {
 		return
 	}
 	if err := cmd.Start(); err != nil {
 		return
 	}
 	sc := bufio.NewScanner(out)
 	sc.Buffer(make([]byte, 64<<10), 1<<20)
 	n := 0
 	for sc.Scan() {
 		var e struct {
 			Action   string `json:"Action"`
 			Time     int64  `json:"time"`
 			TimeNano int64  `json:"timeNano"`
 			Actor    struct {
 				Attributes map[string]string `json:"Attributes"`
 			} `json:"Actor"`
 		}
 		if json.Unmarshal(sc.Bytes(), &e) != nil || e.Action == "" {
 			continue
 		}
 		at := e.Actor.Attributes
 		ev := ContainerEvent{
 			T: e.Time * 1000, Name: at["name"], Image: at["image"],
 			Action: e.Action, ExitCode: at["exitCode"], Signal: at["signal"],
 		}
 		if e.TimeNano > 0 {
 			ev.T = e.TimeNano / 1e6
 		}
 		// health_status comes through as "health_status: healthy"
 		if a, h, ok := strings.Cut(e.Action, ": "); ok {
 			ev.Action, ev.Health = a, h
 		}
 		if ev.Name == "" {
 			continue
 		}
 		s.events.add(ev)
 		n++
 	}
 	_ = cmd.Wait()
 	if n > 0 {
 		log.Printf("docker events: stream ended after %d events; reconnecting", n)
 	}
 }
 // dockerEvents answers from Ham Lab's own history, not Docker's short memory.
 func (s *Server) dockerEvents(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
 	if !dockerRef.MatchString(id) {
 		jsonErr(w, 400, "bad container")
 		return
 	}
 	since := r.URL.Query().Get("since")
 	d := 7 * 24 * time.Hour
 	switch since {
 	case "24h":
 		d = 24 * time.Hour
 	case "30d":
 		d = 30 * 24 * time.Hour
 	default:
 		since = "7d"
 	}
 	ev := s.events.forName(id, time.Now().Add(-d), 500)
 	s.events.mu.Lock()
 	total := len(s.events.ring)
 	s.events.mu.Unlock()
 	jsonOK(w, 200, map[string]any{"events": ev, "since": since, "kept": total, "capacity": eventsKept})
 }
@@ -0,0 +1,206 @@
 package api
 // Following a container's log over a held connection, the way the Services tab follows a
 // journal. The polled `?since=` endpoint next door still exists — it is what a reconnect
 // hands back to pick up from — but the live view holds one `docker logs -f` per viewer
 // instead of asking again every two seconds.
 import (
 	"bufio"
 	"context"
 	"encoding/json"
 	"io"
 	"net/http"
 	"os/exec"
 	"sort"
 	"strconv"
 	"strings"
 	"syscall"
 	"time"
 	"github.com/coder/websocket"
 	"hamlab/internal/priv"
 )
 // logLine is one line of a container's output: the time Docker stamped it with, the text,
 // and which of the two streams it came out of — stderr is where a container says what went
 // wrong, and the page colours it accordingly.
 type logLine struct {
 	T   string `json:"t"`
 	M   string `json:"m"`
 	Err bool   `json:"err,omitempty"`
 	at  string // when to sort it by: its own stamp, or the last one its stream carried
 }
 // splitStamp pulls Docker's RFC 3339 prefix off a line. A line without one (a container
 // that writes bare bytes, or a continuation) keeps the whole text and no time — but is
 // still sorted where its stream put it, behind the last line that did have one.
 func splitStamp(l string, carry *string) logLine {
 	t, m, ok := strings.Cut(l, " ")
 	if _, err := time.Parse(time.RFC3339Nano, t); !ok || err != nil {
 		return logLine{M: l, at: *carry}
 	}
 	*carry = t
 	return logLine{T: t, M: m, at: t}
 }
 // dockerTail streams a container's output as JSON text frames until the client goes away.
 // `?since=<RFC3339>` resumes where a dropped connection left off; otherwise it starts with
 // the last `?n=` lines.
 func (s *Server) dockerTail(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
 	if !dockerRef.MatchString(id) {
 		jsonErr(w, 400, "bad container")
 		return
 	}
 	tool := dockerTool()
 	if tool == "" {
 		jsonErr(w, 500, "neither docker nor podman is installed here")
 		return
 	}
 	args := []string{tool, "logs", "--timestamps", "--follow"}
 	if since := r.URL.Query().Get("since"); since != "" {
 		if _, err := time.Parse(time.RFC3339Nano, since); err != nil {
 			jsonErr(w, 400, "since is an RFC 3339 time")
 			return
 		}
 		args = append(args, "--since", since)
 	} else {
 		args = append(args, "--tail", strconv.Itoa(logCount(r)))
 	}
 	args = append(args, id)
 	if priv.Available() && !dockerReachable() {
 		args = priv.Wrap(args...)
 	}
 	conn, err := websocket.Accept(w, r, &websocket.AcceptOptions{CompressionMode: websocket.CompressionDisabled})
 	if err != nil {
 		return
 	}
 	defer conn.Close(websocket.StatusNormalClosure, "")
 	ctx, cancel := context.WithCancel(r.Context())
 	defer cancel()
 	cmd := exec.CommandContext(ctx, args[0], args[1:]...)
 	// Own process group, so the kill reaches the docker client itself and not just a sudo
 	// wrapper — the same reason the journal tail does it.
 	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
 	cmd.Cancel = func() error {
 		pg := "-" + strconv.Itoa(cmd.Process.Pid)
 		if args[0] != tool { // running under sudo: only sudo can signal root's children
 			k := priv.Wrap("kill", "-9", "--", pg)
 			return exec.Command(k[0], k[1:]...).Run()
 		}
 		return syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
 	}
 	stdout, err := cmd.StdoutPipe()
 	if err != nil {
 		return
 	}
 	stderr, err := cmd.StderrPipe()
 	if err != nil {
 		return
 	}
 	if err := cmd.Start(); err != nil {
 		_ = conn.Write(ctx, websocket.MessageText, mustJSON(logLine{M: err.Error(), Err: true}))
 		return
 	}
 	// Both pipes funnel into one writer: a websocket takes one writer at a time.
 	lines := make(chan logLine, 256)
 	read := func(p io.Reader, isErr bool) {
 		sc := bufio.NewScanner(p)
 		sc.Buffer(make([]byte, 0, 64<<10), 4<<20)
 		carry := ""
 		for sc.Scan() {
 			l := splitStamp(sc.Text(), &carry)
 			l.Err = isErr
 			select {
 			case lines <- l:
 			case <-ctx.Done():
 				return
 			}
 		}
 	}
 	done := make(chan struct{}, 2)
 	go func() { read(stdout, false); done <- struct{}{} }()
 	go func() { read(stderr, true); done <- struct{}{} }()
 	go func() { // the command ends when the container does; closing says so, having said why
 		<-done
 		<-done
 		_ = cmd.Wait()
 		close(lines)
 	}()
 	go func() { // reader: only to notice the close
 		defer cancel()
 		for {
 			if _, _, err := conn.Read(ctx); err != nil {
 				return
 			}
 		}
 	}()
 	ping := time.NewTicker(25 * time.Second)
 	defer ping.Stop()
 	// stdout and stderr are two pipes, and what Docker replays down them arrives in no
 	// particular order. So the opening burst is held for a moment and sorted by the time
 	// Docker stamped each line with; after that the log is live and arrives in order anyway.
 	const burstFor = 300 * time.Millisecond
 	warm := time.NewTimer(burstFor)
 	defer warm.Stop()
 	warmC := warm.C // nil once the burst has been flushed: a nil channel never fires
 	var burst []logLine
 	write := func(l logLine) bool {
 		return conn.Write(ctx, websocket.MessageText, mustJSON(l)) == nil
 	}
 	flush := func() bool {
 		sort.SliceStable(burst, func(i, j int) bool { return burst[i].at < burst[j].at })
 		for _, l := range burst {
 			if !write(l) {
 				return false
 			}
 		}
 		burst = nil
 		return true
 	}
 	for {
 		select {
 		case <-ctx.Done():
 			return
 		case l, ok := <-lines:
 			if !ok { // the container stopped, or never existed: say the last of it and go
 				flush()
 				return
 			}
 			if warmC != nil {
 				burst = append(burst, l)
 				if len(burst) < 10000 { // a container that never stops talking still gets read
 					continue
 				}
 				if !flush() {
 					return
 				}
 				continue
 			}
 			if !write(l) {
 				return
 			}
 		case <-warmC:
 			warmC = nil
 			if !flush() {
 				return
 			}
 		case <-ping.C:
 			pctx, pc := context.WithTimeout(ctx, 10*time.Second)
 			err := conn.Ping(pctx)
 			pc()
 			if err != nil {
 				return
 			}
 		}
 	}
 }
 func mustJSON(v any) []byte { b, _ := json.Marshal(v); return b }
@@ -0,0 +1,351 @@
 package api
 // Scheduled image updates.
 //
 // Dockhand kept three containers up to date nightly — unbound, nginx and pihole, which is
 // DNS and ingress — and that is the one thing it does that Ham Lab did not. This replaces
 // it, and deliberately in the same shape as the ZFS snapshot schedules: a real line in a
 // cron file Ham Lab owns, calling the binary, rather than a ticker inside the daemon.
 // Updating an image is not something that should stop happening because a web console is
 // down, and a cron line is visible to anyone looking at the machine the normal way.
 //
 // A schedule stores everything the run needs — project, compose files, how compose has to
 // be invoked for that stack — so the cron job is self-contained and does not have to
 // rediscover the world before it can act.
 import (
 	"context"
 	"crypto/rand"
 	"encoding/hex"
 	"encoding/json"
 	"errors"
 	"io"
 	"log"
 	"net/http"
 	"os"
 	"path/filepath"
 	"strings"
 	"time"
 	"hamlab/internal/compose"
 	"hamlab/internal/cron"
 	"hamlab/internal/storage"
 	"hamlab/internal/store"
 )
 // UpdateJob is a standing instruction to pull a stack's images, and optionally deploy them.
 type UpdateJob struct {
 	ID      string `json:"id"`
 	Name    string `json:"name"`
 	Stack   string `json:"stack"`
 	Machine string `json:"machine,omitempty"` // for display; the job runs where it was made
 	// Services narrows it to some of the stack's services; empty is all of them.
 	Services []string `json:"services,omitempty"`
 	Cron     string   `json:"cron"`
 	// Apply says whether a newer image is deployed or only fetched. Check-only still pulls,
 	// so the Docker tab shows what is behind without a registry round trip.
 	Apply   bool `json:"apply"`
 	Enabled bool `json:"enabled"`
 	// How to run compose for this stack, captured when the schedule was made.
 	Spec composeSpec `json:"spec"`
 }
 type composeSpec struct {
 	Project string   `json:"project"`
 	Files   []string `json:"files"`
 	Via     string   `json:"via,omitempty"`
 	Remote  string   `json:"remote,omitempty"`
 }
 func (c composeSpec) toSpec(tool string) compose.Spec {
 	return compose.Spec{Project: c.Project, Files: c.Files, Via: c.Via, Remote: c.Remote, Tool: tool}
 }
 const dockerCronFile = "/etc/cron.d/hamlab-docker"
 type dockerSchedState struct {
 	Jobs []UpdateJob `json:"jobs"`
 }
 func (s *Server) dsLoad() dockerSchedState {
 	var raw json.RawMessage
 	s.store.View(func(st *store.State) { raw = st.DockerJobs })
 	var d dockerSchedState
 	if len(raw) > 0 {
 		_ = json.Unmarshal(raw, &d)
 	}
 	if d.Jobs == nil {
 		d.Jobs = []UpdateJob{}
 	}
 	return d
 }
 func (s *Server) dsSave(d dockerSchedState) error {
 	b, err := json.Marshal(d)
 	if err != nil {
 		return err
 	}
 	return s.store.Update(func(st *store.State) { st.DockerJobs = b })
 }
 // dsCronText renders the schedules as cron lines. Same shape as the ZFS file: Ham Lab owns
 // it whole and rewrites it on every change.
 func dsCronText(jobs []UpdateJob, user, exe, config string) (string, error) {
 	var b strings.Builder
 	b.WriteString("# Ham Lab image-update schedules — written by the Docker tab, rewritten whole on every change.\n")
 	b.WriteString("# Edit them there, not here: anything added by hand is lost the next time a schedule changes.\n")
 	b.WriteString("SHELL=/bin/sh\nPATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin\n\n")
 	for i := range jobs {
 		j := &jobs[i]
 		if !j.Enabled {
 			b.WriteString("# " + j.Name + " — paused\n")
 			continue
 		}
 		expr := strings.Join(strings.Fields(j.Cron), " ")
 		if err := cron.ValidSchedule(expr); err != nil {
 			return "", err
 		}
 		what := "pull only"
 		if j.Apply {
 			what = "pull and deploy"
 		}
 		b.WriteString("# " + j.Name + " — " + j.Stack + ", " + what + "\n")
 		b.WriteString(expr + " " + user + " " + exe + " -config " + config + " docker-update " + j.ID + "\n\n")
 	}
 	return b.String(), nil
 }
 func (s *Server) dsSyncCron(ctx context.Context, d dockerSchedState) error {
 	live := false
 	for i := range d.Jobs {
 		if d.Jobs[i].Enabled {
 			live = true
 		}
 	}
 	if !live {
 		return s.storage.RemoveRootFile(ctx, dockerCronFile)
 	}
 	exe, err := os.Executable()
 	if err != nil {
 		return err
 	}
 	exe, _ = filepath.EvalSymlinks(exe)
 	text, err := dsCronText(d.Jobs, s.user, exe, s.cfg.Path())
 	if err != nil {
 		return err
 	}
 	return s.storage.WriteRootFile(ctx, dockerCronFile, text)
 }
 // repoint keeps a schedule working when what deployed its stack has been removed.
 //
 // A schedule captures how to invoke compose at the moment it is made, so the cron job never
 // has to rediscover the world at 3am. But "run compose inside this container" stops being
 // true the day that container is deleted — which is exactly what happens when an
 // orchestrator is retired, and would otherwise break every schedule made before it went.
 // So: if the container named is not running, the files are looked up again, through live
 // mounts and remembered ones both, and compose runs on the host instead.
 func repoint(ctx context.Context, db *store.Store, spec compose.Spec) compose.Spec {
 	if spec.Via == "" || db == nil {
 		return spec
 	}
 	if out, err := dockerRun(ctx, "inspect", "-f", "{{.State.Running}}", spec.Via); err == nil && strings.TrimSpace(string(out)) == "true" {
 		return spec
 	}
 	st := stack{Name: spec.Project, File: spec.Files[0], Files: spec.Files}
 	st.resolve(mountsSeen(ctx, db)) // read-only: this also runs from cron, where the daemon owns state.json
 	if !st.Editable || st.Via != "" {
 		return spec // still only reachable through a container, or not reachable at all
 	}
 	log.Printf("docker schedule %s: %s is gone, running compose here against %s", spec.Project, spec.Via, st.HostFiles[0])
 	spec.Files, spec.Via = st.HostFiles, ""
 	return spec
 }
 // RunUpdate pulls, and deploys when the schedule says to. Shared with the CLI so the button
 // and the cron line cannot drift apart.
 func RunUpdate(ctx context.Context, db *store.Store, j UpdateJob, tool string) (string, error) {
 	if len(j.Spec.Files) == 0 {
 		return "", errors.New("this schedule has no compose file recorded — open it in the Docker tab and save it again")
 	}
 	spec := repoint(ctx, db, j.Spec.toSpec(tool))
 	var log strings.Builder
 	run := func(args ...string) error {
 		cmd := compose.Cmd(ctx, spec, args...)
 		out, err := cmd.CombinedOutput()
 		log.Write(out)
 		return err
 	}
 	args := append([]string{"pull"}, j.Services...)
 	if err := run(args...); err != nil {
 		return log.String(), err
 	}
 	if !j.Apply {
 		return log.String(), nil
 	}
 	up := append([]string{"up", "-d"}, j.Services...)
 	if err := run(up...); err != nil {
 		return log.String(), err
 	}
 	return log.String(), nil
 }
 // ---- endpoints ----
 func (s *Server) dsList(w http.ResponseWriter, _ *http.Request) {
 	d := s.dsLoad()
 	now := time.Now()
 	out := make([]map[string]any, 0, len(d.Jobs))
 	for _, j := range d.Jobs {
 		m := map[string]any{"job": j, "last": storage.ReadRun(s.cfg.DataDir(), "docker-"+j.ID)}
 		if j.Enabled {
 			if n, ok := cron.NextRun(strings.Join(strings.Fields(j.Cron), " "), now); ok {
 				m["next"] = n
 			}
 		}
 		out = append(out, m)
 	}
 	jsonOK(w, 200, map[string]any{"jobs": out, "file": dockerCronFile})
 }
 func (s *Server) dsPut(w http.ResponseWriter, r *http.Request) {
 	var j UpdateJob
 	if err := json.NewDecoder(io.LimitReader(r.Body, 32<<10)).Decode(&j); err != nil {
 		jsonErr(w, 400, "bad request")
 		return
 	}
 	j.Cron = strings.Join(strings.Fields(j.Cron), " ")
 	if err := cron.ValidSchedule(j.Cron); err != nil {
 		jsonErr(w, 400, "schedule: "+err.Error())
 		return
 	}
 	if strings.TrimSpace(j.Stack) == "" {
 		jsonErr(w, 400, "which stack?")
 		return
 	}
 	// resolve the stack now, so the cron job never has to
 	ctx, cancel := context.WithTimeout(r.Context(), 60*time.Second)
 	defer cancel()
 	stacks, err := s.findStacks(ctx)
 	if err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	found := false
 	for _, st := range stacks {
 		if st.Name == j.Stack {
 			// seenFiles is the list as compose itself sees it: host paths normally, the
 			// orchestrator container's own paths when compose runs inside one. Getting this
 			// backwards gives a cron line that works from the console and fails at 3am.
 			j.Spec = composeSpec{Project: st.Name, Files: seenFiles(st, -1, ""), Via: st.Via, Remote: st.Remote}
 			if len(j.Spec.Files) == 0 {
 				jsonErr(w, 409, "that stack's compose file isn't anywhere this console can reach")
 				return
 			}
 			found = true
 		}
 	}
 	if !found {
 		jsonErr(w, 404, "no stack called "+j.Stack+" on this machine")
 		return
 	}
 	if strings.TrimSpace(j.Name) == "" {
 		j.Name = j.Stack
 	}
 	d := s.dsLoad()
 	replaced := false
 	for i := range d.Jobs {
 		if d.Jobs[i].ID == j.ID && j.ID != "" {
 			d.Jobs[i] = j
 			replaced = true
 		}
 	}
 	if !replaced {
 		b := make([]byte, 6)
 		if _, err := rand.Read(b); err != nil {
 			jsonErr(w, 500, err.Error())
 			return
 		}
 		j.ID = hex.EncodeToString(b)
 		d.Jobs = append(d.Jobs, j)
 	}
 	if err := s.dsSave(d); err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	if err := s.dsSyncCron(r.Context(), d); err != nil {
 		jsonErr(w, 500, "saved, but writing "+dockerCronFile+" failed: "+err.Error())
 		return
 	}
 	log.Printf("docker schedule %q saved by %s (%s, %q, %s)", j.Name, clientIP(r), j.Stack, j.Cron, map[bool]string{true: "pull and deploy", false: "pull only"}[j.Apply])
 	jsonOK(w, 200, j)
 }
 func (s *Server) dsDelete(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
 	d := s.dsLoad()
 	kept := d.Jobs[:0]
 	found := false
 	for _, j := range d.Jobs {
 		if j.ID == id {
 			found = true
 			continue
 		}
 		kept = append(kept, j)
 	}
 	if !found {
 		jsonErr(w, 404, "no such schedule")
 		return
 	}
 	d.Jobs = kept
 	if err := s.dsSave(d); err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	if err := s.dsSyncCron(r.Context(), d); err != nil {
 		jsonErr(w, 500, "removed, but writing "+dockerCronFile+" failed: "+err.Error())
 		return
 	}
 	storage.ForgetRun(s.cfg.DataDir(), "docker-"+id)
 	log.Printf("docker schedule %s removed by %s", id, clientIP(r))
 	w.WriteHeader(204)
 }
 func (s *Server) dsRun(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
 	var job UpdateJob
 	found := false
 	for _, j := range s.dsLoad().Jobs {
 		if j.ID == id {
 			job, found = j, true
 		}
 	}
 	if !found {
 		jsonErr(w, 404, "no such schedule")
 		return
 	}
 	// Outliving the request: a pull the browser gave up on still finishes, but not forever.
 	ctx, cancel := context.WithTimeout(context.WithoutCancel(r.Context()), 30*time.Minute)
 	defer cancel()
 	out, err := RunUpdate(ctx, s.store, job, dockerTool())
 	rec := storage.RunRecord{At: time.Now(), Taken: job.Stack}
 	if err != nil {
 		rec.Error = err.Error()
 	}
 	_ = storage.WriteRun(s.cfg.DataDir(), "docker-"+id, rec)
 	stackUpdatesForget("/" + job.Stack)
 	if err != nil {
 		jsonErr(w, 500, err.Error()+"\n"+tail(out, 2000))
 		return
 	}
 	log.Printf("docker schedule %q run by %s", job.Name, clientIP(r))
 	jsonOK(w, 200, map[string]string{"output": tail(out, 4000)})
 }
 func tail(s string, n int) string {
 	if len(s) <= n {
 		return s
 	}
 	return "…" + s[len(s)-n:]
 }
@@ -0,0 +1,193 @@
 package api
 // Guest logins: a temporary account for handing to somebody you want to show around. It is
 // an ordinary account with a clock on it — every tab, every action, on this machine and on
 // every machine this hub manages — that stops working at its expiry (half an hour by
 // default) or the moment it is revoked.
 //
 // Deliberately not a reduced account: Mitchell, 2026-09-27, replacing a read-only version
 // built the same day. He gives these to people he trusts and revokes them otherwise, and a
 // console with half its buttons missing is not the console he wanted to show off. What that
 // means in practice: a guest has the Terminal tab, which is a shell as the user the daemon
 // runs as — root on every host here — and can make an account or a token of its own that
 // outlives the guest login.
 import (
 	"encoding/json"
 	"io"
 	"log"
 	"net/http"
 	"strings"
 	"time"
 	"hamlab/internal/auth"
 	"hamlab/internal/config"
 	"hamlab/internal/store"
 )
 const (
 	guestPrefix     = "guest-"
 	guestTTLDefault = 30 * time.Minute
 	guestTTLMax     = 24 * time.Hour
 	guestReapEvery  = time.Minute
 )
 // ---- the accounts themselves ----
 func (s *Server) listGuests(w http.ResponseWriter, r *http.Request) {
 	now := time.Now()
 	counts := s.sessions.Count()
 	out := []map[string]any{}
 	s.cfg.View(func(c *config.Config) {
 		for _, u := range c.Users {
 			if !u.Guest || u.Expired(now) {
 				continue
 			}
 			out = append(out, map[string]any{
 				"name": u.Name, "created": u.Created, "expires": u.Expires,
 				"issued_by": u.IssuedBy, "sessions": counts[u.Name], "used": !u.LastSeen.IsZero(),
 			})
 		}
 	})
 	jsonOK(w, 200, out)
 }
 // addGuest makes a guest login and hands back its password — the only time it is ever
 // readable, here as everywhere else a password is generated.
 func (s *Server) addGuest(w http.ResponseWriter, r *http.Request) {
 	me := actorOf(r)
 	var body struct {
 		Minutes int `json:"minutes"`
 	}
 	_ = json.NewDecoder(io.LimitReader(r.Body, 4096)).Decode(&body)
 	ttl := guestTTLDefault
 	if body.Minutes > 0 {
 		ttl = time.Duration(body.Minutes) * time.Minute
 	}
 	if ttl > guestTTLMax {
 		jsonErr(w, 400, "a guest login lasts at most 24 hours")
 		return
 	}
 	pw, err := auth.GuestPassword()
 	if err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	hash, err := auth.HashPassword(pw)
 	if err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	var name string
 	for try := 0; try < 8 && name == ""; try++ {
 		suffix, err := auth.GuestSuffix()
 		if err != nil {
 			jsonErr(w, 500, err.Error())
 			return
 		}
 		candidate := guestPrefix + suffix
 		s.cfg.View(func(c *config.Config) {
 			if _, taken := c.User(candidate); !taken {
 				name = candidate
 			}
 		})
 	}
 	if name == "" {
 		jsonErr(w, 500, "could not find a free guest name")
 		return
 	}
 	now := time.Now()
 	expires := now.Add(ttl)
 	if err := s.cfg.Update(func(c *config.Config) {
 		c.Users = append(c.Users, config.User{Name: name, Hash: hash, Created: now, Guest: true, Expires: expires, IssuedBy: me.User})
 	}); err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	log.Printf("guest login %q issued until %s (by %s)", name, expires.Format(time.RFC3339), clientIP(r))
 	jsonOK(w, 200, map[string]any{"name": name, "password": pw, "expires": expires, "minutes": int(ttl.Minutes())})
 }
 // revokeGuest ends a guest login now. No password: it only ever takes access away, and the
 // whole point of the thing is that ending it is one tap.
 func (s *Server) revokeGuest(w http.ResponseWriter, r *http.Request) {
 	name := r.PathValue("name")
 	found := false
 	if err := s.cfg.Update(func(c *config.Config) {
 		kept := c.Users[:0]
 		for _, u := range c.Users {
 			if u.Name == name && u.Guest {
 				found = true
 				continue
 			}
 			kept = append(kept, u)
 		}
 		c.Users = kept
 	}); err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	if !found {
 		jsonErr(w, 404, "no such guest login")
 		return
 	}
 	s.forgetGuest(name)
 	log.Printf("guest login %q revoked (by %s)", name, clientIP(r))
 	w.WriteHeader(204)
 }
 // forgetGuest signs a guest out everywhere and drops the look they left behind.
 func (s *Server) forgetGuest(name string) {
 	s.sessions.RevokeUser(name)
 	_ = s.store.Update(func(st *store.State) { delete(st.Users, name) })
 }
 // watchGuests clears out guest logins whose time is up. Expiry is enforced on every request
 // as well — this is what keeps the config file from collecting them, and what signs a guest
 // out of a console they left open rather than waiting for their next click.
 func (s *Server) watchGuests() {
 	for {
 		s.reapGuests()
 		time.Sleep(guestReapEvery)
 	}
 }
 func (s *Server) reapGuests() {
 	now := time.Now()
 	var gone []string
 	var have bool
 	s.cfg.View(func(c *config.Config) {
 		for _, u := range c.Users {
 			if u.Guest {
 				have = true
 				if u.Expired(now) {
 					gone = append(gone, u.Name)
 				}
 			}
 		}
 	})
 	if !have || len(gone) == 0 {
 		return
 	}
 	expired := map[string]bool{}
 	for _, n := range gone {
 		expired[n] = true
 	}
 	if err := s.cfg.Update(func(c *config.Config) {
 		kept := c.Users[:0]
 		for _, u := range c.Users {
 			if expired[u.Name] {
 				continue
 			}
 			kept = append(kept, u)
 		}
 		c.Users = kept
 	}); err != nil {
 		log.Printf("guests: %v", err)
 		return
 	}
 	for _, n := range gone {
 		s.forgetGuest(n)
 	}
 	log.Printf("guest login(s) expired: %s", strings.Join(gone, ", "))
 }
@@ -0,0 +1,70 @@
 package api
 import (
 	"net/http"
 	"net/http/httptest"
 	"path/filepath"
 	"testing"
 	"time"
 	"hamlab/internal/auth"
 	"hamlab/internal/config"
 	"hamlab/internal/store"
 )
 // A guest login is an ordinary account with a clock on it, so the clock is the thing worth
 // testing: one whose time is up is stopped at the door, whether or not the minute-by-minute
 // sweep has reached their account yet, and their session is dropped on the way — so the
 // console they left open goes back to the login form rather than carrying on.
 func TestExpiredGuestIsRefusedAndSignedOut(t *testing.T) {
 	for _, tc := range []struct {
 		name    string
 		expires time.Time
 		wantOK  bool
 	}{
 		{"still has time", time.Now().Add(10 * time.Minute), true},
 		{"time is up", time.Now().Add(-time.Second), false},
 		{"no expiry at all", time.Time{}, false},
 	} {
 		t.Run(tc.name, func(t *testing.T) {
 			dir := t.TempDir()
 			cfg := config.Default(filepath.Join(dir, "config.json"))
 			if err := cfg.Update(func(c *config.Config) {
 				c.Users = []config.User{{Name: "guest-abcd", Hash: "x", Guest: true, Expires: tc.expires}}
 			}); err != nil {
 				t.Fatal(err)
 			}
 			st, err := store.Open(filepath.Join(dir, "state.json"))
 			if err != nil {
 				t.Fatal(err)
 			}
 			s := &Server{cfg: cfg, store: st, sessions: auth.NewSessions(st)}
 			tok, err := s.sessions.Create("guest-abcd", "10.0.0.5", "curl")
 			if err != nil {
 				t.Fatal(err)
 			}
 			r := httptest.NewRequest("GET", "/api/me", nil)
 			r.AddCookie(&http.Cookie{Name: cookieName, Value: tok})
 			w := httptest.NewRecorder()
 			a, ok, answered := s.cookieActor(w, r)
 			if ok != tc.wantOK {
 				t.Fatalf("ok = %v, want %v", ok, tc.wantOK)
 			}
 			if tc.wantOK {
 				if !a.Guest || a.User != "guest-abcd" {
 					t.Errorf("actor = %+v, want the guest account marked as one", a)
 				}
 				if answered {
 					t.Error("answered the caller for a guest that still has time")
 				}
 				return
 			}
 			if !answered || w.Code != 401 {
 				t.Errorf("answered = %v, code = %d, want true and 401", answered, w.Code)
 			}
 			if _, live := s.sessions.Lookup(tok, "10.0.0.5"); live {
 				t.Error("the session of an expired guest is still live")
 			}
 		})
 	}
 }
@@ -0,0 +1,13 @@
 package api
 // The Hardware tab: what the machine is made of. Read once and remembered — none of it
 // changes while the kernel runs — with ?refresh=1 for the one thing that can, a drive
 // plugged in or pulled out.
 import (
 	"net/http"
 )
 func (s *Server) hardwareReport(w http.ResponseWriter, r *http.Request) {
 	jsonOK(w, 200, s.hardware.Report(r.Context(), r.URL.Query().Get("refresh") == "1"))
 }
@@ -0,0 +1,434 @@
 package api
 // IP-KVMs: a device's own web page, shown inside Ham Lab.
 //
 // The whole difficulty is that Ham Lab is served over TLS and a KVM usually is not, so a
 // browser refuses to put the device's page in an iframe — mixed content, blocked before
 // anything loads. The fix is to serve the device from Ham Lab's own origin, which also
 // means it inherits Ham Lab's login instead of needing a second one, and makes the iframe
 // same-origin.
 //
 // Proxying an arbitrary web app under a path prefix is the hard part, and nothing here
 // knows anything about any particular device. A device UI asks for its assets by absolute
 // path — "/assets/app.js", "/api/whatever" — which a prefix cannot rewrite away. So the
 // prefix is only the entry point: any request Ham Lab does not serve itself is routed to
 // whichever KVM the asking page belongs to, worked out from the Referer, and from a cookie
 // when there is no Referer (a WebSocket upgrade, say). Ham Lab's own routes always win, so
 // a device can never shadow the console.
 import (
 	"crypto/rand"
 	"encoding/hex"
 	"encoding/json"
 	"errors"
 	"io"
 	"log"
 	"net/http"
 	"net/http/httputil"
 	"net/url"
 	"strings"
 	"sync"
 	"time"
 	"hamlab/internal/config"
 	"hamlab/internal/store"
 )
 const kvmPrefix = "/kvm/"
 const kvmCookie = "hamlab_kvm"
 type kvmProxies struct {
 	mu sync.Mutex
 	m  map[string]*httputil.ReverseProxy
 }
 func (s *Server) kvm(id string) (config.KVM, bool) {
 	var found config.KVM
 	ok := false
 	s.cfg.View(func(c *config.Config) {
 		for _, k := range c.KVMs {
 			if k.ID == id {
 				found, ok = k, true
 			}
 		}
 	})
 	return found, ok
 }
 // kvmProxy is the reverse proxy for one device, made once and kept.
 func (s *Server) kvmProxy(k config.KVM) (*httputil.ReverseProxy, error) {
 	s.kvms.mu.Lock()
 	defer s.kvms.mu.Unlock()
 	if s.kvms.m == nil {
 		s.kvms.m = map[string]*httputil.ReverseProxy{}
 	}
 	if p, ok := s.kvms.m[k.ID]; ok {
 		return p, nil
 	}
 	target, err := url.Parse(k.URL)
 	if err != nil || target.Host == "" {
 		return nil, errors.New("this KVM's address is not a URL")
 	}
 	prefix := kvmPrefix + k.ID
 	p := &httputil.ReverseProxy{
 		Director: func(r *http.Request) {
 			r.URL.Scheme, r.URL.Host = target.Scheme, target.Host
 			// Entry through the prefix keeps the path after it; anything else arrived by its
 			// own absolute path and is passed through untouched.
 			if rest, ok := strings.CutPrefix(r.URL.Path, prefix); ok {
 				if rest == "" {
 					rest = "/"
 				}
 				r.URL.Path = rest
 			}
 			r.URL.RawPath = ""
 			r.Host = target.Host
 			// Whether the console itself is on TLS: the response rewriter needs to know, and by
 			// then only the outbound request is in hand.
 			if secure(r) {
 				r.Header.Set(kvmSecureHeader, "1")
 			} else {
 				r.Header.Del(kvmSecureHeader)
 			}
 			// The device gets its own host and none of Ham Lab's session.
 			r.Header.Del("Origin")
 			stripCookie(r, cookieName, kvmCookie)
 			// What is left is the device's own login. A browser that has it teaches the hub; one
 			// that has not is lent what the hub remembers, so the device sees a signed-in client.
 			if own := r.Cookies(); len(own) > 0 {
 				s.rememberKVMCookies(k.ID, own)
 			} else {
 				for name, val := range s.kvmCookies(k.ID) {
 					r.AddCookie(&http.Cookie{Name: name, Value: val})
 				}
 			}
 		},
 		FlushInterval: -1, // a video stream must not be buffered
 		ModifyResponse: func(resp *http.Response) error {
 			// The device's own framing and transport policy would stop it being embedded, or
 			// send a browser off to plain HTTP. Ham Lab is the origin now, and says what it
 			// serves; its own headers are added on the way out.
 			for _, h := range []string{"X-Frame-Options", "Content-Security-Policy", "Content-Security-Policy-Report-Only", "Strict-Transport-Security", "Referrer-Policy"} {
 				resp.Header.Del(h)
 			}
 			// A cookie the device sets by header is kept for Ham Lab's origin — so its Domain
 			// (the device's) goes, Secure follows the console rather than the device, and the
 			// hub remembers it too.
 			if set := resp.Cookies(); len(set) > 0 {
 				https := resp.Request != nil && resp.Request.Header.Get(kvmSecureHeader) == "1"
 				resp.Header.Del("Set-Cookie")
 				for _, c := range set {
 					c.Domain = ""
 					if c.Path == "" {
 						c.Path = "/"
 					}
 					c.Secure = c.Secure && https
 					if v := c.String(); v != "" {
 						resp.Header.Add("Set-Cookie", v)
 					}
 				}
 				s.rememberKVMCookies(k.ID, set)
 			}
 			// A redirect to the device's own root has to come back through the prefix.
 			if loc := resp.Header.Get("Location"); loc != "" {
 				if u, err := url.Parse(loc); err == nil && (u.Host == "" || u.Host == target.Host) && strings.HasPrefix(u.Path, "/") && !strings.HasPrefix(u.Path, prefix) {
 					u.Scheme, u.Host = "", ""
 					u.Path = prefix + u.Path
 					resp.Header.Set("Location", u.String())
 				}
 			}
 			return nil
 		},
 		ErrorHandler: func(w http.ResponseWriter, r *http.Request, err error) {
 			w.Header().Set("Content-Type", "application/json")
 			w.WriteHeader(http.StatusBadGateway)
 			_ = json.NewEncoder(w).Encode(map[string]string{"error": k.Name + " unreachable: " + err.Error()})
 		},
 	}
 	s.kvms.m[k.ID] = p
 	return p, nil
 }
 // kvmSecureHeader rides on the outbound request to say whether the console was reached
 // over TLS; the device ignores it.
 const kvmSecureHeader = "X-Hamlab-Secure"
 // kvmCookies is what the hub remembers of a device's own cookies — its login, in practice.
 func (s *Server) kvmCookies(id string) map[string]string {
 	var out map[string]string
 	s.store.View(func(st *store.State) {
 		if m := st.KVMCookies[id]; len(m) > 0 {
 			out = make(map[string]string, len(m))
 			for k, v := range m {
 				out[k] = v
 			}
 		}
 	})
 	return out
 }
 // rememberKVMCookies keeps the cookies a browser (or the device's own Set-Cookie) carries
 // for a KVM, so that a browser without them can be lent them. A cookie emptied or expired
 // by the device is forgotten. Written only when something changed: this runs on every
 // proxied request.
 func (s *Server) rememberKVMCookies(id string, cs []*http.Cookie) {
 	have := s.kvmCookies(id)
 	changed := false
 	for _, c := range cs {
 		if c.Name == "" || c.Name == cookieName || c.Name == kvmCookie {
 			continue
 		}
 		gone := c.Value == "" || c.MaxAge < 0 || (!c.Expires.IsZero() && c.Expires.Before(time.Now()))
 		if gone {
 			if _, ok := have[c.Name]; ok {
 				changed = true
 			}
 			continue
 		}
 		if have[c.Name] != c.Value {
 			changed = true
 		}
 	}
 	if !changed {
 		return
 	}
 	_ = s.store.Update(func(st *store.State) {
 		if st.KVMCookies == nil {
 			st.KVMCookies = map[string]map[string]string{}
 		}
 		m := st.KVMCookies[id]
 		if m == nil {
 			m = map[string]string{}
 		}
 		for _, c := range cs {
 			if c.Name == "" || c.Name == cookieName || c.Name == kvmCookie {
 				continue
 			}
 			if c.Value == "" || c.MaxAge < 0 || (!c.Expires.IsZero() && c.Expires.Before(time.Now())) {
 				delete(m, c.Name)
 			} else {
 				m[c.Name] = c.Value
 			}
 		}
 		if len(m) == 0 {
 			delete(st.KVMCookies, id)
 		} else {
 			st.KVMCookies[id] = m
 		}
 	})
 }
 // forgetKVMCookies drops a device's remembered login — when the device is removed.
 func (s *Server) forgetKVMCookies(id string) {
 	_ = s.store.Update(func(st *store.State) { delete(st.KVMCookies, id) })
 }
 // lendKVMCookies hands the browser the device cookies the hub remembers and it lacks, as
 // the device's own page would have set them: readable by its scripts, for Ham Lab's whole
 // origin, kept for half a year. The device's SPA checks for its token client-side before it
 // ever asks the server, so lending it only on the wire would still land on the login form.
 func (s *Server) lendKVMCookies(w http.ResponseWriter, r *http.Request, id string) {
 	for name, val := range s.kvmCookies(id) {
 		if _, err := r.Cookie(name); err == nil {
 			continue
 		}
 		http.SetCookie(w, &http.Cookie{Name: name, Value: val, Path: "/", MaxAge: 180 * 24 * 3600, Secure: secure(r), SameSite: http.SameSiteLaxMode})
 	}
 }
 // stripCookie removes named cookies from a request, so Ham Lab's session is never handed to
 // a device and the device never sees which KVM the browser is looking at.
 func stripCookie(r *http.Request, names ...string) {
 	cs := r.Cookies()
 	kept := cs[:0]
 	for _, c := range cs {
 		drop := false
 		for _, n := range names {
 			if c.Name == n {
 				drop = true
 			}
 		}
 		if !drop {
 			kept = append(kept, c)
 		}
 	}
 	r.Header.Del("Cookie")
 	for _, c := range kept {
 		r.AddCookie(c)
 	}
 }
 func (s *Server) forgetKVMProxy(id string) {
 	s.kvms.mu.Lock()
 	delete(s.kvms.m, id)
 	s.kvms.mu.Unlock()
 }
 // serveKVM is the entry point: /kvm/<id>/… . It also drops a cookie naming the device, so
 // that requests the page makes later by absolute path can be traced back to it.
 func (s *Server) serveKVM(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
 	k, ok := s.kvm(id)
 	if !ok {
 		jsonErr(w, 404, "no such KVM")
 		return
 	}
 	p, err := s.kvmProxy(k)
 	if err != nil {
 		jsonErr(w, 400, err.Error())
 		return
 	}
 	http.SetCookie(w, &http.Cookie{Name: kvmCookie, Value: id, Path: "/", HttpOnly: true, Secure: secure(r), SameSite: http.SameSiteLaxMode})
 	s.lendKVMCookies(w, r, id)
 	p.ServeHTTP(w, r)
 }
 // kvmFor works out which device a request belongs to when it arrived by absolute path
 // rather than through the prefix: the Referer names the page that asked, and a cookie
 // covers what has no Referer at all. Nothing device-specific: it never has to know which
 // paths a given KVM uses.
 func (s *Server) kvmFor(r *http.Request) (config.KVM, bool) {
 	if ref := r.Header.Get("Referer"); ref != "" {
 		if u, err := url.Parse(ref); err == nil {
 			if rest, ok := strings.CutPrefix(u.Path, kvmPrefix); ok {
 				if id, _, _ := strings.Cut(rest, "/"); id != "" {
 					if k, ok := s.kvm(id); ok {
 						return k, true
 					}
 				}
 			}
 		}
 		// asked for by a page that is not a KVM: not ours to answer
 		return config.KVM{}, false
 	}
 	if c, err := r.Cookie(kvmCookie); err == nil && c.Value != "" {
 		return s.kvm(c.Value)
 	}
 	return config.KVM{}, false
 }
 // kvmFallback answers a request Ham Lab does not serve itself, when it came from inside a
 // KVM's page. It returns false when the request has nothing to do with a KVM, and the
 // caller carries on with its own 404.
 func (s *Server) kvmFallback(w http.ResponseWriter, r *http.Request) bool {
 	k, ok := s.kvmFor(r)
 	if !ok {
 		return false
 	}
 	p, err := s.kvmProxy(k)
 	if err != nil {
 		return false
 	}
 	p.ServeHTTP(w, r)
 	return true
 }
 func machineLabel(m string) string {
 	if m == "" {
 		return "this host"
 	}
 	return m
 }
 // ---- the list ----
 func (s *Server) listKVMs(w http.ResponseWriter, _ *http.Request) {
 	out := []config.KVM{}
 	s.cfg.View(func(c *config.Config) { out = append(out, c.KVMs...) })
 	jsonOK(w, 200, out)
 }
 // kvmForMachine is the KVM attached to a machine, "" being the hub itself.
 func (s *Server) kvmForMachine(machine string) (config.KVM, bool) {
 	var found config.KVM
 	ok := false
 	s.cfg.View(func(c *config.Config) {
 		for _, k := range c.KVMs {
 			if k.Machine == machine {
 				found, ok = k, true
 			}
 		}
 	})
 	return found, ok
 }
 func (s *Server) addKVM(w http.ResponseWriter, r *http.Request) {
 	var body struct{ Name, URL, Machine string }
 	if err := json.NewDecoder(io.LimitReader(r.Body, 4096)).Decode(&body); err != nil {
 		jsonErr(w, 400, "bad request")
 		return
 	}
 	raw := strings.TrimSpace(body.URL)
 	if raw != "" && !strings.Contains(raw, "://") {
 		raw = "http://" + raw // what you would type: an address, not a scheme
 	}
 	u, err := url.Parse(raw)
 	if err != nil || (u.Scheme != "http" && u.Scheme != "https") || u.Host == "" {
 		jsonErr(w, 400, "address: http(s)://host, or just the host")
 		return
 	}
 	name := strings.TrimSpace(body.Name)
 	if name == "" {
 		name = u.Host
 	}
 	if len(name) > maxNameLen {
 		jsonErr(w, 400, "that name is too long")
 		return
 	}
 	b := make([]byte, 6)
 	if _, err := rand.Read(b); err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	machine := strings.TrimSpace(body.Machine)
 	k := config.KVM{ID: hex.EncodeToString(b), Name: name, URL: u.Scheme + "://" + u.Host + strings.TrimSuffix(u.Path, "/"), Machine: machine, Added: time.Now()}
 	// One per machine: a second one for the same host replaces the first rather than sitting
 	// beside it, since the tab shows the machine you are looking at.
 	var replaced string
 	if err := s.cfg.Update(func(c *config.Config) {
 		kept := c.KVMs[:0]
 		for _, old := range c.KVMs {
 			if old.Machine == machine {
 				replaced = old.ID
 				continue
 			}
 			kept = append(kept, old)
 		}
 		c.KVMs = append(kept, k)
 	}); err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	if replaced != "" {
 		s.forgetKVMProxy(replaced)
 	}
 	log.Printf("kvm %q added for %s (%s) by %s", k.Name, machineLabel(machine), k.URL, clientIP(r))
 	jsonOK(w, 201, k)
 }
 func (s *Server) removeKVM(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
 	found := false
 	if err := s.cfg.Update(func(c *config.Config) {
 		kept := c.KVMs[:0]
 		for _, k := range c.KVMs {
 			if k.ID == id {
 				found = true
 				continue
 			}
 			kept = append(kept, k)
 		}
 		c.KVMs = kept
 	}); err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	if !found {
 		jsonErr(w, 404, "no such KVM")
 		return
 	}
 	s.forgetKVMProxy(id)
 	s.forgetKVMCookies(id)
 	log.Printf("kvm %s removed by %s", id, clientIP(r))
 	w.WriteHeader(204)
 }
@@ -0,0 +1,123 @@
 package api
 import (
 	"net/http"
 	"net/http/httptest"
 	"net/url"
 	"path/filepath"
 	"strings"
 	"testing"
 	"hamlab/internal/auth"
 	"hamlab/internal/config"
 	"hamlab/internal/store"
 )
 // A KVM's own login is a cookie its page sets in one browser. The hub remembers it from the
 // requests that carry it, lends it to a browser that lacks it — both on the wire to the
 // device and back to the browser, since the device's page checks for it client-side — and
 // forgets it when the device says it is gone or the KVM is removed.
 func TestKVMCookiesAreRememberedAndLent(t *testing.T) {
 	var seen []string // the Cookie header the fake device received, per request
 	device := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		seen = append(seen, r.Header.Get("Cookie"))
 		if r.URL.Path == "/api/auth/login" {
 			// a device that sets its token by header, with its own domain and Secure
 			http.SetCookie(w, &http.Cookie{Name: "nano-kvm-token", Value: "tok-1", Domain: "10.0.0.217", Secure: true})
 		}
 		if r.URL.Path == "/api/auth/logout" {
 			http.SetCookie(w, &http.Cookie{Name: "nano-kvm-token", Value: "", MaxAge: -1})
 		}
 		w.WriteHeader(200)
 	}))
 	defer device.Close()
 	dir := t.TempDir()
 	cfg := config.Default(filepath.Join(dir, "config.json"))
 	if err := cfg.Update(func(c *config.Config) {
 		c.KVMs = []config.KVM{{ID: "abc123", Name: "test", URL: device.URL, Machine: ""}}
 	}); err != nil {
 		t.Fatal(err)
 	}
 	st, err := store.Open(filepath.Join(dir, "state.json"))
 	if err != nil {
 		t.Fatal(err)
 	}
 	s := &Server{cfg: cfg, store: st, sessions: auth.NewSessions(st)}
 	k, _ := s.kvm("abc123")
 	p, err := s.kvmProxy(k)
 	if err != nil {
 		t.Fatal(err)
 	}
 	ask := func(path string, cookies ...*http.Cookie) *httptest.ResponseRecorder {
 		r := httptest.NewRequest("GET", path, nil)
 		r.SetPathValue("id", "abc123")
 		for _, c := range cookies {
 			r.AddCookie(c)
 		}
 		w := httptest.NewRecorder()
 		if strings.HasPrefix(path, kvmPrefix) {
 			s.serveKVM(w, r)
 		} else {
 			p.ServeHTTP(w, r)
 		}
 		return w
 	}
 	// 1. the device sets its token by header: it reaches the browser without the device's
 	//    domain or Secure (plain-http console), and the hub remembers it
 	w := ask("/api/auth/login", &http.Cookie{Name: cookieName, Value: "hamlab-session"})
 	var set *http.Cookie
 	for _, c := range w.Result().Cookies() {
 		if c.Name == "nano-kvm-token" {
 			set = c
 		}
 	}
 	if set == nil || set.Domain != "" || set.Secure || set.Value != "tok-1" || set.Path != "/" {
 		t.Fatalf("rewritten Set-Cookie = %+v, want nano-kvm-token=tok-1 for / without Domain or Secure", set)
 	}
 	if got := s.kvmCookies("abc123"); got["nano-kvm-token"] != "tok-1" {
 		t.Fatalf("remembered = %v, want the token", got)
 	}
 	if strings.Contains(seen[0], "hamlab_session") {
 		t.Fatalf("the device was handed Ham Lab's session: %q", seen[0])
 	}
 	// 2. a browser that has the cookie teaches a newer value
 	ask("/api/vm/info", &http.Cookie{Name: "nano-kvm-token", Value: "tok-2"}, &http.Cookie{Name: cookieName, Value: "x"})
 	if got := s.kvmCookies("abc123"); got["nano-kvm-token"] != "tok-2" {
 		t.Fatalf("remembered = %v, want tok-2", got)
 	}
 	if seen[1] != "nano-kvm-token=tok-2" {
 		t.Fatalf("device saw %q, want the browser's own cookie only", seen[1])
 	}
 	// 3. a browser without it is lent it: on the wire, and back as a Set-Cookie it can read
 	w = ask(kvmPrefix+"abc123/", &http.Cookie{Name: cookieName, Value: "x"})
 	if seen[2] != "nano-kvm-token=tok-2" {
 		t.Fatalf("device saw %q from a browser with no cookie, want the remembered token", seen[2])
 	}
 	lent := false
 	for _, c := range w.Result().Cookies() {
 		if c.Name == "nano-kvm-token" && c.Value == "tok-2" && !c.HttpOnly && c.Path == "/" && c.MaxAge > 0 {
 			lent = true
 		}
 	}
 	if !lent {
 		t.Fatalf("browser was not lent the token: %v", w.Result().Cookies())
 	}
 	// 4. the device logging out forgets it
 	ask("/api/auth/logout", &http.Cookie{Name: "nano-kvm-token", Value: "tok-2"})
 	if got := s.kvmCookies("abc123"); got != nil {
 		t.Fatalf("after logout remembered = %v, want nothing", got)
 	}
 	// 5. removing the KVM clears whatever is left
 	s.rememberKVMCookies("abc123", []*http.Cookie{{Name: "nano-kvm-token", Value: "tok-3"}})
 	s.forgetKVMCookies("abc123")
 	if got := s.kvmCookies("abc123"); got != nil {
 		t.Fatalf("after remove remembered = %v, want nothing", got)
 	}
 	_ = url.Parse
 }
@@ -0,0 +1,123 @@
 package api
 // The Network tab: the machine's interfaces, how the internet looks from here, and the speed
 // test. One machine at a time, like every other tab — the test runs on the machine in the
 // switcher, so the number is that machine's line and not the hub's.
 import (
 	"encoding/json"
 	"log"
 	"net/http"
 	"time"
 	"hamlab/internal/network"
 	"hamlab/internal/sensors"
 	"hamlab/internal/store"
 )
 // networkOverview is everything the tab draws at once: the interfaces, the outside view, the
 // last test and the ones before it.
 func (s *Server) networkOverview(w http.ResponseWriter, _ *http.Request) {
 	run := s.net.Test.State()
 	jsonOK(w, 200, map[string]any{
 		"report":  network.Look(s.hostName()),
 		"outside": s.net.Outside(),
 		"run":     run,
 		"history": s.net.Test.History(),
 		"method":  network.Method,
 	})
 }
 func (s *Server) speedtestState(w http.ResponseWriter, _ *http.Request) {
 	jsonOK(w, 200, s.net.Test.State())
 }
 func (s *Server) speedtestStart(w http.ResponseWriter, r *http.Request) {
 	run, err := s.net.Test.Start()
 	if err != nil {
 		jsonErr(w, 409, err.Error())
 		return
 	}
 	log.Printf("speed test started on %s (by %s)", s.hostName(), clientIP(r))
 	jsonOK(w, 200, run)
 }
 func (s *Server) speedtestCancel(w http.ResponseWriter, r *http.Request) {
 	if !s.net.Test.Cancel() {
 		jsonErr(w, 404, "no speed test is running")
 		return
 	}
 	log.Printf("speed test cancelled on %s (by %s)", s.hostName(), clientIP(r))
 	w.WriteHeader(204)
 }
 // ---- keeping the results ----
 // savedNetwork is what survives a restart: the recent results, so the history is a history
 // rather than whatever has happened since the daemon last started.
 type savedNetwork struct {
 	Results []network.Result `json:"results,omitempty"`
 }
 func (s *Server) loadNetwork() []network.Result {
 	var saved savedNetwork
 	var raw json.RawMessage
 	s.store.View(func(st *store.State) { raw = st.Network })
 	if len(raw) > 0 {
 		if err := json.Unmarshal(raw, &saved); err != nil {
 			log.Printf("network: kept results unreadable (%v); starting a fresh history", err)
 		}
 	}
 	return saved.Results
 }
 // keepResult is called when a test finishes: append it to what is on disk.
 func (s *Server) keepResult(r network.Result) {
 	if r.Error != "" {
 		log.Printf("speed test on %s failed: %s", s.hostName(), r.Error)
 	} else {
 		log.Printf("speed test on %s: down %s, up %s, latency %.1f ms via %s",
 			s.hostName(), network.Mbits(r.DownMbps), network.Mbits(r.UpMbps), r.LatencyMs, r.Colo)
 	}
 	results := s.net.Test.History()
 	b, err := json.Marshal(savedNetwork{Results: results})
 	if err != nil {
 		return
 	}
 	if err := s.store.Update(func(st *store.State) { st.Network = b }); err != nil {
 		log.Printf("network: %v", err)
 	}
 }
 // ---- the speed test as sensors ----
 // netSensors publishes the last speed test as ordinary sensors, so the line can be a widget,
 // a graph, an alert or a Prometheus series like anything else Ham Lab measures. Each one's
 // timestamp is when the test ran, which is how a stale reading says so.
 type netSensors struct{ w *network.Watcher }
 func (n netSensors) Sensors() []sensors.Sensor {
 	var out []sensors.Sensor
 	if o := n.w.Outside(); o.IP != "" {
 		at := o.At.UnixMilli()
 		out = append(out, sensors.Sensor{ID: "net.internet.ip", Name: "Public address", Kind: "state", Value: o.IP, T: at})
 		if o.Colo != "" {
 			out = append(out, sensors.Sensor{ID: "net.internet.edge", Name: "Internet edge", Kind: "state", Value: o.Colo, T: at})
 		}
 	}
 	run := n.w.Test.State()
 	if run.Result == nil || run.Result.Error != "" {
 		return out
 	}
 	r := run.Result
 	at := r.At.UnixMilli()
 	return append(out,
 		sensors.Sensor{ID: "net.internet.download", Name: "Download", Kind: "rate", Unit: "Mbps", Value: r.DownMbps, T: at, Min: sensors.F(0)},
 		sensors.Sensor{ID: "net.internet.upload", Name: "Upload", Kind: "rate", Unit: "Mbps", Value: r.UpMbps, T: at, Min: sensors.F(0)},
 		sensors.Sensor{ID: "net.internet.latency", Name: "Latency", Kind: "duration", Unit: "ms", Value: r.LatencyMs, T: at, Min: sensors.F(0)},
 		sensors.Sensor{ID: "net.internet.jitter", Name: "Jitter", Kind: "duration", Unit: "ms", Value: r.JitterMs, T: at, Min: sensors.F(0)},
 		sensors.Sensor{ID: "net.internet.tested", Name: "Speed test age", Kind: "duration", Unit: "s", Value: round1(time.Since(r.At).Seconds()), T: at, Min: sensors.F(0)},
 	)
 }
 func round1(v float64) float64 { return float64(int64(v*10+0.5)) / 10 }
@@ -33,7 +33,10 @@ import (
 	"hamlab/internal/config"
 	"hamlab/internal/cron"
 	"hamlab/internal/files"
 	"hamlab/internal/hardware"
 	"hamlab/internal/incus"
 	"hamlab/internal/metrics"
 	"hamlab/internal/network"
 	"hamlab/internal/nodes"
 	"hamlab/internal/priv"
 	"hamlab/internal/sensors"
@@ -67,8 +70,13 @@ type Server struct {
 	tasks       *tasks.Manager
 	updates     *updates.Manager
 	storage     *storage.Manager
 	hardware    *hardware.Manager
 	space       *space.Manager
 	net         *network.Watcher
 	update      updateState
 	events      eventLog
 	kvms        kvmProxies
 	virt        *incus.Manager
 	fsLocal     files.FS
 	fsSudo      files.FS // nil without passwordless sudo
 	nodes       *nodes.Manager
@@ -85,7 +93,7 @@ type Server struct {
 func New(cfg *config.Config, sessions *auth.Sessions, mgr *term.Manager, col *metrics.Collector, st *store.Store, svc *services.Manager, cr *cron.Manager, tm *tasks.Manager, up *updates.Manager, sto *storage.Manager, fsLocal, fsSudo files.FS, nm *nodes.Manager, fingerprint string, ui fs.FS, version string) *Server {
 	s := &Server{cfg: cfg, sessions: sessions, mgr: mgr, metrics: col, sensors: sensors.NewRegistry(col), store: st, svc: svc, cron: cr, tasks: tm, updates: up, storage: sto, fsLocal: fsLocal, fsSudo: fsSudo, nodes: nm, fingerprint: fingerprint, ui: ui, version: version,
-		started: time.Now(), user: os.Getenv("USER")}
+		hardware: hardware.New(), started: time.Now(), user: os.Getenv("USER")}
 	self, err := os.Executable()
 	if err != nil {
 		self = "hamlab"
@@ -98,7 +106,13 @@ func New(cfg *config.Config, sessions *auth.Sessions, mgr *term.Manager, col *me
 	if s.user == "" {
 		s.user = "?"
 	}
 	s.virt = incus.New(cfg.DataDir(), s.user, version)
 	s.migrateUsers()
 	// The speed test and the outside view. Its history outlives a restart, and its last result
 	// is published as sensors, so the line is graphable and alertable like anything else.
 	s.net = network.NewWatcher(s.keepResult)
 	s.net.Test.Load(s.loadNetwork())
 	s.sensors.Add(netSensors{s.net})
 	s.sensors.Add(newStatusSensors(s))
 	s.alerts = newAlertEngine(s)
 	// The saved dashboard layout carries the sampling interval; apply it on boot.
@@ -136,8 +150,18 @@ func New(cfg *config.Config, sessions *auth.Sessions, mgr *term.Manager, col *me
 		}
 		go s.keepFleetInStep()
 	}
 	go s.watchGuests()
 	go s.watchForUpdates(context.Background())
 	go s.watchDockerEvents(context.Background())
 	s.mux = s.routes()
 	return s
 }
 // routes is the whole HTTP surface. A method of its own, rather than inline in New, so a
 // test can build the real router without the machinery behind it — which is what the guest
 // gate is checked against, since it decides a request by asking this router where it would go.
 func (s *Server) routes() *http.ServeMux {
 	m := http.NewServeMux()
 	m.HandleFunc("GET /healthz", func(w http.ResponseWriter, _ *http.Request) { w.WriteHeader(204) })
 	m.HandleFunc("POST /api/login", s.login)
@@ -184,11 +208,19 @@ func New(cfg *config.Config, sessions *auth.Sessions, mgr *term.Manager, col *me
 	m.HandleFunc("POST /api/docker/images/prune", s.auth(s.imagePrune))
 	m.HandleFunc("POST /api/docker/volumes/remove", s.auth(s.volumeRemove))
 	m.HandleFunc("POST /api/docker/volumes/prune", s.auth(s.volumePrune))
 	m.HandleFunc("POST /api/docker/networks/remove", s.auth(s.networkRemove))
 	m.HandleFunc("POST /api/docker/networks/prune", s.auth(s.networkPrune))
 	m.HandleFunc("GET /api/docker/stacks/{name}", s.auth(s.stackFile))
 	m.HandleFunc("PUT /api/docker/stacks/{name}", s.auth(s.stackSave))
 	m.HandleFunc("POST /api/docker/stacks/{name}/{action}", s.auth(s.stackAction))
 	m.HandleFunc("GET /api/docker/stacks/{name}/diff", s.auth(s.stackDiff))
 	m.HandleFunc("GET /api/docker/stacks/{name}/updates", s.auth(s.stackUpdates))
 	m.HandleFunc("GET /api/docker/updates", s.auth(s.machineUpdates))
 	m.HandleFunc("GET /api/docker/schedules", s.auth(s.dsList))
 	m.HandleFunc("PUT /api/docker/schedules", s.auth(s.dsPut))
 	m.HandleFunc("DELETE /api/docker/schedules/{id}", s.auth(s.dsDelete))
 	m.HandleFunc("POST /api/docker/schedules/{id}/run", s.auth(s.dsRun))
 	m.HandleFunc("GET /api/nodes/{node}/docker/updates", s.auth(s.machineUpdates))
 	// The same, asked of a node through this hub: the hub answers for the stacks whose files
 	// it holds (deployed to the node over tcp) and passes everything else to the node.
 	m.HandleFunc("GET /api/nodes/{node}/docker/stacks", s.auth(s.stacksList))
@@ -204,7 +236,26 @@ func New(cfg *config.Config, sessions *auth.Sessions, mgr *term.Manager, col *me
 	m.HandleFunc("GET /api/nodes/{node}/tasks/{id}/stream", s.auth(s.taskOnHub(s.streamTask)))
 	m.HandleFunc("POST /api/docker/containers/{id}/{action}", s.auth(s.dockerAction))
 	m.HandleFunc("GET /api/docker/containers/{id}/logs", s.auth(s.dockerLogs))
 	m.HandleFunc("GET /api/docker/containers/{id}/logs/tail", s.auth(s.dockerTail))
 	m.HandleFunc("GET /api/docker/containers/{id}/inspect", s.auth(s.dockerInspect))
 	m.HandleFunc("GET /api/docker/containers/{id}/health", s.auth(s.dockerHealth))
 	m.HandleFunc("GET /api/docker/containers/{id}/events", s.auth(s.dockerEvents))
 	m.HandleFunc("POST /api/docker/containers/{id}/exec", s.auth(s.dockerExec))
 	m.HandleFunc("POST /api/docker/containers/{id}/limits", s.auth(s.dockerUpdate))
 	m.HandleFunc("GET /api/virt", s.auth(s.virtOverview))
 	m.HandleFunc("POST /api/virt/install", s.auth(s.virtInstall))
 	m.HandleFunc("GET /api/virt/images/remote", s.auth(s.virtRemoteImages))
 	m.HandleFunc("POST /api/virt/images/download", s.auth(s.virtImageDownload))
 	m.HandleFunc("POST /api/virt/images/import", s.auth(s.virtImageImport))
 	m.HandleFunc("POST /api/virt/images/publish", s.auth(s.virtImagePublish))
 	m.HandleFunc("POST /api/virt/images/delete", s.auth(s.virtImageDelete))
 	m.HandleFunc("POST /api/virt/instances", s.auth(s.virtCreate))
 	m.HandleFunc("GET /api/virt/instances/{name}", s.auth(s.virtInstance))
 	m.HandleFunc("POST /api/virt/instances/{name}/rename", s.auth(s.virtRename))
 	m.HandleFunc("PUT /api/virt/instances/{name}/config", s.auth(s.virtConfig))
 	m.HandleFunc("POST /api/virt/instances/{name}/snapshots", s.auth(s.virtSnapshotCreate))
 	m.HandleFunc("POST /api/virt/instances/{name}/snapshots/{snap}/{action}", s.auth(s.virtSnapshotAction))
 	m.HandleFunc("POST /api/virt/instances/{name}/{action}", s.auth(s.virtAction))
 	m.HandleFunc("GET /api/alerts", s.auth(s.alertsGet))
 	m.HandleFunc("PUT /api/alerts", s.auth(s.alertsPut))
 	m.HandleFunc("POST /api/alerts/test", s.auth(s.alertsTest))
@@ -245,6 +296,15 @@ func New(cfg *config.Config, sessions *auth.Sessions, mgr *term.Manager, col *me
 	m.HandleFunc("POST /api/files/rename", s.auth(s.filesRename))
 	m.HandleFunc("POST /api/files/delete", s.auth(s.filesDelete))
 	m.HandleFunc("POST /api/files/chmod", s.auth(s.filesChmod))
 	m.HandleFunc("GET /api/hardware", s.auth(s.hardwareReport))
 	m.HandleFunc("GET /api/network", s.auth(s.networkOverview))
 	m.HandleFunc("GET /api/vpn", s.auth(s.vpnReport))
 	m.HandleFunc("POST /api/vpn/{iface}/peers", s.auth(s.addVPNPeer))
 	m.HandleFunc("PATCH /api/vpn/{iface}/peers/{section}", s.auth(s.editVPNPeer))
 	m.HandleFunc("DELETE /api/vpn/{iface}/peers/{section}", s.auth(s.removeVPNPeer))
 	m.HandleFunc("GET /api/network/speedtest", s.auth(s.speedtestState))
 	m.HandleFunc("POST /api/network/speedtest", s.auth(s.speedtestStart))
 	m.HandleFunc("POST /api/network/speedtest/cancel", s.auth(s.speedtestCancel))
 	m.HandleFunc("GET /api/storage", s.auth(s.storageOverview))
 	m.HandleFunc("GET /api/storage/smart/{dev}", s.auth(s.storageSmart))
 	m.HandleFunc("POST /api/storage/smart/{dev}/test/{kind}", s.auth(s.storageSmartTest))
@@ -252,9 +312,9 @@ func New(cfg *config.Config, sessions *auth.Sessions, mgr *term.Manager, col *me
 	m.HandleFunc("POST /api/storage/zfs/scrub/{pool}/{action}", s.auth(s.zfsScrub))
 	m.HandleFunc("POST /api/storage/zfs/snapshot", s.auth(s.zfsSnapshot))
 	m.HandleFunc("GET /api/zfs", s.auth(s.zfsOverview))
 	m.HandleFunc("GET /api/zfs/arc", s.auth(s.zfsARC))
 	m.HandleFunc("PUT /api/zfs/schedules", s.auth(s.zfsPolicyPut))
 	m.HandleFunc("DELETE /api/zfs/schedules/{id}", s.auth(s.zfsPolicyDelete))
 	m.HandleFunc("GET /api/zfs/schedules/{id}/expired", s.auth(s.zfsPreview))
 	m.HandleFunc("POST /api/zfs/schedules/{id}/run", s.auth(s.zfsRunNow))
 	m.HandleFunc("POST /api/zfs/snapshots/delete", s.auth(s.zfsSnapshotDelete))
 	m.HandleFunc("GET /api/space", s.auth(s.spaceOverview))
@@ -271,9 +331,13 @@ func New(cfg *config.Config, sessions *auth.Sessions, mgr *term.Manager, col *me
 	m.HandleFunc("POST /api/users", s.auth(s.addUser))
 	m.HandleFunc("DELETE /api/users/{name}", s.auth(s.removeUser))
 	m.HandleFunc("POST /api/users/{name}/password", s.auth(s.resetPassword))
 	m.HandleFunc("GET /api/guests", s.auth(s.listGuests))
 	m.HandleFunc("POST /api/guests", s.auth(s.addGuest))
 	m.HandleFunc("DELETE /api/guests/{name}", s.auth(s.revokeGuest))
 	m.HandleFunc("GET /api/sessions", s.auth(s.listSessions))
 	m.HandleFunc("DELETE /api/sessions/{id}", s.auth(s.revokeSession))
 	m.HandleFunc("POST /api/sessions/revoke-others", s.auth(s.revokeOtherSessions))
 	m.HandleFunc("POST /api/sessions/revoke-idle", s.auth(s.revokeIdleSessions))
 	m.HandleFunc("GET /api/prefs/ui", s.auth(s.getUIPrefs))
 	m.HandleFunc("PUT /api/prefs/ui", s.auth(s.putUIPrefs))
 	// Go runtime profiles, behind the login, for chasing memory/CPU (bug 14).
@@ -285,10 +349,22 @@ func New(cfg *config.Config, sessions *auth.Sessions, mgr *term.Manager, col *me
 	m.HandleFunc("PUT /api/prefs/sampler", s.auth(s.putSampler))
 	m.HandleFunc("GET /api/prefs/history", s.auth(s.getHistoryWindow))
 	m.HandleFunc("PUT /api/prefs/history", s.auth(s.putHistoryWindow))
-	m.HandleFunc("/api/", func(w http.ResponseWriter, _ *http.Request) { jsonErr(w, 404, "not found") })
+	m.HandleFunc("GET /api/kvms", s.auth(s.listKVMs))
 	m.HandleFunc("POST /api/kvms", s.auth(s.addKVM))
 	m.HandleFunc("DELETE /api/kvms/{id}", s.auth(s.removeKVM))
 	m.HandleFunc("/kvm/{id}/{rest...}", s.auth(s.serveKVM))
 	m.HandleFunc("/kvm/{id}", s.auth(s.serveKVM))
 	// A path Ham Lab does not serve may still belong to a KVM's page, which asks for its
 	// own assets by absolute path. Ham Lab's own routes matched first, so this can only
 	// catch what would otherwise have been a 404.
 	m.HandleFunc("/api/", s.auth(func(w http.ResponseWriter, r *http.Request) {
 		if s.kvmFallback(w, r) {
 			return
 		}
 		jsonErr(w, 404, "not found")
 	}))
 	m.HandleFunc("/", s.static)
-	s.mux = m
+	return m
 	return s
 }
 // sensorPath reports whether a path belongs to the read-only sensor API.
@@ -308,9 +384,16 @@ func sensorPath(p string) bool {
 func (s *Server) Handler() http.Handler {
 	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		h := w.Header()
-		// The console must never be framed. An embedded widget exists to be: it is a page
+		// The console must never be framed. Two exceptions: an embedded widget exists to be
-		// that can only read sensors, and only with a read-only token in its URL.
+		// (a page that can only read sensors, and only with a read-only token in its URL),
-		if r.URL.Path != "/embed/widget" {
+		// and a proxied KVM is framed by the console itself — 'none' forbids that even from
 		// the same origin, so those paths say 'self' instead of opening up.
 		switch {
 		case r.URL.Path == "/embed/widget":
 		case strings.HasPrefix(r.URL.Path, kvmPrefix):
 			h.Set("X-Frame-Options", "SAMEORIGIN")
 			h.Set("Content-Security-Policy", "frame-ancestors 'self'")
 		default:
 			h.Set("X-Frame-Options", "DENY")
 			h.Set("Content-Security-Policy", "frame-ancestors 'none'")
 		}
@@ -356,14 +439,56 @@ func (s *Server) setCookie(w http.ResponseWriter, r *http.Request, tok string, m
 	})
 }
 // signedIn reports whether a request carries a live browser session, without the
 // side effects of the auth middleware.
 func (s *Server) signedIn(r *http.Request) bool {
 	c, err := r.Cookie(cookieName)
 	if err != nil {
 		return false
 	}
 	_, ok := s.sessions.Lookup(c.Value, clientKey(r))
 	return ok
 }
 // cookieActor resolves the session cookie to who is calling. answered means the caller has
 // already been sent an error and the handler must not run: the account behind a live session
 // has gone away, or a guest login's time is up — so a console left open on a visitor's phone
 // stops at the expiry, not at whatever they click next.
 func (s *Server) cookieActor(w http.ResponseWriter, r *http.Request) (a actor, ok, answered bool) {
 	c, err := r.Cookie(cookieName)
 	if err != nil {
 		return actor{}, false, false
 	}
 	ses, live := s.sessions.Lookup(c.Value, clientKey(r))
 	if !live {
 		return actor{}, false, false
 	}
 	var u config.User
 	var found bool
 	s.cfg.View(func(cf *config.Config) { u, found = cf.User(ses.User) })
 	if !found || u.Expired(time.Now()) {
 		s.sessions.Delete(c.Value)
 		s.setCookie(w, r, "", -1)
 		if found {
 			jsonErr(w, 401, "this guest login has expired")
 		} else {
 			jsonErr(w, 401, "that account no longer exists")
 		}
 		return actor{}, false, true
 	}
 	return actor{User: ses.User, Session: ses.ID, token: c.Value, Guest: u.Guest, Expires: u.Expires}, true, false
 }
 // auth accepts the browser session cookie or a bearer API token issued to a hub.
 func (s *Server) auth(next http.HandlerFunc) http.HandlerFunc {
 	return func(w http.ResponseWriter, r *http.Request) {
-		if c, err := r.Cookie(cookieName); err == nil {
+		a, ok, answered := s.cookieActor(w, r)
-			if ses, ok := s.sessions.Lookup(c.Value, clientKey(r)); ok {
+		if answered {
-				next(w, withActor(r, actor{User: ses.User, Session: ses.ID, token: c.Value}))
+			return
-				return
+		}
-			}
+		if ok {
 			next(w, withActor(r, a))
 			return
 		}
 		if h := r.Header.Get("Authorization"); strings.HasPrefix(h, "Bearer ") {
 			var hashes, names []string
@@ -397,11 +522,11 @@ func (s *Server) sensorAuth(next http.HandlerFunc) http.HandlerFunc {
 	return func(w http.ResponseWriter, r *http.Request) {
 		ok, trusted := false, false
 		var patterns []string
-		if c, err := r.Cookie(cookieName); err == nil {
+		if a, live, answered := s.cookieActor(w, r); answered {
-			if ses, valid := s.sessions.Lookup(c.Value, clientKey(r)); valid {
+			return
-				ok, trusted = true, true
+		} else if live {
-				r = withActor(r, actor{User: ses.User, Session: ses.ID, token: c.Value})
+			ok, trusted = true, true
-			}
+			r = withActor(r, a)
 		}
 		if !ok {
 			presented, fromURL := strings.TrimPrefix(r.Header.Get("Authorization"), "Bearer "), false
@@ -870,11 +995,11 @@ func (s *Server) login(w http.ResponseWriter, r *http.Request) {
 // name is allowed only while there is exactly one account — what a client from before
 // user names sends, and what a hub pairing with a one-user node may leave out.
 func (s *Server) checkLogin(username, password string) (string, bool) {
-	var users []config.User
+	var users, real []config.User
-	s.cfg.View(func(c *config.Config) { users = append(users, c.Users...) })
+	s.cfg.View(func(c *config.Config) { users, real = append(users, c.Users...), c.Accounts() })
 	username = strings.TrimSpace(username)
-	if username == "" && len(users) == 1 {
+	if username == "" && len(real) == 1 { // guest logins are not the account an empty name means
-		username = users[0].Name
+		username = real[0].Name
 	}
 	if username == "" {
 		auth.CheckUser(nil, "", password) // take as long as a real check
@@ -1009,8 +1134,14 @@ func (s *Server) me(w http.ResponseWriter, r *http.Request) {
 			addr = h // the address this console listens on, for the machine menu
 		}
 	})
 	me := actorOf(r)
 	var guestUntil any
 	if me.Guest {
 		guestUntil = me.Expires
 	}
 	jsonOK(w, 200, map[string]any{
-		"host": s.hostName(), "user": s.user, "account": actorOf(r).User, "version": s.version,
+		"host": s.hostName(), "user": s.user, "account": me.User, "version": s.version,
 		"guest": me.Guest, "guest_until": guestUntil,
 		"load": load, "uptime": uptime, "service_started": s.started,
 		"reboot_needed": rebootErr == nil, "fingerprint": s.fingerprint, "arch": runtime.GOARCH, "addr": addr,
 		"update_available": updateVersion,
@@ -2446,6 +2577,12 @@ func (s *Server) static(w http.ResponseWriter, r *http.Request) {
 			return
 		}
 	}
 	// A path with no file of ours behind it may belong to a KVM's page, which asks for its
 	// own assets by absolute path. Checked here rather than in the router because this
 	// handler is the catch-all — and gated on a session, since this handler is public.
 	if s.signedIn(r) && s.kvmFallback(w, r) {
 		return
 	}
 	// SPA fallback
 	w.Header().Set("Cache-Control", "no-cache")
 	http.ServeFileFS(w, r, s.ui, "index.html")
@@ -31,7 +31,7 @@ type spaceTarget struct {
 	Layout string      `json:"layout,omitempty"` // pool: raidz2, mirror, …
 	Health string      `json:"health,omitempty"` // pool: ONLINE, DEGRADED, …
 	Size   uint64      `json:"size"`
-	Used   uint64      `json:"used,omitempty"` // pool: allocated
+	Used   uint64      `json:"used,omitempty"`  // pool: allocated
 	Disks  []string    `json:"disks,omitempty"` // pool: its member disks
 	Roots  []string    `json:"roots"`           // the mountpoints a scan walks
 	Scan   *space.Scan `json:"scan,omitempty"`  // the last scan, if any
@@ -12,6 +12,12 @@ package api
 // relative paths, the .env beside them and the labels all stay exactly as they were
 // deployed — else on the host.
 //
 // Those mounts are also written down. An orchestrator can be removed, and when it is, the
 // containers it left behind still carry labels naming paths inside a filesystem that no
 // longer exists; without a record of where that filesystem was mounted, every stack on the
 // machine would become unfindable at once. A remembered mount resolves with no container to
 // run in, which is right: compose runs on the host against the host's copy of the file.
 //
 // A stack can also be deployed *from another machine*: Dockhand on the hub drives a node's
 // Docker over tcp, so the node runs the containers and the hub holds the files. The node
 // alone can only list such a stack read-only. The hub, asked about one of its nodes, fills
@@ -39,7 +45,9 @@ import (
 	"sync"
 	"time"
 	"hamlab/internal/compose"
 	"hamlab/internal/priv"
 	"hamlab/internal/store"
 )
 type stack struct {
@@ -89,6 +97,100 @@ func stackMounts(ctx context.Context) []mountMap {
 	return maps
 }
 // mountsFor is stackMounts plus memory. An orchestrator that deployed everything here and
 // has since been removed takes the only record of where its files really are with it: the
 // containers it left behind are labelled with paths inside *its* filesystem. So every map
 // seen is written down, and a path that no live container explains is looked for in what
 // was seen before. A remembered map resolves with no `via`, which is correct — the
 // container is gone, so compose runs on the host, against the host's copy of the file.
 func mountsFor(ctx context.Context, db *store.Store) []mountMap {
 	live := stackMounts(ctx)
 	rememberMounts(db, live)
 	return append(live, recallMounts(db, live)...)
 }
 // mountsSeen is the same without writing anything down. The daemon owns state.json; a cron
 // job that wrote to it would be clobbered by the next thing the daemon saved, so out there
 // the record is read and never added to.
 func mountsSeen(ctx context.Context, db *store.Store) []mountMap {
 	live := stackMounts(ctx)
 	return append(live, recallMounts(db, live)...)
 }
 // rememberMounts folds what is mounted now into what was mounted before, and forgets a map
 // whose source has gone from disk — that one is not coming back, and keeping it would make
 // a stack look editable when its file no longer exists.
 //
 // This is on the path of every stack listing, which the Docker tab polls, so it reads first
 // and only writes when the set actually changed.
 func rememberMounts(db *store.Store, live []mountMap) {
 	if len(live) == 0 {
 		return // docker is down or answered nothing; that is not evidence of anything
 	}
 	key := func(c, d string) string { return c + "\x00" + d }
 	var have []store.RememberedMount
 	db.View(func(st *store.State) { have = append([]store.RememberedMount(nil), st.Mounts...) })
 	at := map[string]int{}
 	for i, m := range have {
 		at[key(m.Container, m.Dest)] = i
 	}
 	now, add, moved := time.Now(), 0, false
 	for _, m := range live {
 		i, ok := at[key(m.container, m.dest)]
 		if !ok {
 			add++
 			continue
 		}
 		if have[i].Src != m.src {
 			moved = true
 		}
 	}
 	// Pruning is the expensive half — a stat, or a sudo test(1), per remembered map — so it
 	// only happens when something else has already made a write necessary.
 	if add == 0 && !moved {
 		return
 	}
 	_ = db.Update(func(st *store.State) {
 		at := map[string]int{}
 		for i, m := range st.Mounts {
 			at[key(m.Container, m.Dest)] = i
 		}
 		for _, m := range live {
 			if i, ok := at[key(m.container, m.dest)]; ok {
 				st.Mounts[i].Src, st.Mounts[i].Seen = m.src, now
 				continue
 			}
 			st.Mounts = append(st.Mounts, store.RememberedMount{Container: m.container, Dest: m.dest, Src: m.src, Seen: now})
 		}
 		kept := st.Mounts[:0]
 		for _, m := range st.Mounts {
 			if readable(m.Src) {
 				kept = append(kept, m)
 			}
 		}
 		st.Mounts = kept
 	})
 }
 // recallMounts is what was seen before and is not running now.
 func recallMounts(db *store.Store, live []mountMap) []mountMap {
 	var remembered []store.RememberedMount
 	db.View(func(st *store.State) { remembered = append([]store.RememberedMount(nil), st.Mounts...) })
 	now := map[string]bool{}
 	for _, m := range live {
 		now[m.container+"\x00"+m.dest] = true
 	}
 	var out []mountMap
 	for _, m := range remembered {
 		if now[m.Container+"\x00"+m.Dest] {
 			continue
 		}
 		out = append(out, mountMap{"", filepath.Clean(m.Dest), filepath.Clean(m.Src)}) // no container: run compose here
 	}
 	sort.Slice(out, func(i, j int) bool { return len(out[i].dest) > len(out[j].dest) })
 	return out
 }
 func readable(p string) bool {
 	if _, err := os.Stat(p); err == nil {
 		return true
@@ -172,7 +274,7 @@ func (s *Server) findStacks(ctx context.Context) ([]stack, error) {
 			}
 		}
 	}
-	maps := stackMounts(ctx)
+	maps := mountsFor(ctx, s.store)
 	stacks := []stack{}
 	for _, st := range by {
 		sort.Strings(st.Containers)
@@ -353,7 +455,7 @@ func (s *Server) hubStacks(ctx context.Context, node string) ([]stack, error) {
 			continue
 		}
 		if remote == "" {
-			remote, maps = s.nodeDocker(ctx, node), stackMounts(ctx)
+			remote, maps = s.nodeDocker(ctx, node), mountsFor(ctx, s.store)
 		}
 		if remote == "-" {
 			continue
@@ -511,24 +613,13 @@ func (s *Server) stackFile(w http.ResponseWriter, r *http.Request) {
 // composeCmd is `docker compose …` for this stack, run where its files make sense and
 // pointed at the Docker that runs its containers. files are given as compose will see them.
 func composeCmd(ctx context.Context, st stack, files []string, args ...string) *exec.Cmd {
-	tool := dockerTool()
+	return compose.Cmd(ctx, specOf(st, files), args...)
-	base := []string{"compose", "-p", st.Name}
+}
-	for _, f := range files {
+
-		base = append(base, "-f", f)
+// specOf is a stack as the compose runner needs it — and as a scheduled update stores it,
-	}
+// so the cron job and the console invoke compose identically.
-	if st.Via != "" {
+func specOf(st stack, files []string) compose.Spec {
-		pre := []string{"exec"}
+	return compose.Spec{Project: st.Name, Files: files, Via: st.Via, Remote: st.Remote, Tool: dockerTool()}
 		if st.Remote != "" {
 			pre = append(pre, "-e", "DOCKER_HOST="+st.Remote)
 		}
 		return exec.CommandContext(ctx, tool, append(append(append(pre, st.Via, "docker"), base...), args...)...)
 	}
 	cmd := exec.CommandContext(ctx, tool, append(base, args...)...)
 	cmd.Dir = filepath.Dir(files[0])
 	if st.Remote != "" {
 		cmd.Env = append(os.Environ(), "DOCKER_HOST="+st.Remote)
 	}
 	return cmd
 }
 // seenFiles is the stack's file list as compose runs it, with file i replaced by a candidate.
@@ -734,6 +825,12 @@ func (s *Server) stackAction(w http.ResponseWriter, r *http.Request) {
 		script = base("up -d --remove-orphans") + remember
 	case "update":
 		script = base("pull") + " && " + base("up -d --remove-orphans") + remember
 	case "recreate":
 		// The way out of a container Docker will no longer start — most often one pinned to a
 		// network that has since been recreated, which answers "network … not found" to every
 		// ordinary deploy. `up -d` leaves such a container alone because its definition has
 		// not changed; only --force-recreate builds it again.
 		script = base("up -d --force-recreate --remove-orphans") + remember
 	case "restart":
 		script = base("restart")
 	case "stop":
@@ -741,7 +838,7 @@ func (s *Server) stackAction(w http.ResponseWriter, r *http.Request) {
 	case "down": // remove its containers and networks; its files, volumes and bind-mounted data stay
 		script = base("down --remove-orphans")
 	default:
-		jsonErr(w, 400, "deploy, update, restart, stop or down")
+		jsonErr(w, 400, "deploy, update, recreate, restart, stop or down")
 		return
 	}
 	if t, busy := s.tasks.RunningOf("stack:" + st.Name); busy {
@@ -788,6 +885,103 @@ var (
 func stackUpdatesForget(key string) { updMu.Lock(); delete(updCache, key); updMu.Unlock() }
 // checkImages asks the registry whether each image is still the one we hold. remote is a
 // DOCKER_HOST for a machine whose images live elsewhere, "" for this one. Each image is one
 // registry request, which Docker Hub counts — so this is only ever called when asked for.
 func checkImages(ctx context.Context, images []string, remote string) []imageCheck {
 	local := func(args ...string) ([]byte, error) { // the Docker that holds the images
 		if remote != "" {
 			args = append([]string{"-H", remote}, args...)
 		}
 		return dockerRun(ctx, args...)
 	}
 	res := []imageCheck{}
 	for _, img := range images {
 		ic := imageCheck{Image: img, State: "unknown"}
 		out, err := local("image", "inspect", "-f", "{{json .RepoDigests}}", img)
 		var have []string
 		if err != nil || json.Unmarshal(bytes.TrimSpace(out), &have) != nil || len(have) == 0 {
 			ic.Detail = "built or loaded here, not pulled from a registry"
 			res = append(res, ic)
 			continue
 		}
 		rem, err := dockerRun(ctx, "buildx", "imagetools", "inspect", img, "--format", "{{.Manifest.Digest}}")
 		digest := strings.TrimSpace(string(rem))
 		if i := strings.LastIndex(digest, "\n"); i >= 0 {
 			digest = digest[i+1:]
 		}
 		if err != nil || !strings.HasPrefix(digest, "sha256:") {
 			ic.Detail = "the registry didn't answer"
 			if err != nil {
 				ic.Detail += ": " + err.Error()
 			}
 			res = append(res, ic)
 			continue
 		}
 		ic.State = "update"
 		for _, h := range have {
 			if strings.HasSuffix(h, "@"+digest) {
 				ic.State = "current"
 			}
 		}
 		res = append(res, ic)
 	}
 	return res
 }
 // machineUpdates checks every image on this machine in one pass, so the container list can
 // be coloured and the fleet can be asked "what is behind?" without walking stack by stack.
 func (s *Server) machineUpdates(w http.ResponseWriter, r *http.Request) {
 	node := r.PathValue("node")
 	ctx, cancel := context.WithTimeout(r.Context(), 10*time.Minute)
 	defer cancel()
 	stacks, err := s.findStacks(ctx)
 	if node != "" {
 		stacks, err = s.hubStacks(ctx, node)
 	}
 	if err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	// one entry per image, however many stacks use it
 	type want struct{ remote string }
 	imgs := map[string]want{}
 	byImage := map[string][]string{} // image -> stacks
 	for _, st := range stacks {
 		for _, img := range st.Images {
 			imgs[img] = want{remote: st.Remote}
 			byImage[img] = append(byImage[img], st.Name)
 		}
 	}
 	refresh := r.URL.Query().Get("refresh") == "1"
 	key := "machine:" + node
 	updMu.Lock()
 	c, ok := updCache[key]
 	updMu.Unlock()
 	if ok && time.Since(c.at) < 6*time.Hour && !refresh {
 		jsonOK(w, 200, map[string]any{"images": c.res, "checked": c.at.UnixMilli(), "stacks": byImage})
 		return
 	}
 	if !refresh {
 		jsonOK(w, 200, map[string]any{"images": []imageCheck{}, "checked": 0, "stacks": byImage})
 		return
 	}
 	var list []string
 	remote := ""
 	for img, wnt := range imgs {
 		list = append(list, img)
 		if wnt.remote != "" {
 			remote = wnt.remote
 		}
 	}
 	sort.Strings(list)
 	res := checkImages(ctx, list, remote)
 	updMu.Lock()
 	updCache[key] = updAnswer{time.Now(), res}
 	updMu.Unlock()
 	jsonOK(w, 200, map[string]any{"images": res, "checked": time.Now().UnixMilli(), "stacks": byImage})
 }
 // stackUpdates asks each image's registry for the digest its tag points at now and compares
 // it with the digest that was pulled. It is asked for, never polled: every answer is a
 // registry request, and Docker Hub counts those. Answers are kept for six hours.
@@ -816,43 +1010,7 @@ func (s *Server) stackUpdates(w http.ResponseWriter, r *http.Request) {
 		jsonOK(w, 200, map[string]any{"images": []imageCheck{}, "checked": 0})
 		return
 	}
-	local := func(args ...string) ([]byte, error) { // the Docker that holds the images
+	res := checkImages(ctx, st.Images, st.Remote)
 		if st.Remote != "" {
 			args = append([]string{"-H", st.Remote}, args...)
 		}
 		return dockerRun(ctx, args...)
 	}
 	res := []imageCheck{}
 	for _, img := range st.Images {
 		ic := imageCheck{Image: img, State: "unknown"}
 		out, err := local("image", "inspect", "-f", "{{json .RepoDigests}}", img)
 		var have []string
 		if err != nil || json.Unmarshal(bytes.TrimSpace(out), &have) != nil || len(have) == 0 {
 			ic.Detail = "built or loaded here, not pulled from a registry"
 			res = append(res, ic)
 			continue
 		}
 		remote, err := dockerRun(ctx, "buildx", "imagetools", "inspect", img, "--format", "{{.Manifest.Digest}}")
 		digest := strings.TrimSpace(string(remote))
 		if i := strings.LastIndex(digest, "\n"); i >= 0 {
 			digest = digest[i+1:]
 		}
 		if err != nil || !strings.HasPrefix(digest, "sha256:") {
 			ic.Detail = "the registry didn't answer"
 			if err != nil {
 				ic.Detail += ": " + err.Error()
 			}
 			res = append(res, ic)
 			continue
 		}
 		ic.State = "update"
 		for _, h := range have {
 			if strings.HasSuffix(h, "@"+digest) {
 				ic.State = "current"
 			}
 		}
 		res = append(res, ic)
 	}
 	updMu.Lock()
 	updCache[key] = updAnswer{time.Now(), res}
 	updMu.Unlock()
@@ -0,0 +1,121 @@
 package api
 import (
 	"context"
 	"os"
 	"path/filepath"
 	"testing"
 	"hamlab/internal/compose"
 	"hamlab/internal/store"
 )
 // A stack deployed by an orchestrator running in a container records the path its compose
 // file had inside that container. Resolving it has to work both while the orchestrator is
 // running — compose is then run inside it — and after it has been removed, where the same
 // mount is only remembered and compose has to run on the host instead.
 func TestResolveThroughMounts(t *testing.T) {
 	root := t.TempDir()
 	dir := filepath.Join(root, "orch", "stacks", "app")
 	if err := os.MkdirAll(dir, 0o755); err != nil {
 		t.Fatal(err)
 	}
 	file := filepath.Join(dir, "compose.yaml")
 	if err := os.WriteFile(file, []byte("services: {}\n"), 0o644); err != nil {
 		t.Fatal(err)
 	}
 	inner := "/app/data/stacks/app/compose.yaml"
 	for _, tc := range []struct {
 		name, container, wantVia string
 	}{
 		{"orchestrator running", "orch", "orch"},
 		{"orchestrator removed", "", ""},
 	} {
 		t.Run(tc.name, func(t *testing.T) {
 			st := stack{Name: "app", File: inner, Files: []string{inner}}
 			st.resolve([]mountMap{{tc.container, "/app/data", filepath.Join(root, "orch")}})
 			if !st.Editable {
 				t.Fatalf("not editable: %s", st.Note)
 			}
 			if st.HostFile != file {
 				t.Errorf("host file = %q, want %q", st.HostFile, file)
 			}
 			if st.Via != tc.wantVia {
 				t.Errorf("via = %q, want %q", st.Via, tc.wantVia)
 			}
 		})
 	}
 	// Nothing explains the path: the stack is listed, but read-only and says why.
 	st := stack{Name: "app", File: inner, Files: []string{inner}}
 	st.resolve(nil)
 	if st.Editable || st.Note == "" {
 		t.Errorf("unresolvable stack should be read-only with a note, got editable=%v note=%q", st.Editable, st.Note)
 	}
 }
 // Every file of a multi-file stack must resolve through the same place. Redeploying from
 // the first file alone would silently drop whatever the override set.
 func TestResolveRefusesSplitStack(t *testing.T) {
 	root := t.TempDir()
 	dir := filepath.Join(root, "orch", "stacks", "app")
 	if err := os.MkdirAll(dir, 0o755); err != nil {
 		t.Fatal(err)
 	}
 	if err := os.WriteFile(filepath.Join(dir, "compose.yaml"), []byte("services: {}\n"), 0o644); err != nil {
 		t.Fatal(err)
 	}
 	st := stack{Name: "app", File: "/app/data/stacks/app/compose.yaml", Files: []string{
 		"/app/data/stacks/app/compose.yaml",
 		"/app/data/stacks/app/override.yaml", // never written: not on disk
 	}}
 	st.resolve([]mountMap{{"orch", "/app/data", filepath.Join(root, "orch")}})
 	if st.Editable {
 		t.Error("a stack with a file that cannot be found should not be editable")
 	}
 }
 // A schedule made while an orchestrator was running says "run compose inside that
 // container". When the container is gone the schedule must still work, against the same
 // files on the host — otherwise retiring the orchestrator silently breaks every scheduled
 // update made before it went.
 func TestRepointSurvivesAMissingContainer(t *testing.T) {
 	root := t.TempDir()
 	dir := filepath.Join(root, "orch", "stacks", "app")
 	if err := os.MkdirAll(dir, 0o755); err != nil {
 		t.Fatal(err)
 	}
 	host := filepath.Join(dir, "compose.yaml")
 	if err := os.WriteFile(host, []byte("services: {}\n"), 0o644); err != nil {
 		t.Fatal(err)
 	}
 	db, err := store.Open(filepath.Join(root, "state.json"))
 	if err != nil {
 		t.Fatal(err)
 	}
 	if err := db.Update(func(st *store.State) {
 		st.Mounts = []store.RememberedMount{{Container: "gone-orchestrator", Dest: "/app/data", Src: filepath.Join(root, "orch")}}
 	}); err != nil {
 		t.Fatal(err)
 	}
 	in := compose.Spec{Project: "app", Files: []string{"/app/data/stacks/app/compose.yaml"}, Via: "gone-orchestrator", Tool: "docker"}
 	got := repoint(context.Background(), db, in)
 	if got.Via != "" {
 		t.Errorf("via = %q, want it dropped", got.Via)
 	}
 	if len(got.Files) != 1 || got.Files[0] != host {
 		t.Errorf("files = %v, want [%s]", got.Files, host)
 	}
 	// Nothing remembered: the schedule is left exactly as it was rather than pointed at a
 	// guess. It will fail loudly, which is better than deploying the wrong file.
 	empty, err := store.Open(filepath.Join(root, "empty.json"))
 	if err != nil {
 		t.Fatal(err)
 	}
 	if got := repoint(context.Background(), empty, in); got.Via != in.Via || len(got.Files) != 1 || got.Files[0] != in.Files[0] {
 		t.Errorf("unresolvable spec was changed: %+v", got)
 	}
 }
@@ -12,6 +12,7 @@ import (
 	"io"
 	"log"
 	"net/http"
 	"strconv"
 	"strings"
 	"time"
@@ -26,6 +27,10 @@ type actor struct {
 	Session string // session ID, "" for a token
 	token   string // the raw cookie token, so a handler can spare the current session
 	Token   string // token name, "" for a session
 	// Guest marks a temporary look-around login, and Expires is when it stops working. What
 	// a guest may reach is decided in the middleware, not here; see internal/api/guests.go.
 	Guest   bool
 	Expires time.Time
 }
 type ctxKey int
@@ -44,6 +49,8 @@ func actorOf(r *http.Request) actor {
 // who names the caller for a log line: "core@10.0.0.5" for a person, "token:hub:core@…" for a hub.
 func (a actor) String() string {
 	switch {
 	case a.User != "" && a.Guest:
 		return "guest:" + a.User
 	case a.User != "":
 		return a.User
 	case a.Token != "":
@@ -59,7 +66,7 @@ func (s *Server) listUsers(w http.ResponseWriter, r *http.Request) {
 	counts := s.sessions.Count()
 	var out []map[string]any
 	s.cfg.View(func(c *config.Config) {
-		for _, u := range c.Users {
+		for _, u := range c.Accounts() { // guest logins are listed on their own, see listGuests
 			var seen any
 			if !u.LastSeen.IsZero() {
 				seen = u.LastSeen
@@ -84,6 +91,10 @@ func (s *Server) addUser(w http.ResponseWriter, r *http.Request) {
 		jsonErr(w, 400, "a user name is 1–32 lower-case letters, digits, . _ or -, starting with a letter or digit")
 		return
 	}
 	if strings.HasPrefix(name, guestPrefix) {
 		jsonErr(w, 400, "names starting with "+guestPrefix+" belong to guest logins")
 		return
 	}
 	if len(body.Password) < auth.MinPasswordLen {
 		jsonErr(w, 400, "password must be at least 12 characters")
 		return
@@ -141,7 +152,7 @@ func (s *Server) removeUser(w http.ResponseWriter, r *http.Request) {
 	}
 	found, last := false, false
 	if err := s.cfg.Update(func(c *config.Config) {
-		if len(c.Users) <= 1 {
+		if len(c.Accounts()) <= 1 { // guest logins don't keep the last real account alive
 			last = true
 			return
 		}
@@ -254,6 +265,29 @@ func (s *Server) revokeSession(w http.ResponseWriter, r *http.Request) {
 	w.WriteHeader(204)
 }
 // revokeIdleSessions drops the caller's sessions that have not been used for ?hours (24 by
 // default): the scripts and headless browsers that signed in once and never came back,
 // without touching the phone that was used this morning.
 func (s *Server) revokeIdleSessions(w http.ResponseWriter, r *http.Request) {
 	a := actorOf(r)
 	if a.User == "" {
 		jsonErr(w, 403, "only a signed-in user may do that")
 		return
 	}
 	hours := 24
 	if h := r.URL.Query().Get("hours"); h != "" {
 		n, err := strconv.Atoi(h)
 		if err != nil || n < 1 || n > 24*365 {
 			jsonErr(w, 400, "hours is a whole number from 1 to 8760")
 			return
 		}
 		hours = n
 	}
 	n := s.sessions.RevokeIdle(a.User, a.token, time.Duration(hours)*time.Hour)
 	log.Printf("%d idle session(s) of %q revoked (idle > %dh, by %s)", n, a.User, hours, clientIP(r))
 	jsonOK(w, 200, map[string]int{"revoked": n})
 }
 func (s *Server) revokeOtherSessions(w http.ResponseWriter, r *http.Request) {
 	a := actorOf(r)
 	if a.User == "" {
@@ -0,0 +1,490 @@
 package api
 // The Virtualization tab: LXC containers and KVM virtual machines through Incus, per
 // machine. Three page states — ready, repairable, installable, or unavailable with the
 // reason — and every action lands on the machine that was asked. Long work (install, create,
 // image downloads, publish) runs as a followed task; consoles are terminal sessions like
 // any other, so they reattach and work through the hub's proxy.
 import (
 	"context"
 	"encoding/json"
 	"fmt"
 	"io"
 	"log"
 	"net/http"
 	"os"
 	"path/filepath"
 	"strings"
 	"time"
 	"hamlab/internal/incus"
 )
 func readBody(r *http.Request, v any, max int64) error {
 	return json.NewDecoder(io.LimitReader(r.Body, max)).Decode(v)
 }
 // virtOverview is the page: the machine's state, and when Incus answers, everything on it.
 func (s *Server) virtOverview(w http.ResponseWriter, r *http.Request) {
 	st := s.virt.Detect(r.Context())
 	out := map[string]any{"status": st}
 	if st.State == "ready" {
 		ctx, cancel := context.WithTimeout(r.Context(), 40*time.Second)
 		defer cancel()
 		inst, err := s.virt.Instances(ctx)
 		if err != nil {
 			out["error"] = err.Error()
 		}
 		out["instances"] = inst
 		imgs, _ := s.virt.Images(ctx)
 		out["images"] = imgs
 		out["pools"] = s.virt.Pools(ctx)
 		out["networks"] = s.virt.Networks(ctx)
 	}
 	out["library"] = map[string]any{"images": s.virt.Library(s.virt.ImagesDir()), "isos": s.virt.Library(s.virt.ISOsDir())}
 	jsonOK(w, 200, out)
 }
 func (s *Server) virtInstall(w http.ResponseWriter, r *http.Request) {
 	var o incus.InstallOpts
 	if err := readBody(r, &o, 16<<10); err != nil {
 		jsonErr(w, 400, "bad request")
 		return
 	}
 	st := s.virt.Detect(r.Context())
 	if st.State != "installable" && st.State != "repair" {
 		jsonErr(w, 409, "nothing to install here: "+st.State)
 		return
 	}
 	if err := s.virt.EnsureDirs(); err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	script, lay, err := s.virt.InstallScript(st, o)
 	if err != nil {
 		jsonErr(w, 400, err.Error())
 		return
 	}
 	if err := s.virt.SaveLayout(lay); err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	// kept beside the layout, so what was run is on record
 	_ = os.WriteFile(filepath.Join(s.virt.Dir(), "install.sh"), []byte(script), 0o700)
 	t, err := s.tasks.Start("virt:install", "install Incus", s.virt.ShellCmd(r.Context(), script))
 	if err != nil {
 		jsonErr(w, 409, err.Error())
 		return
 	}
 	log.Printf("virt: install started as task %s — %s pool, %s (from %s)", t.ID, lay.Driver, lay.Subnet, clientIP(r))
 	jsonOK(w, 202, t.Snapshot(false))
 }
 func (s *Server) virtReady(w http.ResponseWriter, r *http.Request) bool {
 	if st := s.virt.Detect(r.Context()); st.State != "ready" {
 		jsonErr(w, 409, "Incus isn't ready on this machine")
 		return false
 	}
 	return true
 }
 // ---- images ----
 func (s *Server) virtRemoteImages(w http.ResponseWriter, r *http.Request) {
 	if !s.virtReady(w, r) {
 		return
 	}
 	list, err := s.virt.Remote(r.Context(), r.URL.Query().Get("refresh") == "1")
 	if err != nil {
 		jsonErr(w, 502, err.Error())
 		return
 	}
 	if t := r.URL.Query().Get("type"); t == "container" || t == "virtual-machine" {
 		var f []incus.Image
 		for _, im := range list {
 			if im.Type == t {
 				f = append(f, im)
 			}
 		}
 		list = f
 	}
 	if list == nil {
 		list = []incus.Image{}
 	}
 	jsonOK(w, 200, map[string]any{"images": list})
 }
 func (s *Server) virtImageDownload(w http.ResponseWriter, r *http.Request) {
 	var body struct {
 		Alias string `json:"alias"`
 	}
 	_ = readBody(r, &body, 4096)
 	alias := strings.TrimSpace(body.Alias)
 	if !incus.ValidImage(alias) || strings.Contains(alias, ":") {
 		jsonErr(w, 400, "alias: like debian/13 or ubuntu/24.04/cloud")
 		return
 	}
 	if !s.virtReady(w, r) {
 		return
 	}
 	local := strings.ReplaceAll(alias, "/", "-")
 	// copied into the local store under a flat alias, then exported to the library so the
 	// file is Ham Lab's whether or not Incus keeps its cache
 	script := `set -e; incus image copy "images:$1" local: --alias "$2"; cd "$3" && incus image export "$2" "$2" && echo "exported to $3/$2"`
 	t, err := s.tasks.Start("virt:image", "download "+alias, s.virt.ShellCmd(r.Context(), script, alias, local, s.virt.ImagesDir()))
 	if err != nil {
 		jsonErr(w, 409, err.Error())
 		return
 	}
 	log.Printf("virt: image %s download as task %s (from %s)", alias, t.ID, clientIP(r))
 	jsonOK(w, 202, t.Snapshot(false))
 }
 func (s *Server) virtImageImport(w http.ResponseWriter, r *http.Request) {
 	var body struct {
 		Path  string `json:"path"`
 		Alias string `json:"alias"`
 		Type  string `json:"type"` // container | virtual-machine
 	}
 	_ = readBody(r, &body, 8192)
 	alias := strings.TrimSpace(body.Alias)
 	if !incus.ValidImage(alias) || strings.Contains(alias, ":") {
 		jsonErr(w, 400, "alias: letters, digits, dots, dashes")
 		return
 	}
 	p := filepath.Clean(body.Path)
 	st, err := os.Stat(p)
 	if !filepath.IsAbs(p) || err != nil || !st.Mode().IsRegular() {
 		jsonErr(w, 400, "path: a file on this machine")
 		return
 	}
 	if !s.virtReady(w, r) {
 		return
 	}
 	dest := filepath.Join(s.virt.ImagesDir(), filepath.Base(p))
 	var script string
 	if body.Type == "virtual-machine" {
 		// a bare disk image needs the metadata Incus wants beside it; a small tarball Ham Lab writes
 		script = `set -e; [ "$1" = "$2" ] || cp -n "$1" "$2"; META="$3/$4.metadata.tar.gz"; T=$(mktemp -d); printf 'architecture: %s\ncreation_date: %s\nproperties:\n  description: "%s"\n  os: custom\n' "$5" "$(date +%s)" "$4" > "$T/metadata.yaml"; tar -C "$T" -czf "$META" metadata.yaml; rm -rf "$T"; incus image import "$META" "$2" --alias "$4"; echo "imported $4 from $2"`
 	} else {
 		script = `set -e; [ "$1" = "$2" ] || cp -n "$1" "$2"; incus image import "$2" --alias "$4"; echo "imported $4 from $2"`
 	}
 	t, err := s.tasks.Start("virt:image", "import "+alias, s.virt.ShellCmd(r.Context(), script, p, dest, s.virt.ImagesDir(), alias, incus.IncusArch()))
 	if err != nil {
 		jsonErr(w, 409, err.Error())
 		return
 	}
 	log.Printf("virt: image import %s from %s as task %s (from %s)", alias, p, t.ID, clientIP(r))
 	jsonOK(w, 202, t.Snapshot(false))
 }
 func (s *Server) virtImagePublish(w http.ResponseWriter, r *http.Request) {
 	var body struct {
 		Instance string `json:"instance"`
 		Alias    string `json:"alias"`
 	}
 	_ = readBody(r, &body, 4096)
 	alias := strings.TrimSpace(body.Alias)
 	if !incus.ValidName(body.Instance) || !incus.ValidImage(alias) || strings.Contains(alias, ":") {
 		jsonErr(w, 400, "an instance and an alias")
 		return
 	}
 	if !s.virtReady(w, r) {
 		return
 	}
 	script := `set -e; incus publish "$1" --alias "$2"; cd "$3" && incus image export "$2" "$2" && echo "exported to $3/$2"`
 	t, err := s.tasks.Start("virt:image", "publish "+body.Instance+" as "+alias, s.virt.ShellCmd(r.Context(), script, body.Instance, alias, s.virt.ImagesDir()))
 	if err != nil {
 		jsonErr(w, 409, err.Error())
 		return
 	}
 	log.Printf("virt: publish %s as %s, task %s (from %s)", body.Instance, alias, t.ID, clientIP(r))
 	jsonOK(w, 202, t.Snapshot(false))
 }
 func (s *Server) virtImageDelete(w http.ResponseWriter, r *http.Request) {
 	var body struct {
 		Ref string `json:"ref"`
 	}
 	_ = readBody(r, &body, 4096)
 	if !incus.ValidImage(body.Ref) {
 		jsonErr(w, 400, "bad image")
 		return
 	}
 	if !s.virtReady(w, r) {
 		return
 	}
 	ctx, cancel := context.WithTimeout(r.Context(), 60*time.Second)
 	defer cancel()
 	if out, err := s.virt.Cmd(ctx, "image", "delete", body.Ref).CombinedOutput(); err != nil {
 		jsonErr(w, 409, strings.TrimPrefix(strings.TrimSpace(string(out)), "Error: "))
 		return
 	}
 	log.Printf("virt: image %s deleted (from %s)", body.Ref, clientIP(r))
 	w.WriteHeader(204)
 }
 // ---- instances ----
 func (s *Server) virtCreate(w http.ResponseWriter, r *http.Request) {
 	var sp incus.CreateSpec
 	if err := readBody(r, &sp, 512<<10); err != nil {
 		jsonErr(w, 400, "bad request")
 		return
 	}
 	st := s.virt.Detect(r.Context())
 	if st.State != "ready" {
 		jsonErr(w, 409, "Incus isn't ready on this machine")
 		return
 	}
 	args, err := incus.CreateArgs(sp, st.KVM)
 	if err != nil {
 		jsonErr(w, 400, err.Error())
 		return
 	}
 	if inst, err := s.virt.Instances(r.Context()); err == nil {
 		for _, in := range inst {
 			if in.Name == sp.Name {
 				jsonErr(w, 409, "there is already an instance called "+sp.Name)
 				return
 			}
 		}
 	}
 	kind := "container"
 	if sp.VM {
 		kind = "VM"
 	}
 	t, err := s.tasks.Start("virt:create", fmt.Sprintf("create %s %s", kind, sp.Name), s.virt.Cmd(r.Context(), args...))
 	if err != nil {
 		jsonErr(w, 409, err.Error())
 		return
 	}
 	log.Printf("virt: create %s %s from %s as task %s (from %s)", kind, sp.Name, sp.Image, t.ID, clientIP(r))
 	jsonOK(w, 202, t.Snapshot(false))
 }
 func (s *Server) virtInstance(w http.ResponseWriter, r *http.Request) {
 	name := r.PathValue("name")
 	if !incus.ValidName(name) {
 		jsonErr(w, 400, "bad name")
 		return
 	}
 	list, err := s.virt.Instances(r.Context())
 	if err != nil {
 		jsonErr(w, 502, err.Error())
 		return
 	}
 	for _, in := range list {
 		if in.Name == name {
 			jsonOK(w, 200, in)
 			return
 		}
 	}
 	jsonErr(w, 404, "no such instance")
 }
 // virtAction is start / stop / restart / kill / delete, and the two consoles.
 func (s *Server) virtAction(w http.ResponseWriter, r *http.Request) {
 	name, action := r.PathValue("name"), r.PathValue("action")
 	if !incus.ValidName(name) {
 		jsonErr(w, 400, "bad name")
 		return
 	}
 	if !s.virtReady(w, r) {
 		return
 	}
 	var args []string
 	switch action {
 	case "start":
 		args = []string{"start", name}
 	case "stop":
 		args = []string{"stop", name}
 	case "restart":
 		args = []string{"restart", name}
 	case "kill":
 		args = []string{"stop", "--force", name}
 	case "delete":
 		var body struct {
 			Confirm string `json:"confirm"`
 			Force   bool   `json:"force"`
 		}
 		_ = readBody(r, &body, 4096)
 		if body.Confirm != name {
 			jsonErr(w, 400, "type the instance's name to delete it")
 			return
 		}
 		args = []string{"delete", name}
 		if body.Force {
 			args = []string{"delete", "--force", name}
 		}
 	case "console", "exec":
 		s.virtTerminal(w, r, name, action)
 		return
 	default:
 		jsonErr(w, 400, "unknown action")
 		return
 	}
 	ctx, cancel := context.WithTimeout(r.Context(), 3*time.Minute)
 	defer cancel()
 	if out, err := s.virt.Cmd(ctx, args...).CombinedOutput(); err != nil {
 		jsonErr(w, 409, strings.TrimPrefix(strings.TrimSpace(string(out)), "Error: "))
 		return
 	}
 	log.Printf("virt: %s %s (from %s)", action, name, clientIP(r))
 	w.WriteHeader(204)
 }
 // virtTerminal opens the serial console or a shell as a terminal session.
 func (s *Server) virtTerminal(w http.ResponseWriter, r *http.Request, name, kind string) {
 	var argv []string
 	title := ""
 	if kind == "console" {
 		argv = s.virt.Argv(r.Context(), "console", name)
 		title = "console " + name
 	} else {
 		argv = s.virt.Argv(r.Context(), "exec", name, "--", "sh", "-c", "if command -v bash >/dev/null 2>&1; then exec bash -l; else exec sh -l; fi")
 		title = "sh " + name
 	}
 	sess, err := s.mgr.CreateCmd("", title, argv)
 	if err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	log.Printf("virt: %s of %s as terminal %s (from %s)", kind, name, sess.ID, clientIP(r))
 	jsonOK(w, 200, map[string]string{"id": sess.ID, "name": title})
 }
 func (s *Server) virtRename(w http.ResponseWriter, r *http.Request) {
 	name := r.PathValue("name")
 	var body struct {
 		Name string `json:"name"`
 	}
 	_ = readBody(r, &body, 4096)
 	if !incus.ValidName(name) || !incus.ValidName(body.Name) {
 		jsonErr(w, 400, "bad name")
 		return
 	}
 	if !s.virtReady(w, r) {
 		return
 	}
 	ctx, cancel := context.WithTimeout(r.Context(), time.Minute)
 	defer cancel()
 	if out, err := s.virt.Cmd(ctx, "rename", name, body.Name).CombinedOutput(); err != nil {
 		jsonErr(w, 409, strings.TrimPrefix(strings.TrimSpace(string(out)), "Error: "))
 		return
 	}
 	log.Printf("virt: renamed %s to %s (from %s)", name, body.Name, clientIP(r))
 	w.WriteHeader(204)
 }
 // virtConfig sets limits, the snapshot schedule, or a description — the config keys the page
 // edits, each validated for its shape. An empty value unsets the key.
 func (s *Server) virtConfig(w http.ResponseWriter, r *http.Request) {
 	name := r.PathValue("name")
 	var body map[string]string
 	if err := readBody(r, &body, 16<<10); err != nil || !incus.ValidName(name) {
 		jsonErr(w, 400, "bad request")
 		return
 	}
 	var set, unset []string
 	for k, v := range body {
 		v = strings.TrimSpace(v)
 		ok := false
 		switch k {
 		case "limits.cpu":
 			ok = incus.ValidCPU(v)
 		case "limits.memory":
 			ok = incus.ValidSize(v)
 		case "snapshots.schedule":
 			ok = incus.ValidCron(v)
 		case "snapshots.expiry":
 			ok = incus.ValidExpiry(v)
 		case "description":
 			ok = len(v) <= 200 && !strings.ContainsAny(v, "\n\r")
 		case "boot.autostart":
 			ok = v == "" || v == "true" || v == "false"
 		}
 		if !ok {
 			jsonErr(w, 400, "bad value for "+k)
 			return
 		}
 		if v == "" {
 			unset = append(unset, k)
 		} else {
 			set = append(set, k+"="+v)
 		}
 	}
 	if !s.virtReady(w, r) {
 		return
 	}
 	ctx, cancel := context.WithTimeout(r.Context(), time.Minute)
 	defer cancel()
 	if len(set) > 0 {
 		if out, err := s.virt.Cmd(ctx, append([]string{"config", "set", name}, set...)...).CombinedOutput(); err != nil {
 			jsonErr(w, 409, strings.TrimPrefix(strings.TrimSpace(string(out)), "Error: "))
 			return
 		}
 	}
 	for _, k := range unset {
 		_, _ = s.virt.Cmd(ctx, "config", "unset", name, k).CombinedOutput() // unsetting what isn't set is fine
 	}
 	log.Printf("virt: config %s set %v unset %v (from %s)", name, set, unset, clientIP(r))
 	w.WriteHeader(204)
 }
 func (s *Server) virtSnapshotCreate(w http.ResponseWriter, r *http.Request) {
 	name := r.PathValue("name")
 	var body struct {
 		Name string `json:"name"`
 	}
 	_ = readBody(r, &body, 4096)
 	snap := strings.TrimSpace(body.Name)
 	if snap == "" {
 		snap = time.Now().Format("2006-01-02_15.04.05")
 	}
 	if !incus.ValidName(name) || !incus.ValidSnap(snap) {
 		jsonErr(w, 400, "bad name")
 		return
 	}
 	if !s.virtReady(w, r) {
 		return
 	}
 	ctx, cancel := context.WithTimeout(r.Context(), 5*time.Minute)
 	defer cancel()
 	if out, err := s.virt.Cmd(ctx, "snapshot", "create", name, snap).CombinedOutput(); err != nil {
 		jsonErr(w, 409, strings.TrimPrefix(strings.TrimSpace(string(out)), "Error: "))
 		return
 	}
 	log.Printf("virt: snapshot %s/%s (from %s)", name, snap, clientIP(r))
 	jsonOK(w, 200, map[string]string{"name": snap})
 }
 func (s *Server) virtSnapshotAction(w http.ResponseWriter, r *http.Request) {
 	name, snap, action := r.PathValue("name"), r.PathValue("snap"), r.PathValue("action")
 	if !incus.ValidName(name) || !incus.ValidSnap(snap) {
 		jsonErr(w, 400, "bad name")
 		return
 	}
 	var args []string
 	switch action {
 	case "restore":
 		args = []string{"snapshot", "restore", name, snap}
 	case "delete":
 		args = []string{"snapshot", "delete", name, snap}
 	default:
 		jsonErr(w, 400, "unknown action")
 		return
 	}
 	if !s.virtReady(w, r) {
 		return
 	}
 	ctx, cancel := context.WithTimeout(r.Context(), 10*time.Minute)
 	defer cancel()
 	if out, err := s.virt.Cmd(ctx, args...).CombinedOutput(); err != nil {
 		jsonErr(w, 409, strings.TrimPrefix(strings.TrimSpace(string(out)), "Error: "))
 		return
 	}
 	log.Printf("virt: snapshot %s %s/%s (from %s)", action, name, snap, clientIP(r))
 	w.WriteHeader(204)
 }
@@ -0,0 +1,115 @@
 package api
 // The VPN tab: this machine's WireGuard tunnels, who is on them, and — where the tunnel is
 // one this machine configures — adding, renaming and removing the devices that dial in.
 //
 // Per machine like every other tab: the tunnels shown and edited are the ones on the machine
 // in the switcher, which is the only honest way to show them, since the inbound server and an
 // outbound tunnel are different machines here.
 import (
 	"context"
 	"encoding/json"
 	"errors"
 	"io"
 	"log"
 	"net/http"
 	"strings"
 	"hamlab/internal/vpn"
 )
 func (s *Server) vpnReport(w http.ResponseWriter, r *http.Request) {
 	jsonOK(w, 200, vpn.Look(r.Context()))
 }
 // tunnelOf finds the tunnel a request names, so every editing handler starts from the live
 // state rather than from whatever the page last saw.
 func (s *Server) tunnelOf(ctx context.Context, w http.ResponseWriter, name string) (vpn.Tunnel, bool) {
 	for _, t := range vpn.Look(ctx).Tunnels {
 		if t.Name == name {
 			if !t.Managed {
 				jsonErr(w, 400, vpn.ErrUnmanaged.Error())
 				return t, false
 			}
 			return t, true
 		}
 	}
 	jsonErr(w, 404, "no such tunnel on this machine")
 	return vpn.Tunnel{}, false
 }
 // addVPNPeer makes a device and hands back its profile. The profile holds a private key that
 // is generated for this call and stored nowhere, so it is the only time it can be read — the
 // same bargain as any password this console generates.
 func (s *Server) addVPNPeer(w http.ResponseWriter, r *http.Request) {
 	t, ok := s.tunnelOf(r.Context(), w, r.PathValue("iface"))
 	if !ok {
 		return
 	}
 	var body vpn.NewPeer
 	if err := json.NewDecoder(io.LimitReader(r.Body, 8<<10)).Decode(&body); err != nil {
 		jsonErr(w, 400, "bad request")
 		return
 	}
 	made, err := vpn.AddPeer(r.Context(), t, body)
 	if err != nil {
 		code := 400
 		if errors.Is(err, vpn.ErrUnmanaged) {
 			code = 409
 		}
 		jsonErr(w, code, err.Error())
 		return
 	}
 	// The name and address are worth a log line; the profile is emphatically not.
 	log.Printf("vpn: device %q added to %s as %s (by %s)", made.Name, t.Name, made.Address, clientIP(r))
 	jsonOK(w, 200, made)
 }
 func (s *Server) editVPNPeer(w http.ResponseWriter, r *http.Request) {
 	t, ok := s.tunnelOf(r.Context(), w, r.PathValue("iface"))
 	if !ok {
 		return
 	}
 	var body struct{ Name, Address string }
 	if err := json.NewDecoder(io.LimitReader(r.Body, 4<<10)).Decode(&body); err != nil {
 		jsonErr(w, 400, "bad request")
 		return
 	}
 	section := r.PathValue("section")
 	if err := vpn.EditPeer(r.Context(), t, section, body.Name, body.Address); err != nil {
 		jsonErr(w, 400, err.Error())
 		return
 	}
 	log.Printf("vpn: device %s on %s changed (by %s)", section, t.Name, clientIP(r))
 	w.WriteHeader(204)
 }
 func (s *Server) removeVPNPeer(w http.ResponseWriter, r *http.Request) {
 	t, ok := s.tunnelOf(r.Context(), w, r.PathValue("iface"))
 	if !ok {
 		return
 	}
 	section := r.PathValue("section")
 	var body struct{ Confirm string }
 	_ = json.NewDecoder(io.LimitReader(r.Body, 4<<10)).Decode(&body)
 	// The name has to be typed back. Removing a device is the one action here that cannot be
 	// undone — the key is gone and the profile on the device stops working — and it is one tap
 	// away from a list of devices that all look alike.
 	var want string
 	for _, p := range t.Peers {
 		if p.Section == section {
 			want = p.Name
 		}
 	}
 	if want != "" && !strings.EqualFold(strings.TrimSpace(body.Confirm), want) {
 		jsonErr(w, 400, "type the device's name to remove it")
 		return
 	}
 	if err := vpn.DeletePeer(r.Context(), t, section); err != nil {
 		jsonErr(w, 400, err.Error())
 		return
 	}
 	log.Printf("vpn: device %s (%s) removed from %s (by %s)", section, want, t.Name, clientIP(r))
 	w.WriteHeader(204)
 }
@@ -136,6 +136,17 @@ func (s *Server) zfsOverview(w http.ResponseWriter, r *http.Request) {
 	})
 }
 // zfsARC is the cache: the sampler's live ARC (size, the hit rate over the last interval)
 // beside the storage package's fuller picture (since-boot counters, tunables, and each
 // pool's cache/log/special devices or the lack of them).
 func (s *Server) zfsARC(w http.ResponseWriter, r *http.Request) {
 	out := map[string]any{"detail": s.storage.ARC(r.Context()), "live": nil}
 	if smp, ok := s.metrics.Latest(); ok && smp.ARC != nil {
 		out["live"] = smp.ARC
 	}
 	jsonOK(w, 200, out)
 }
 func (s *Server) zfsPolicyPut(w http.ResponseWriter, r *http.Request) {
 	var p storage.Policy
 	if err := json.NewDecoder(io.LimitReader(r.Body, 16<<10)).Decode(&p); err != nil {
@@ -217,24 +228,6 @@ func (s *Server) zfsPolicy(id string) (storage.Policy, bool) {
 	return storage.Policy{}, false
 }
 // zfsPreview is what this schedule would destroy if it ran now. Nothing is deleted.
 func (s *Server) zfsPreview(w http.ResponseWriter, r *http.Request) {
 	p, ok := s.zfsPolicy(r.PathValue("id"))
 	if !ok {
 		jsonErr(w, 404, "no such schedule")
 		return
 	}
 	doomed, err := s.storage.Doomed(r.Context(), &p)
 	if err != nil {
 		jsonErr(w, 500, err.Error())
 		return
 	}
 	if doomed == nil {
 		doomed = []storage.Snapshot{}
 	}
 	jsonOK(w, 200, map[string]any{"expired": doomed})
 }
 func (s *Server) zfsRunNow(w http.ResponseWriter, r *http.Request) {
 	id := r.PathValue("id")
 	p, ok := s.zfsPolicy(id)
@@ -28,6 +28,10 @@ const (
 	argonKeyLen  = 32
 	SessionTTL = 30 * 24 * time.Hour
 	// SessionIdle is how long a session may go unused before it is dropped, whatever its
 	// expiry says. Every script, curl and headless browser that signs in leaves a session;
 	// a phone that is used keeps its own alive by being used.
 	SessionIdle = 7 * 24 * time.Hour
 )
 func HashPassword(pw string) (string, error) {
@@ -113,11 +117,14 @@ var dummyHash = func() string {
 	return h
 }()
-// CheckUser verifies a password for one of users. A missing user fails in the same time
+// CheckUser verifies a password for one of users. A missing user — or a guest login whose
-// as a wrong password.
+// time is up — fails in the same time as a wrong password.
 func CheckUser(users []config.User, name, pw string) bool {
 	for _, u := range users {
 		if u.Name == name {
 			if u.Expired(time.Now()) {
 				break
 			}
 			return VerifyPassword(u.Hash, pw)
 		}
 	}
@@ -125,6 +132,51 @@ func CheckUser(users []config.User, name, pw string) bool {
 	return false
 }
 // ---- guest logins ----
 // guestAlphabet leaves out every letter and digit that can be misread in a screenshot or
 // over the phone: i, l, o, 0 and 1. A guest login is read aloud and typed by hand.
 const guestAlphabet = "abcdefghjkmnpqrstuvwxyz23456789"
 // randomFrom returns n characters of guestAlphabet, chosen without modulo bias.
 func randomFrom(n int) (string, error) {
 	out := make([]byte, 0, n)
 	buf := make([]byte, n)
 	limit := byte(256 - 256%len(guestAlphabet)) // the top of the last whole run of the alphabet
 	for len(out) < n {
 		if _, err := rand.Read(buf); err != nil {
 			return "", err
 		}
 		for _, b := range buf {
 			if b >= limit {
 				continue // would favour the first few letters; ask for another byte instead
 			}
 			out = append(out, guestAlphabet[int(b)%len(guestAlphabet)])
 			if len(out) == n {
 				break
 			}
 		}
 	}
 	return string(out), nil
 }
 // GuestPassword is a password for a guest login: four groups of four, ~78 bits, in an
 // alphabet with nothing ambiguous in it.
 func GuestPassword() (string, error) {
 	var groups []string
 	for i := 0; i < 4; i++ {
 		g, err := randomFrom(4)
 		if err != nil {
 			return "", err
 		}
 		groups = append(groups, g)
 	}
 	return strings.Join(groups, "-"), nil
 }
 // GuestSuffix is the tail of a generated guest name.
 func GuestSuffix() (string, error) { return randomFrom(4) }
 // ---- sessions ----
 // Sessions is a token store backed by the state file. Expiry is sliding.
@@ -141,12 +193,26 @@ func NewSessions(st *store.Store) *Sessions {
 func pruneExpired(state *store.State) {
 	now := time.Now()
 	for tok, ses := range state.Sessions {
-		if ses.Expires.Before(now) {
+		if !live(ses, now) {
 			delete(state.Sessions, tok)
 		}
 	}
 }
 // live is whether a session still counts: unexpired, and used within SessionIdle. A
 // session from before visits were recorded is judged by its creation, or failing that by
 // its expiry alone.
 func live(ses store.Session, now time.Time) bool {
 	if !ses.Expires.After(now) {
 		return false
 	}
 	seen := ses.LastSeen
 	if seen.IsZero() {
 		seen = ses.Created
 	}
 	return seen.IsZero() || now.Sub(seen) <= SessionIdle
 }
 func randomHex(n int) (string, error) {
 	b := make([]byte, n)
 	if _, err := rand.Read(b); err != nil {
@@ -186,7 +252,7 @@ func (s *Sessions) Lookup(tok, ip string) (store.Session, bool) {
 	ok, touch := false, false
 	s.st.View(func(state *store.State) {
 		found, have := state.Sessions[tok]
-		if have && found.Expires.After(time.Now()) {
+		if have && live(found, time.Now()) {
 			ses, ok = found, true
 			touch = time.Until(found.Expires) < SessionTTL-time.Hour || time.Since(found.LastSeen) > time.Hour
 		}
@@ -216,7 +282,7 @@ func (s *Sessions) List(user string) []store.Session {
 	s.st.View(func(state *store.State) {
 		now := time.Now()
 		for _, ses := range state.Sessions {
-			if ses.User == user && ses.Expires.After(now) {
+			if ses.User == user && live(ses, now) {
 				out = append(out, ses)
 			}
 		}
@@ -231,7 +297,7 @@ func (s *Sessions) Count() map[string]int {
 	s.st.View(func(state *store.State) {
 		now := time.Now()
 		for _, ses := range state.Sessions {
-			if ses.Expires.After(now) {
+			if live(ses, now) {
 				n[ses.User]++
 			}
 		}
@@ -267,14 +333,33 @@ func (s *Sessions) RevokeOthers(user, keep string) int {
 	return n
 }
 // RevokeIdle ends every session of user not seen for longer than idle, except the one
 // holding keep. Returns how many.
 func (s *Sessions) RevokeIdle(user, keep string, idle time.Duration) int {
 	n := 0
 	now := time.Now()
 	_ = s.st.Update(func(state *store.State) {
 		for tok, ses := range state.Sessions {
 			seen := ses.LastSeen
 			if seen.IsZero() {
 				seen = ses.Created
 			}
 			if ses.User == user && tok != keep && now.Sub(seen) > idle {
 				delete(state.Sessions, tok)
 				n++
 			}
 		}
 	})
 	return n
 }
 // RevokeUser ends every session of user (the account is going away).
 func (s *Sessions) RevokeUser(user string) { s.RevokeOthers(user, "") }
 // Limiter is a single global backoff: after 5 failures, each further failure
 // doubles a lockout starting at 30s, capped at 10 minutes. Single-user box.
 // Limiter locks out a client after repeated failed logins, per client address, with a
 // global backstop so a botnet can't burn CPU on argon2 indefinitely. A lockout for one
-// address never blocks another.
+// address never blocks another. Past perClientLock failures, each further failure doubles a
 // lockout starting at 30s, capped at 10 minutes.
 type Limiter struct {
 	mu      sync.Mutex
 	clients map[string]*bucket
@@ -290,7 +375,11 @@ type bucket struct {
 var ErrLocked = errors.New("too many failed logins")
 const (
-	perClientLock = 5   // failures before a client is locked
+	// perClientLock is how many failures one address gets before it is made to wait. Set high
 	// on purpose: a guest login is a generated password read off a message and typed by hand
 	// on a phone, and being locked out for fumbling it is worse than the brute-force risk —
 	// which the doubling backoff past this point, and the global backstop, still cover.
 	perClientLock = 20
 	globalLock    = 100 // failures (all clients) in the window before everyone waits a minute
 	globalWindow  = 5 * time.Minute
 )
@@ -0,0 +1,50 @@
 // Package compose runs docker compose the way a stack was deployed — which is not always
 // simply "on this host". A stack deployed by an orchestrator running inside its own
 // container records paths in that container's filesystem, so compose has to run in there;
 // a stack deployed to another machine over tcp needs DOCKER_HOST pointed at it. Both the
 // console and the scheduled updates it writes must invoke it identically, or a redeploy
 // from one would not match the other's labels.
 package compose
 import (
 	"context"
 	"os"
 	"os/exec"
 	"path/filepath"
 )
 // Spec is everything needed to run compose for one stack.
 type Spec struct {
 	Project string   // -p
 	Files   []string // -f, in order
 	Via     string   // a container to run compose inside; "" is this host
 	Remote  string   // a DOCKER_HOST; "" is this machine's Docker
 	Tool    string   // "docker" or "podman"
 }
 // Cmd builds the command. args are compose's own, e.g. "pull" or "up", "-d".
 func Cmd(ctx context.Context, s Spec, args ...string) *exec.Cmd {
 	tool := s.Tool
 	if tool == "" {
 		tool = "docker"
 	}
 	base := []string{"compose", "-p", s.Project}
 	for _, f := range s.Files {
 		base = append(base, "-f", f)
 	}
 	if s.Via != "" {
 		pre := []string{"exec"}
 		if s.Remote != "" {
 			pre = append(pre, "-e", "DOCKER_HOST="+s.Remote)
 		}
 		return exec.CommandContext(ctx, tool, append(append(append(pre, s.Via, "docker"), base...), args...)...)
 	}
 	cmd := exec.CommandContext(ctx, tool, append(base, args...)...)
 	if len(s.Files) > 0 {
 		cmd.Dir = filepath.Dir(s.Files[0])
 	}
 	if s.Remote != "" {
 		cmd.Env = append(os.Environ(), "DOCKER_HOST="+s.Remote)
 	}
 	return cmd
 }
@@ -46,6 +46,22 @@ type User struct {
 	Hash     string    `json:"hash"`
 	Created  time.Time `json:"created"`
 	LastSeen time.Time `json:"last_seen,omitempty"`
 	// Guest marks a temporary login made to show somebody the console: it may only read,
 	// it expires by itself at Expires, and it is not one of the machine's accounts — it is
 	// left out of the account list, does not count towards "the last account", and cannot
 	// be the one an empty user name means. See internal/api/guests.go.
 	Guest bool `json:"guest,omitempty"`
 	// Expires is when a guest login stops working. A guest without one has already
 	// expired, which is the safe way to read a half-written file.
 	Expires time.Time `json:"expires,omitempty"`
 	// IssuedBy is the account that made this guest login.
 	IssuedBy string `json:"issued_by,omitempty"`
 }
 // Expired reports whether a guest login's time is up. A real account never expires; a
 // guest with no expiry at all counts as expired rather than as forever.
 func (u User) Expired(now time.Time) bool {
 	return u.Guest && (u.Expires.IsZero() || now.After(u.Expires))
 }
 // UpdateSource is where this machine looks for a newer Ham Lab, and how often. The field
@@ -61,15 +77,33 @@ type UpdateSource struct {
 	Check string `json:"check,omitempty"`
 }
 // KVM is an IP-KVM: a device with a web page that gives screen, keyboard and mouse access
 // to a machine, including when that machine has no OS running. Ham Lab holds the list on
 // the hub and proxies each one, because a KVM matters exactly when the machine it is
 // attached to cannot answer for itself.
 type KVM struct {
 	ID   string `json:"id"`
 	Name string `json:"name"`
 	// URL is the device's own address, as you would type it into a browser on the LAN.
 	URL string `json:"url"`
 	// Machine is the host this KVM is attached to, and is the key: one KVM per machine, shown
 	// on that machine's KVM tab. Empty means the hub itself. The list lives on the hub even
 	// for a node's KVM, because a KVM earns its keep when the machine is down and cannot
 	// answer for itself.
 	Machine string    `json:"machine"`
 	Added   time.Time `json:"added"`
 }
 type Config struct {
 	Listen string `json:"listen"`
 	// PasswordHash is the single login of installs before users existed. On load it becomes
 	// the first user (see MigrateUsers) and is then dropped from the file.
-	PasswordHash string `json:"password_hash,omitempty"`
+	PasswordHash string  `json:"password_hash,omitempty"`
-	Users        []User `json:"users,omitempty"`
+	Users        []User  `json:"users,omitempty"`
 	TLS          string  `json:"tls,omitempty"` // "auto" → self-signed cert in the data dir; "" → plain HTTP
 	Tokens       []Token `json:"tokens,omitempty"`
 	Nodes        []Node  `json:"nodes,omitempty"`
 	KVMs         []KVM   `json:"kvms,omitempty"`
 	// Updates: where a newer build comes from. See UpdateSource.
 	Updates UpdateSource `json:"update,omitempty"`
@@ -127,6 +161,18 @@ func (c *Config) MigrateUsers(name string) bool {
 	return true
 }
 // Accounts is every real account: the machine's users, without the guest logins.
 // Call under View/Update or on a config nobody else holds.
 func (c *Config) Accounts() []User {
 	out := make([]User, 0, len(c.Users))
 	for _, u := range c.Users {
 		if !u.Guest {
 			out = append(out, u)
 		}
 	}
 	return out
 }
 // User finds an account by name. Call under View/Update or on a config nobody else holds.
 func (c *Config) User(name string) (User, bool) {
 	for _, u := range c.Users {
@@ -0,0 +1,843 @@
 // Package hardware is what is physically in the machine: the board it is, its processor,
 // the memory modules in its slots, the drives attached to it and the graphics in it.
 //
 // None of this changes while the machine is running, so it is read once and remembered.
 // Everything here is read from what Debian and its kin already have — dmidecode for the
 // parts only the firmware knows, lscpu, lspci, sysfs — and every reader is allowed to
 // come back empty: a machine with no dmidecode, no PCI bus or no root still gets a page,
 // with a note saying what could not be read and why.
 package hardware
 import (
 	"bufio"
 	"context"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"regexp"
 	"sort"
 	"strconv"
 	"strings"
 	"sync"
 	"time"
 	"hamlab/internal/priv"
 	"hamlab/internal/storage"
 )
 type System struct {
 	Model  string `json:"model,omitempty"` // "GMKtec NucBox G3"
 	Serial string `json:"serial,omitempty"`
 	Family string `json:"family,omitempty"`
 	Kernel string `json:"kernel,omitempty"`
 	Arch   string `json:"arch,omitempty"`
 	Virt   string `json:"virt,omitempty"` // "kvm", "lxc", … when this is not real hardware
 }
 // Slot is an expansion slot as the firmware describes it. Worth treating with suspicion on
 // a mini PC, where the table is often copied from a reference design and lists slots the
 // case has no room for.
 type Slot struct {
 	Designation string `json:"designation"`
 	Type        string `json:"type,omitempty"`
 	Usage       string `json:"usage,omitempty"` // Available | In Use | Unknown
 	Length      string `json:"length,omitempty"`
 	Width       string `json:"width,omitempty"`
 	Bus         string `json:"bus,omitempty"`    // PCI address of what is in it
 	Device      string `json:"device,omitempty"` // and what that turns out to be
 }
 // Board is the motherboard and the firmware on it.
 type Board struct {
 	Maker        string `json:"maker,omitempty"`
 	Product      string `json:"product,omitempty"`
 	Version      string `json:"version,omitempty"`
 	Serial       string `json:"serial,omitempty"`
 	AssetTag     string `json:"asset_tag,omitempty"`
 	Type         string `json:"type,omitempty"`     // "Motherboard", "Processor Module", …
 	Location     string `json:"location,omitempty"` // where it sits in the chassis
 	Chassis      string `json:"chassis,omitempty"`
 	ChassisMaker string `json:"chassis_maker,omitempty"`
 	BIOS         string `json:"bios,omitempty"` // vendor and version
 	BIOSDate     string `json:"bios_date,omitempty"`
 	BIOSRevision string `json:"bios_revision,omitempty"`
 	ROMSize      string `json:"rom_size,omitempty"`
 	UEFI         bool   `json:"uefi,omitempty"`
 	Slots        []Slot `json:"slots,omitempty"`
 }
 type Cache struct {
 	Level string `json:"level"` // L1d, L1i, L2, L3
 	Size  string `json:"size"`  // as lscpu puts it: "6 MiB (1 instance)"
 }
 type CPU struct {
 	Model          string   `json:"model,omitempty"`
 	Vendor         string   `json:"vendor,omitempty"`
 	Arch           string   `json:"arch,omitempty"`
 	Sockets        int      `json:"sockets,omitempty"`
 	SocketNames    []string `json:"socket_names,omitempty"` // what the firmware calls each socket
 	Cores          int      `json:"cores,omitempty"`
 	Threads        int      `json:"threads,omitempty"`
 	ThreadsPerCore int      `json:"threads_per_core,omitempty"`
 	MinMHz         float64  `json:"min_mhz,omitempty"`
 	MaxMHz         float64  `json:"max_mhz,omitempty"`
 	Family         string   `json:"family,omitempty"`
 	ModelID        string   `json:"model_id,omitempty"`
 	Stepping       string   `json:"stepping,omitempty"`
 	Microcode      string   `json:"microcode,omitempty"`
 	Virtualization string   `json:"virtualization,omitempty"`
 	Caches         []Cache  `json:"caches,omitempty"`
 	Flags          []string `json:"flags,omitempty"` // the handful anyone looks for
 }
 type Module struct {
 	Locator    string `json:"locator"` // the slot, as the board labels it
 	Bank       string `json:"bank,omitempty"`
 	Empty      bool   `json:"empty,omitempty"`
 	Size       uint64 `json:"size,omitempty"`
 	Type       string `json:"type,omitempty"`        // DDR4
 	FormFactor string `json:"form_factor,omitempty"` // SODIMM
 	SpeedMTs   int    `json:"speed_mts,omitempty"`   // what the module can do
 	RunningMTs int    `json:"running_mts,omitempty"` // what the board is running it at
 	Maker      string `json:"maker,omitempty"`
 	Part       string `json:"part,omitempty"`
 	Serial     string `json:"serial,omitempty"`
 	Rank       int    `json:"rank,omitempty"`
 	Voltage    string `json:"voltage,omitempty"`
 }
 type Memory struct {
 	Total       uint64   `json:"total"`                  // what the kernel can use
 	Installed   uint64   `json:"installed,omitempty"`    // what the modules add up to
 	MaxCapacity uint64   `json:"max_capacity,omitempty"` // what the board will take
 	Slots       int      `json:"slots,omitempty"`
 	Filled      int      `json:"filled,omitempty"`
 	ECC         string   `json:"ecc,omitempty"`
 	Modules     []Module `json:"modules,omitempty"`
 }
 type Drive struct {
 	Name       string `json:"name"`
 	Kind       string `json:"kind"` // nvme | ssd | hdd | usb | emmc | sd | other
 	Model      string `json:"model,omitempty"`
 	Vendor     string `json:"vendor,omitempty"`
 	Serial     string `json:"serial,omitempty"`
 	Firmware   string `json:"firmware,omitempty"`
 	Tran       string `json:"tran,omitempty"`
 	Size       uint64 `json:"size"`
 	Rotational bool   `json:"rotational,omitempty"`
 	Removable  bool   `json:"removable,omitempty"`
 }
 type GPU struct {
 	Model  string `json:"model,omitempty"`
 	Vendor string `json:"vendor,omitempty"`
 	Slot   string `json:"slot,omitempty"` // PCI address, when it is on a bus
 	Driver string `json:"driver,omitempty"`
 	Node   string `json:"node,omitempty"` // card0
 	Memory uint64 `json:"memory,omitempty"`
 	Class  string `json:"class,omitempty"` // VGA | 3D | Display
 }
 type Report struct {
 	At     time.Time `json:"at"`
 	System System    `json:"system"`
 	Board  Board     `json:"board"`
 	CPU    CPU       `json:"cpu"`
 	Memory Memory    `json:"memory"`
 	Drives []Drive   `json:"drives"`
 	GPUs   []GPU     `json:"gpus"`
 	Notes  []string  `json:"notes,omitempty"` // what could not be read, and why
 }
 // Manager remembers the last report. Hardware does not change under a running kernel —
 // the one thing that would, a drive plugged in, is what Refresh is for.
 type Manager struct {
 	mu   sync.Mutex
 	last *Report
 	at   time.Time
 }
 func New() *Manager { return &Manager{} }
 const keepFor = 10 * time.Minute
 func (m *Manager) Report(ctx context.Context, refresh bool) *Report {
 	m.mu.Lock()
 	if !refresh && m.last != nil && time.Since(m.at) < keepFor {
 		r := m.last
 		m.mu.Unlock()
 		return r
 	}
 	m.mu.Unlock()
 	r := collect(ctx)
 	m.mu.Lock()
 	m.last, m.at = r, time.Now()
 	m.mu.Unlock()
 	return r
 }
 func collect(ctx context.Context) *Report {
 	ctx, cancel := context.WithTimeout(ctx, 25*time.Second)
 	defer cancel()
 	r := &Report{At: time.Now()}
 	dmi, note := readDMI(ctx)
 	if note != "" {
 		r.Notes = append(r.Notes, note)
 	}
 	r.System = system(ctx, dmi)
 	r.Board = board(dmi, pciNames(ctx))
 	r.CPU = cpu(ctx, dmi)
 	r.Memory = memory(dmi)
 	if len(r.Memory.Modules) == 0 && note == "" {
 		r.Notes = append(r.Notes, "memory modules: the firmware's table says nothing about this machine's slots")
 	}
 	r.Drives = drives(ctx)
 	r.GPUs = gpus(ctx)
 	if len(r.GPUs) == 0 {
 		r.Notes = append(r.Notes, "graphics: nothing on the PCI bus and no kernel display device")
 	}
 	return r
 }
 // ---- dmidecode ----
 // record is one entry of the firmware's table: its DMI type and its key/value lines.
 type record struct {
 	Type int
 	V    map[string]string
 }
 // listKey holds every doubly-indented line of a record — dmidecode's lists, which have no
 // key of their own: BIOS characteristics, board features, slot characteristics.
 const listKey = "\x00list"
 var handleRe = regexp.MustCompile(`^Handle 0x[0-9A-Fa-f]+, DMI type (\d+)`)
 // readDMI runs dmidecode once for every type the page wants and parses the lot. It needs
 // root; without it the page loses the board, the memory modules and the socket names, and
 // says so rather than pretending the machine has none.
 func readDMI(ctx context.Context) ([]record, string) {
 	tool, err := exec.LookPath("dmidecode")
 	if err != nil {
 		if _, e := os.Stat("/usr/sbin/dmidecode"); e != nil {
 			return nil, "the board and the memory modules need dmidecode, which isn't installed here"
 		}
 		tool = "/usr/sbin/dmidecode"
 	}
 	args := []string{tool, "-t", "bios", "-t", "system", "-t", "baseboard", "-t", "chassis", "-t", "slot", "-t", "processor", "-t", "memory"}
 	if os.Geteuid() != 0 {
 		if !priv.Available() {
 			return nil, "the board and the memory modules come from the firmware's table, which only root may read"
 		}
 		args = priv.Wrap(args...)
 	}
 	out, err := exec.CommandContext(ctx, args[0], args[1:]...).Output()
 	if err != nil {
 		return nil, "the firmware's table could not be read: " + strings.TrimSpace(err.Error())
 	}
 	return parse(string(out)), ""
 }
 // parse turns dmidecode's output into records: a handle line starts one, and the single
 // indented "Key: Value" lines under it are its fields. Doubly-indented lines are the
 // continuation of a list (a processor's flags), which nothing here reads.
 func parse(text string) []record {
 	var recs []record
 	var cur *record
 	sc := bufio.NewScanner(strings.NewReader(text))
 	sc.Buffer(make([]byte, 0, 64<<10), 2<<20)
 	for sc.Scan() {
 		line := sc.Text()
 		if m := handleRe.FindStringSubmatch(line); m != nil {
 			t, _ := strconv.Atoi(m[1])
 			recs = append(recs, record{Type: t, V: map[string]string{}})
 			cur = &recs[len(recs)-1]
 			continue
 		}
 		if cur == nil || !strings.HasPrefix(line, "\t") {
 			if strings.TrimSpace(line) == "" {
 				cur = nil
 			}
 			continue
 		}
 		if strings.HasPrefix(line, "\t\t") { // a list under the field above it, e.g. BIOS characteristics
 			cur.V[listKey] += strings.TrimSpace(line) + "\n"
 			continue
 		}
 		k, v, ok := strings.Cut(strings.TrimSpace(line), ":")
 		if !ok {
 			continue
 		}
 		cur.V[k] = strings.TrimSpace(v)
 	}
 	return recs
 }
 func (r record) get(keys ...string) string {
 	for _, k := range keys {
 		if v := r.V[k]; v != "" && !unset(v) {
 			return v
 		}
 	}
 	return ""
 }
 // unset is every way the firmware has of saying "nobody filled this in".
 func unset(v string) bool {
 	switch strings.ToLower(v) {
 	case "", "unknown", "not specified", "not provided", "none", "default string", "to be filled by o.e.m.",
 		"system manufacturer", "system product name", "to be filled by oem", "n/a", "empty", "no module installed",
 		"<out of spec>", "other", "not applicable":
 		return true
 	}
 	return false
 }
 func first(recs []record, t int) record {
 	for _, r := range recs {
 		if r.Type == t {
 			return r
 		}
 	}
 	return record{Type: t, V: map[string]string{}}
 }
 func all(recs []record, t int) []record {
 	var out []record
 	for _, r := range recs {
 		if r.Type == t {
 			out = append(out, r)
 		}
 	}
 	return out
 }
 // ---- system ----
 func system(ctx context.Context, dmi []record) System {
 	sys := first(dmi, 1)
 	s := System{
 		Model:  join(sys.get("Manufacturer"), sys.get("Product Name")),
 		Serial: sys.get("Serial Number"),
 		Family: sys.get("Family"),
 		Arch:   runtimeArch(),
 	}
 	s.Kernel = readFile("/proc/sys/kernel/osrelease")
 	if s.Model == "" { // no SMBIOS: a Pi, a Mac, anything with a device tree
 		s.Model = strings.TrimRight(readFile("/sys/firmware/devicetree/base/model"), "\x00")
 	}
 	if out, err := exec.CommandContext(ctx, "systemd-detect-virt").Output(); err == nil {
 		if v := strings.TrimSpace(string(out)); v != "" && v != "none" {
 			s.Virt = v
 		}
 	}
 	return s
 }
 // board reads the motherboard, the chassis it sits in, the firmware on it and the
 // expansion slots the firmware claims. All of it is SMBIOS, so a machine without one — a
 // Mac, a Pi — gets an empty board, and the page says so rather than inventing one.
 func board(dmi []record, pci map[string]string) Board {
 	b, ch, bios := first(dmi, 2), first(dmi, 3), first(dmi, 0)
 	out := Board{
 		Maker:        b.get("Manufacturer"),
 		Product:      b.get("Product Name"),
 		Version:      b.get("Version"),
 		Serial:       b.get("Serial Number"),
 		AssetTag:     b.get("Asset Tag"),
 		Type:         b.get("Type"),
 		Location:     b.get("Location In Chassis"),
 		Chassis:      ch.get("Type"),
 		ChassisMaker: ch.get("Manufacturer"),
 		BIOS:         join(bios.get("Vendor"), bios.get("Version")),
 		BIOSDate:     bios.get("Release Date"),
 		BIOSRevision: bios.get("BIOS Revision"),
 	}
 	// "ROM Size: 0 MB" is a board that never filled the field in, not a board with no ROM
 	if rom := bios.get("ROM Size"); rom != "" && !strings.HasPrefix(rom, "0 ") {
 		out.ROMSize = rom
 	}
 	if out.Product == out.Maker {
 		out.Product = "" // a board named after its maker says nothing twice
 	}
 	out.UEFI = strings.Contains(bios.V[listKey], "UEFI") // a characteristics line, not a field
 	for _, sl := range all(dmi, 9) {
 		s := Slot{
 			Designation: sl.get("Designation"),
 			Type:        sl.get("Type"),
 			Usage:       sl.get("Current Usage"),
 			Length:      sl.get("Length"),
 			Width:       sl.get("Data Bus Width"),
 		}
 		// every slot on some boards claims 0000:00:00.0, which is the host bridge and
 		// certainly not what is plugged into that slot
 		if bus := sl.get("Bus Address"); bus != "" && bus != "0000:00:00.0" {
 			s.Bus, s.Device = bus, pci[bus]
 		}
 		if s.Designation != "" {
 			out.Slots = append(out.Slots, s)
 		}
 	}
 	return out
 }
 func join(parts ...string) string {
 	var keep []string
 	for _, p := range parts {
 		if p = strings.TrimSpace(p); p != "" && !unset(p) {
 			keep = append(keep, p)
 		}
 	}
 	// "GMKtec GMKtec" reads badly: drop a maker the product already names
 	if len(keep) == 2 && strings.HasPrefix(strings.ToLower(keep[1]), strings.ToLower(keep[0])) {
 		return keep[1]
 	}
 	return strings.Join(keep, " ")
 }
 func readFile(p string) string {
 	b, err := os.ReadFile(p)
 	if err != nil {
 		return ""
 	}
 	return strings.TrimSpace(string(b))
 }
 func runtimeArch() string {
 	if out, err := exec.Command("uname", "-m").Output(); err == nil {
 		return strings.TrimSpace(string(out))
 	}
 	return ""
 }
 // ---- processor ----
 // worthShowing is the short list of processor features anyone actually asks about; the
 // full flag list runs to hundreds and belongs in a terminal.
 var worthShowing = []string{"aes", "avx", "avx2", "avx512f", "sha_ni", "vmx", "svm", "sse4_2", "rdrand", "hypervisor"}
 func cpu(ctx context.Context, dmi []record) CPU {
 	c := CPU{}
 	f := lscpu(ctx)
 	c.Model = f["Model name"]
 	c.Vendor = f["Vendor ID"]
 	c.Arch = f["Architecture"]
 	c.Sockets = atoi(f["Socket(s)"])
 	c.ThreadsPerCore = atoi(f["Thread(s) per core"])
 	c.Threads = atoi(f["CPU(s)"])
 	c.Cores = atoi(f["Core(s) per socket"]) * maxInt(1, c.Sockets)
 	c.MinMHz = atof(f["CPU min MHz"])
 	c.MaxMHz = atof(f["CPU max MHz"])
 	c.Family, c.ModelID, c.Stepping = f["CPU family"], f["Model"], f["Stepping"]
 	c.Virtualization = f["Virtualization"]
 	for _, l := range []string{"L1d", "L1i", "L2", "L3"} {
 		if v := f[l+" cache"]; v != "" {
 			c.Caches = append(c.Caches, Cache{Level: l, Size: v})
 		}
 	}
 	have := map[string]bool{}
 	for _, fl := range strings.Fields(f["Flags"]) {
 		have[fl] = true
 	}
 	for _, fl := range worthShowing {
 		if have[fl] {
 			c.Flags = append(c.Flags, fl)
 		}
 	}
 	for _, p := range all(dmi, 4) {
 		if s := p.get("Socket Designation"); s != "" {
 			c.SocketNames = append(c.SocketNames, s)
 		}
 		if c.Model == "" {
 			c.Model = p.get("Version", "Family")
 		}
 		if c.MaxMHz == 0 {
 			if mhz := atof(strings.TrimSuffix(p.get("Max Speed"), " MHz")); mhz > 0 {
 				c.MaxMHz = mhz
 			}
 		}
 	}
 	if c.Sockets == 0 {
 		c.Sockets = len(c.SocketNames)
 	}
 	// A processor whose cores are not all alike — Apple's, Intel's P and E cores — is
 	// reported by lscpu as several groups, and "core(s) per socket" then counts one group
 	// rather than the chip. Without SMT every thread is a core, so the thread count is the
 	// honest answer.
 	if c.ThreadsPerCore == 1 && c.Threads > c.Cores {
 		c.Cores = c.Threads
 	}
 	// /proc/cpuinfo is the only place the running microcode revision is written down
 	if b, err := os.ReadFile("/proc/cpuinfo"); err == nil {
 		for _, line := range strings.Split(string(b), "\n") {
 			if k, v, ok := strings.Cut(line, ":"); ok && strings.TrimSpace(k) == "microcode" {
 				c.Microcode = strings.TrimSpace(v)
 				break
 			}
 		}
 	}
 	if c.Model == "" {
 		c.Model = "unknown processor"
 	}
 	return c
 }
 func lscpu(ctx context.Context) map[string]string {
 	f := map[string]string{}
 	out, err := exec.CommandContext(ctx, "lscpu").Output()
 	if err != nil {
 		// no util-linux: /proc/cpuinfo still names the thing and counts it
 		if b, err := os.ReadFile("/proc/cpuinfo"); err == nil {
 			n := 0
 			for _, line := range strings.Split(string(b), "\n") {
 				k, v, ok := strings.Cut(line, ":")
 				k, v = strings.TrimSpace(k), strings.TrimSpace(v)
 				switch {
 				case ok && strings.EqualFold(k, "model name"):
 					f["Model name"] = v
 				case ok && k == "processor":
 					n++
 				}
 			}
 			f["CPU(s)"] = strconv.Itoa(n)
 		}
 		return f
 	}
 	for _, line := range strings.Split(string(out), "\n") {
 		k, v, ok := strings.Cut(line, ":")
 		if !ok || strings.HasPrefix(line, " ") {
 			continue
 		}
 		f[strings.TrimSpace(k)] = strings.TrimSpace(v)
 	}
 	return f
 }
 func atoi(s string) int { n, _ := strconv.Atoi(strings.TrimSpace(s)); return n }
 func atof(s string) float64 {
 	v, _ := strconv.ParseFloat(strings.TrimSpace(s), 64)
 	return v
 }
 func maxInt(a, b int) int {
 	if a > b {
 		return a
 	}
 	return b
 }
 // leadingInt reads the number off the front of "3200 MT/s".
 func leadingInt(s string) int {
 	f := strings.Fields(s)
 	if len(f) == 0 {
 		return 0
 	}
 	return atoi(f[0])
 }
 // ---- memory ----
 var sizeRe = regexp.MustCompile(`^(\d+)\s*(kB|KB|MB|GB|TB)$`)
 // dmiSize turns dmidecode's "8 GB" / "16384 MB" into bytes.
 func dmiSize(s string) uint64 {
 	m := sizeRe.FindStringSubmatch(strings.TrimSpace(s))
 	if m == nil {
 		return 0
 	}
 	n, _ := strconv.ParseUint(m[1], 10, 64)
 	switch strings.ToUpper(m[2]) {
 	case "KB":
 		return n << 10
 	case "MB":
 		return n << 20
 	case "GB":
 		return n << 30
 	case "TB":
 		return n << 40
 	}
 	return 0
 }
 func memory(dmi []record) Memory {
 	m := Memory{}
 	if b, err := os.ReadFile("/proc/meminfo"); err == nil {
 		for _, line := range strings.Split(string(b), "\n") {
 			if k, v, ok := strings.Cut(line, ":"); ok && k == "MemTotal" {
 				m.Total = uint64(atoi(strings.TrimSuffix(strings.TrimSpace(v), " kB"))) << 10
 				break
 			}
 		}
 	}
 	for _, a := range all(dmi, 16) {
 		m.MaxCapacity += dmiSize(a.get("Maximum Capacity"))
 		switch e := a.V["Error Correction Type"]; {
 		case e == "None":
 			m.ECC = "non-ECC" // the array says so outright; worth knowing, not worth hiding
 		case e != "" && !unset(e):
 			m.ECC = e
 		}
 	}
 	for _, d := range all(dmi, 17) {
 		mod := Module{
 			Locator: d.get("Locator"),
 			Bank:    d.get("Bank Locator"),
 			Size:    dmiSize(d.get("Size")),
 		}
 		m.Slots++
 		if mod.Locator == "" {
 			mod.Locator = "slot " + strconv.Itoa(m.Slots)
 		}
 		if mod.Size == 0 {
 			mod.Empty = true
 			m.Modules = append(m.Modules, mod)
 			continue
 		}
 		m.Filled++
 		mod.Type = d.get("Type")
 		mod.FormFactor = d.get("Form Factor")
 		mod.SpeedMTs = leadingInt(d.get("Speed"))
 		mod.RunningMTs = leadingInt(d.get("Configured Memory Speed", "Configured Clock Speed"))
 		mod.Maker = makerName(d.get("Manufacturer"), d.V["Module Manufacturer ID"])
 		mod.Part = d.get("Part Number")
 		mod.Serial = d.get("Serial Number")
 		mod.Rank = atoi(d.get("Rank"))
 		mod.Voltage = d.get("Configured Voltage")
 		m.Installed += mod.Size
 		m.Modules = append(m.Modules, mod)
 	}
 	return m
 }
 // jedec maps the memory makers who turn up in a homelab to their names: SPD carries a
 // JEDEC id, and dmidecode prints the raw code when the module's string field is empty.
 var jedec = map[string]string{
 	"0X2C00": "Micron", "0X802C": "Micron",
 	"0XCE00": "Samsung", "0X80CE": "Samsung", "0X00CE": "Samsung",
 	"0XAD00": "SK hynix", "0X80AD": "SK hynix",
 	"0X9801": "Kingston", "0X7F98": "Kingston", "0X0198": "Kingston",
 	"0X859B": "Crucial", "0X04CB": "A-DATA", "0X04CD": "G.Skill",
 	"0X029E": "Corsair", "0X0B45": "Wilk Elektronik", "0X830B": "Wilk Elektronik",
 	"0X0443": "Ramaxel", "0X0194": "Smart Modular",
 }
 // makerName prefers what the module says it is; a bare JEDEC id becomes a name when it is
 // one we know, and stays the id when it is not — a wrong name is worse than a number.
 func makerName(v, id string) string {
 	if v != "" && !unset(v) && !strings.HasPrefix(v, "0x") {
 		return v
 	}
 	code := v
 	if code == "" {
 		code = id
 	}
 	if n, ok := jedec[strings.ToUpper(code)]; ok {
 		return n
 	}
 	if v != "" && !unset(v) {
 		return v
 	}
 	return ""
 }
 // ---- drives ----
 func drives(ctx context.Context) []Drive {
 	devs, err := storage.Devices(ctx)
 	if err != nil {
 		return nil
 	}
 	out := make([]Drive, 0, len(devs))
 	for _, d := range devs {
 		out = append(out, Drive{
 			Name: d.Name, Kind: d.Kind, Model: d.Model, Vendor: d.Vendor, Serial: d.Serial,
 			Tran: d.Tran, Size: d.Size, Rotational: d.Rotational, Removable: d.Removable,
 			Firmware: firmwareOf(d.Name),
 		})
 	}
 	return out
 }
 // firmwareOf reads a drive's firmware revision. An NVMe device is folded to its
 // controller ("nvme0"), which has no /sys/block entry of its own — the revision is on the
 // controller in /sys/class/nvme.
 func firmwareOf(name string) string {
 	return firstNonEmpty(
 		readFile("/sys/block/"+name+"/device/firmware_rev"),
 		readFile("/sys/block/"+name+"/device/rev"),
 		readFile("/sys/class/nvme/"+name+"/firmware_rev"),
 	)
 }
 func firstNonEmpty(v ...string) string {
 	for _, s := range v {
 		if s != "" {
 			return s
 		}
 	}
 	return ""
 }
 // ---- graphics ----
 // lspci -mm quotes every field, so a device name can hold the spaces and brackets it
 // likes: `"VGA compatible controller [0300]" "Intel Corporation [8086]" "Alder Lake-N …"`.
 var quoted = regexp.MustCompile(`"([^"]*)"`)
 // stripID drops the trailing PCI id from `Alder Lake-N [UHD Graphics] [46d1]`.
 func stripID(s string) string {
 	if i := strings.LastIndex(s, " ["); i > 0 && strings.HasSuffix(s, "]") {
 		return strings.TrimSpace(s[:i])
 	}
 	return s
 }
 func gpus(ctx context.Context) []GPU {
 	var out []GPU
 	seen := map[string]bool{} // PCI addresses already accounted for
 	if _, err := exec.LookPath("lspci"); err == nil {
 		drivers := pciDrivers(ctx)
 		if b, err := exec.CommandContext(ctx, "lspci", "-D", "-mm", "-nn").Output(); err == nil {
 			for _, line := range strings.Split(string(b), "\n") {
 				slot, _, _ := strings.Cut(line, " ")
 				fields := quoted.FindAllStringSubmatch(line, -1)
 				if len(fields) < 3 {
 					continue
 				}
 				kind := displayClass(fields[0][1])
 				if kind == "" {
 					continue
 				}
 				seen[slot] = true
 				g := GPU{Slot: slot, Class: kind, Vendor: stripID(fields[1][1]), Model: stripID(fields[2][1]), Driver: drivers[slot]}
 				g.Node, g.Memory = drmFor(slot)
 				out = append(out, g)
 			}
 		}
 	}
 	// A GPU that is not on a PCI bus at all — an Apple Silicon Mac, a Pi, anything with a
 	// device tree — is only visible as a kernel display device.
 	cards, _ := filepath.Glob("/sys/class/drm/card[0-9]")
 	for _, c := range cards {
 		dev, err := filepath.EvalSymlinks(c + "/device")
 		if err != nil || seen[filepath.Base(dev)] {
 			continue
 		}
 		g := GPU{Node: filepath.Base(c), Class: "Display"}
 		if drv, err := filepath.EvalSymlinks(dev + "/driver"); err == nil {
 			g.Driver = filepath.Base(drv)
 		}
 		g.Model = deviceTreeName(dev)
 		if g.Model == "" && g.Driver != "" {
 			g.Model = g.Driver + " display"
 		}
 		if g.Model == "" {
 			continue
 		}
 		g.Memory = vram(dev)
 		out = append(out, g)
 	}
 	sort.SliceStable(out, func(i, j int) bool { return out[i].Slot < out[j].Slot })
 	return out
 }
 // pciNames maps a PCI address to a readable name, so a slot the firmware says is in use
 // can say what is in it.
 func pciNames(ctx context.Context) map[string]string {
 	m := map[string]string{}
 	if _, err := exec.LookPath("lspci"); err != nil {
 		return m
 	}
 	out, err := exec.CommandContext(ctx, "lspci", "-D", "-mm", "-nn").Output()
 	if err != nil {
 		return m
 	}
 	for _, line := range strings.Split(string(out), "\n") {
 		slot, _, _ := strings.Cut(line, " ")
 		f := quoted.FindAllStringSubmatch(line, -1)
 		if slot == "" || len(f) < 3 {
 			continue
 		}
 		m[slot] = strings.TrimSpace(stripID(f[1][1]) + " " + stripID(f[2][1]))
 	}
 	return m
 }
 func displayClass(c string) string {
 	switch {
 	case strings.Contains(c, "[0300]"):
 		return "VGA"
 	case strings.Contains(c, "[0302]"):
 		return "3D"
 	case strings.Contains(c, "[0301]"), strings.Contains(c, "[0380]"):
 		return "Display"
 	}
 	return ""
 }
 // pciDrivers maps a PCI address to the kernel module driving it.
 func pciDrivers(ctx context.Context) map[string]string {
 	m := map[string]string{}
 	out, err := exec.CommandContext(ctx, "lspci", "-D", "-k").Output()
 	if err != nil {
 		return m
 	}
 	slot := ""
 	for _, line := range strings.Split(string(out), "\n") {
 		if !strings.HasPrefix(line, "\t") && !strings.HasPrefix(line, " ") {
 			slot, _, _ = strings.Cut(line, " ")
 			continue
 		}
 		if v, ok := strings.CutPrefix(strings.TrimSpace(line), "Kernel driver in use:"); ok && slot != "" {
 			m[slot] = strings.TrimSpace(v)
 		}
 	}
 	return m
 }
 // drmFor finds the kernel display device for a PCI address, and its video memory when the
 // driver publishes it (amdgpu does; Intel's integrated graphics have none of their own).
 func drmFor(slot string) (string, uint64) {
 	cards, _ := filepath.Glob("/sys/class/drm/card[0-9]")
 	for _, c := range cards {
 		dev, err := filepath.EvalSymlinks(c + "/device")
 		if err != nil || filepath.Base(dev) != slot {
 			continue
 		}
 		return filepath.Base(c), vram(dev)
 	}
 	return "", 0
 }
 func vram(dev string) uint64 {
 	for _, f := range []string{"mem_info_vram_total", "mem_info_vis_vram_total"} {
 		if v := readFile(dev + "/" + f); v != "" {
 			n, _ := strconv.ParseUint(v, 10, 64)
 			return n
 		}
 	}
 	return 0
 }
 // deviceTreeName reads what a device-tree machine calls its graphics: the compatible
 // string, tidied — "apple,agx-g13g" becomes "Apple agx-g13g".
 func deviceTreeName(dev string) string {
 	c := readFile(dev + "/of_node/compatible")
 	if c == "" {
 		return ""
 	}
 	c = strings.TrimRight(strings.Split(c, "\x00")[0], "\x00")
 	maker, rest, ok := strings.Cut(c, ",")
 	if !ok {
 		return c
 	}
 	if maker != "" {
 		maker = strings.ToUpper(maker[:1]) + maker[1:]
 	}
 	return strings.TrimSpace(maker + " " + rest)
 }
@@ -0,0 +1,103 @@
 package hardware
 import "testing"
 func TestDMISize(t *testing.T) {
 	for in, want := range map[string]uint64{
 		"8 GB": 8 << 30, "16384 MB": 16384 << 20, "512 kB": 512 << 10, "1 TB": 1 << 40,
 		"No Module Installed": 0, "": 0, "8 GiB": 0,
 	} {
 		if got := dmiSize(in); got != want {
 			t.Errorf("dmiSize(%q) = %d, want %d", in, got, want)
 		}
 	}
 }
 func TestMakerName(t *testing.T) {
 	cases := []struct{ v, id, want string }{
 		{"Samsung", "", "Samsung"},                         // the module says who it is
 		{"0x0B45", "Bank 12, Hex 0x45", "Wilk Elektronik"}, // only a JEDEC id, and one we know
 		{"0xFFFF", "", "0xFFFF"},                           // one we don't: the number, not a guess
 		{"Unknown", "", ""},
 		{"", "0xCE00", "Samsung"},
 	}
 	for _, c := range cases {
 		if got := makerName(c.v, c.id); got != c.want {
 			t.Errorf("makerName(%q, %q) = %q, want %q", c.v, c.id, got, c.want)
 		}
 	}
 }
 func TestStripIDAndDisplayClass(t *testing.T) {
 	if got := stripID("Alder Lake-N [UHD Graphics] [46d1]"); got != "Alder Lake-N [UHD Graphics]" {
 		t.Errorf("stripID = %q", got)
 	}
 	if got := stripID("no brackets here"); got != "no brackets here" {
 		t.Errorf("stripID left a plain name alone wrong: %q", got)
 	}
 	if displayClass(`VGA compatible controller [0300]`) != "VGA" {
 		t.Error("a VGA controller should be graphics")
 	}
 	if displayClass(`Ethernet controller [0200]`) != "" {
 		t.Error("a network card is not graphics")
 	}
 }
 func TestLeadingIntAndJoin(t *testing.T) {
 	if leadingInt("3200 MT/s") != 3200 || leadingInt("") != 0 || leadingInt("Unknown") != 0 {
 		t.Error("leadingInt")
 	}
 	if got := join("GMKtec", "GMKtec NucBox G3"); got != "GMKtec NucBox G3" {
 		t.Errorf("a maker the product already names should not be said twice: %q", got)
 	}
 	if got := join("Dell Inc.", "OptiPlex 7040"); got != "Dell Inc. OptiPlex 7040" {
 		t.Errorf("join = %q", got)
 	}
 	if got := join("To Be Filled By O.E.M.", "System Product Name"); got != "" {
 		t.Errorf("firmware boilerplate is not a model: %q", got)
 	}
 }
 const sample = `# dmidecode 3.6
 Handle 0x0028, DMI type 17, 92 bytes
 Memory Device
 	Array Handle: 0x0027
 	Size: 8 GB
 	Form Factor: SODIMM
 	Locator: Controller0-ChannelA-DIMM0
 	Type: DDR4
 	Speed: 3200 MT/s
 	Manufacturer: 0x0B45
 	Part Number: WPBH32D408SWD-8G
 	Configured Memory Speed: 2933 MT/s
 	Rank: 1
 Handle 0x0029, DMI type 17, 92 bytes
 Memory Device
 	Size: No Module Installed
 	Locator: Controller1-ChannelA-DIMM0
 	Type: Unknown
 `
 // parseRecords is what readDMI does once dmidecode has answered; this exercises it
 // without needing root or the tool.
 func TestMemoryFromRecords(t *testing.T) {
 	recs := parse(sample)
 	m := memory(recs)
 	if m.Slots != 2 || m.Filled != 1 {
 		t.Fatalf("slots %d filled %d, want 2 and 1", m.Slots, m.Filled)
 	}
 	mod := m.Modules[0]
 	if mod.Size != 8<<30 || mod.Type != "DDR4" || mod.Maker != "Wilk Elektronik" || mod.Part != "WPBH32D408SWD-8G" {
 		t.Errorf("first module read wrong: %+v", mod)
 	}
 	if mod.SpeedMTs != 3200 || mod.RunningMTs != 2933 {
 		t.Errorf("a module running below its rating should say both: %d / %d", mod.RunningMTs, mod.SpeedMTs)
 	}
 	if !m.Modules[1].Empty || m.Modules[1].Locator != "Controller1-ChannelA-DIMM0" {
 		t.Errorf("empty slot read wrong: %+v", m.Modules[1])
 	}
 	if m.Installed != 8<<30 {
 		t.Errorf("installed = %d", m.Installed)
 	}
 }
@@ -0,0 +1,90 @@
 package incus
 import (
 	"os/exec"
 	"strings"
 	"testing"
 )
 // The install script is generated, so a quoting slip would only show on a live machine.
 // Build it for the shapes of machine we have and let sh check the syntax.
 func TestInstallScriptShapes(t *testing.T) {
 	if _, err := exec.LookPath("sh"); err != nil {
 		t.Skip("no sh")
 	}
 	m := New("/home/core/apps/hamlab", "core", "test")
 	cases := []struct {
 		name string
 		st   Status
 		o    InstallOpts
 		want []string
 	}{
 		{"grid: zfs, apt, docker", Status{State: "installable", PM: "apt", Arch: "x86_64", Drivers: []string{"zfs", "btrfs", "dir"}, Docker: true, ZFSHost: true, DiskFree: 700 << 30, MemTotal: 62 << 30},
 			InstallOpts{Driver: "zfs", PoolSizeGiB: 300, Subnet: "172.16.100.1/24", DockerRules: true, ARCMaxGiB: 16},
 			[]string{"apt-get", "zpool create", "incus storage create default zfs source=hamlab-virt", "ipv4.address='172.16.100.1/24'", "DOCKER-USER", "zfs_arc_max", "usermod -aG incus-admin 'core'", "SupplementaryGroups=incus-admin", "systemd-run"}},
 		{"core: btrfs, apt, docker, no tools yet", Status{State: "installable", PM: "apt", Arch: "x86_64", Drivers: []string{"btrfs", "dir"}, Docker: true, DiskFree: 170 << 30, MemTotal: 8 << 30},
 			InstallOpts{Driver: "btrfs", PoolSizeGiB: 40, Subnet: "172.16.100.1/24", DockerRules: true},
 			[]string{"btrfs-progs", "mkfs.btrfs", "/etc/fstab", "incus storage create default btrfs source="}},
 		{"mac: dir, dnf, arm64", Status{State: "installable", PM: "dnf", Arch: "aarch64", Drivers: []string{"dir"}, DiskFree: 200 << 30, MemTotal: 15 << 30},
 			InstallOpts{Driver: "dir", Subnet: "172.16.101.1/24"},
 			[]string{"dnf -y install incus qemu-system-aarch64-core edk2-aarch64", "incus storage create default dir source="}},
 		{"repair: already installed", Status{State: "repair", PM: "apt", Arch: "x86_64", Drivers: []string{"dir"}, DiskFree: 10 << 30, MemTotal: 4 << 30},
 			InstallOpts{Driver: "dir", Subnet: "172.16.100.1/24"},
 			[]string{"already installed"}},
 	}
 	for _, c := range cases {
 		script, lay, err := m.InstallScript(c.st, c.o)
 		if err != nil {
 			t.Fatalf("%s: %v", c.name, err)
 		}
 		for _, w := range c.want {
 			if !strings.Contains(script, w) {
 				t.Errorf("%s: script lacks %q", c.name, w)
 			}
 		}
 		if lay.Driver != c.o.Driver || lay.Subnet != c.o.Subnet {
 			t.Errorf("%s: layout %+v", c.name, lay)
 		}
 		cmd := exec.Command("sh", "-n")
 		cmd.Stdin = strings.NewReader(script)
 		if out, err := cmd.CombinedOutput(); err != nil {
 			t.Errorf("%s: sh -n: %v\n%s", c.name, err, out)
 		}
 	}
 }
 func TestInstallScriptRefusals(t *testing.T) {
 	m := New("/x", "core", "test")
 	st := Status{State: "installable", PM: "apt", Drivers: []string{"btrfs", "dir"}, DiskFree: 20 << 30, MemTotal: 8 << 30}
 	for _, o := range []InstallOpts{
 		{Driver: "zfs", PoolSizeGiB: 10, Subnet: "172.16.100.1/24"},                  // not offered here
 		{Driver: "btrfs", PoolSizeGiB: 50, Subnet: "172.16.100.1/24"},                // more than is free
 		{Driver: "btrfs", PoolSizeGiB: 10, Subnet: "172.16.100.0/24"},                // the network address, not the bridge's
 		{Driver: "btrfs", PoolSizeGiB: 10, Subnet: "10.0.0.1/8"},                     // too wide
 		{Driver: "btrfs", PoolSizeGiB: 10, Subnet: "172.16.100.1/24", ARCMaxGiB: 64}, // more RAM than there is
 	} {
 		if _, _, err := m.InstallScript(st, o); err == nil {
 			t.Errorf("accepted %+v", o)
 		}
 	}
 }
 func TestCreateArgs(t *testing.T) {
 	args, err := CreateArgs(CreateSpec{Name: "web-1", Image: "images:debian/13", VM: true, CPU: "2", Memory: "2GiB", Disk: "20GiB", Network: "incusbr0", Start: true}, true)
 	if err != nil {
 		t.Fatal(err)
 	}
 	want := "launch images:debian/13 web-1 --vm -c limits.cpu=2 -c limits.memory=2GiB -d root,size=20GiB -n incusbr0"
 	if got := strings.Join(args, " "); got != want {
 		t.Fatalf("got %q", got)
 	}
 	if _, err := CreateArgs(CreateSpec{Name: "web-1", Image: "images:debian/13", VM: true}, false); err == nil {
 		t.Fatal("a VM without KVM must be refused")
 	}
 	if _, err := CreateArgs(CreateSpec{Name: "1bad", Image: "images:debian/13"}, true); err == nil {
 		t.Fatal("a name starting with a digit must be refused")
 	}
 	if _, err := CreateArgs(CreateSpec{Name: "ok", Image: "images:debian/13", Memory: "2 gigs"}, true); err == nil {
 		t.Fatal("a malformed size must be refused")
 	}
 }
@@ -0,0 +1,130 @@
 package incus
 // The bridge subnet. Incus's own `auto` picks a random 10.x.y.0/24, and on a LAN that is
 // itself 10.0.0.0/8 (this one) every such pick is on-link and collides. So Ham Lab chooses:
 // the first /24 that nothing on the machine already routes, is not an address it holds, and
 // is outside Docker's configured pools — tried in an order that also stays clear of Docker's
 // *defaults* (172.17–172.31 and 192.168.0.0/16 carved into /20s), which a Docker that has no
 // daemon.json will hand out later without asking anyone.
 import (
 	"context"
 	"encoding/json"
 	"fmt"
 	"net/netip"
 	"os"
 	"os/exec"
 	"strings"
 	"time"
 )
 // candidates in order of preference: 172.16.0.0/16 is the one RFC 1918 block Docker never
 // takes by default; then the top of 192.168/16, which Docker reaches last; then the rest.
 func candidates() []netip.Prefix {
 	var out []netip.Prefix
 	for i := 100; i <= 254; i++ {
 		out = append(out, netip.MustParsePrefix(fmt.Sprintf("172.16.%d.0/24", i)))
 	}
 	for i := 254; i >= 200; i-- {
 		out = append(out, netip.MustParsePrefix(fmt.Sprintf("192.168.%d.0/24", i)))
 	}
 	for i := 1; i < 100; i++ {
 		out = append(out, netip.MustParsePrefix(fmt.Sprintf("172.16.%d.0/24", i)))
 	}
 	return out
 }
 func overlaps(a, b netip.Prefix) bool {
 	return a.Contains(b.Addr()) || b.Contains(a.Addr())
 }
 // PickSubnet returns the first candidate /24 clear of every prefix in taken.
 func PickSubnet(taken []netip.Prefix) netip.Prefix {
 	for _, c := range candidates() {
 		clear := true
 		for _, t := range taken {
 			if overlaps(c, t) {
 				clear = false
 				break
 			}
 		}
 		if clear {
 			return c
 		}
 	}
 	return netip.MustParsePrefix("172.16.100.0/24") // every candidate taken: not a machine this will run on
 }
 // Gateway is the bridge's own address in a subnet: its first host, as CIDR, the way
 // Incus's ipv4.address wants it.
 func Gateway(p netip.Prefix) string {
 	return fmt.Sprintf("%s/%d", p.Masked().Addr().Next(), p.Bits())
 }
 // TakenPrefixes reads what the machine already uses: its routes, its addresses, and the
 // pools Docker is configured to carve networks from.
 func TakenPrefixes(ctx context.Context) []netip.Prefix {
 	var out []netip.Prefix
 	add := func(s string) {
 		if p, err := netip.ParsePrefix(s); err == nil && p.Addr().Is4() {
 			out = append(out, p.Masked())
 		} else if a, err := netip.ParseAddr(s); err == nil && a.Is4() {
 			out = append(out, netip.PrefixFrom(a, 32))
 		}
 	}
 	c, cancel := context.WithTimeout(ctx, 5*time.Second)
 	defer cancel()
 	if b, err := exec.CommandContext(c, "ip", "-j", "-4", "route", "show").Output(); err == nil {
 		var routes []struct{ Dst string }
 		if json.Unmarshal(b, &routes) == nil {
 			for _, r := range routes {
 				if r.Dst != "" && r.Dst != "default" {
 					add(r.Dst)
 				}
 			}
 		}
 	}
 	if b, err := exec.CommandContext(c, "ip", "-j", "-4", "addr", "show").Output(); err == nil {
 		var links []struct {
 			AddrInfo []struct {
 				Local     string `json:"local"`
 				Prefixlen int    `json:"prefixlen"`
 			} `json:"addr_info"`
 		}
 		if json.Unmarshal(b, &links) == nil {
 			for _, l := range links {
 				for _, a := range l.AddrInfo {
 					add(fmt.Sprintf("%s/%d", a.Local, a.Prefixlen))
 				}
 			}
 		}
 	}
 	for _, p := range dockerPools() {
 		add(p)
 	}
 	return out
 }
 // dockerPools reads default-address-pools from Docker's daemon.json, and falls back to
 // Docker's built-in defaults when the file names none.
 func dockerPools() []string {
 	var cfg struct {
 		Pools []struct {
 			Base string `json:"base"`
 		} `json:"default-address-pools"`
 	}
 	if b, err := os.ReadFile("/etc/docker/daemon.json"); err == nil && json.Unmarshal(b, &cfg) == nil && len(cfg.Pools) > 0 {
 		var out []string
 		for _, p := range cfg.Pools {
 			out = append(out, p.Base)
 		}
 		return out
 	}
 	if _, err := exec.LookPath("docker"); err != nil {
 		return nil
 	}
 	return []string{"172.17.0.0/12", "192.168.0.0/16"}
 }
 // String form used in the layout file and on the page.
 func prefixString(p netip.Prefix) string { return strings.TrimSpace(p.Masked().String()) }
@@ -0,0 +1,41 @@
 package incus
 import (
 	"net/netip"
 	"testing"
 )
 func TestPickSubnetAvoidsTaken(t *testing.T) {
 	taken := []netip.Prefix{
 		netip.MustParsePrefix("10.0.0.0/8"),      // the LAN, a /8
 		netip.MustParsePrefix("172.16.100.0/24"), // first choice gone
 		netip.MustParsePrefix("172.16.101.0/24"), // and the second
 		netip.MustParsePrefix("172.28.0.0/14"),   // Docker's configured pool
 		netip.MustParsePrefix("192.168.80.0/20"), // a Docker network
 	}
 	got := PickSubnet(taken)
 	if got.String() != "172.16.102.0/24" {
 		t.Fatalf("got %s", got)
 	}
 	if Gateway(got) != "172.16.102.1/24" {
 		t.Fatalf("gateway %s", Gateway(got))
 	}
 }
 func TestPickSubnetWholeBlockTaken(t *testing.T) {
 	// 172.16.0.0/12 covers every 172.16.x candidate; the picker must move to 192.168
 	got := PickSubnet([]netip.Prefix{netip.MustParsePrefix("172.16.0.0/12")})
 	if got.String() != "192.168.254.0/24" {
 		t.Fatalf("got %s", got)
 	}
 }
 func TestOverlapsEitherWay(t *testing.T) {
 	a := netip.MustParsePrefix("172.16.5.0/24")
 	if !overlaps(a, netip.MustParsePrefix("172.16.0.0/12")) || !overlaps(netip.MustParsePrefix("172.16.0.0/12"), a) {
 		t.Fatal("a /24 inside a /12 overlaps both ways")
 	}
 	if overlaps(a, netip.MustParsePrefix("172.16.6.0/24")) {
 		t.Fatal("neighbours do not overlap")
 	}
 }
@@ -0,0 +1,113 @@
 package metrics
 // The ZFS ARC — the read cache in RAM — read live from /proc/spl/kstat/zfs/arcstats on a
 // machine that has the module loaded, and left out of the sample everywhere else. The kstat
 // file is counters since boot; the hit rate here is over the sampling interval, which is the
 // number that says whether the cache is doing its job *now*. The lifetime figures and the
 // pool's caching tiers (L2ARC, SLOG, special vdevs) are the storage package's business.
 import (
 	"strconv"
 	"strings"
 	"time"
 )
 const arcstatsPath = "/proc/spl/kstat/zfs/arcstats"
 // ARC is the read cache as one sample sees it. Sizes are bytes.
 type ARC struct {
 	Size   uint64 `json:"size"`
 	Target uint64 `json:"target"` // c: what the ARC is currently allowed to hold
 	Min    uint64 `json:"min"`    // c_min: the floor it will not shrink below
 	Max    uint64 `json:"max"`    // c_max: the ceiling (zfs_arc_max, or half of RAM)
 	// what is in it
 	MRU          uint64 `json:"mru"`      // recently used once
 	MFU          uint64 `json:"mfu"`      // used more than once
 	Anon         uint64 `json:"anon"`     // in flight: dirty, not yet on disk
 	Data         uint64 `json:"data"`     // file contents
 	Metadata     uint64 `json:"metadata"` // dnodes, indirect blocks, dbufs, bonus buffers
 	MetaUsed     uint64 `json:"meta_used"`
 	Headers      uint64 `json:"headers"`
 	Compressed   uint64 `json:"compressed"`
 	Uncompressed uint64 `json:"uncompressed"`
 	MRUGhost     uint64 `json:"mru_ghost"` // what it recently evicted and still remembers
 	MFUGhost     uint64 `json:"mfu_ghost"`
 	// L2ARC, the cache devices (0 when there are none)
 	L2Size  uint64 `json:"l2_size"`
 	L2ASize uint64 `json:"l2_asize"` // as stored on the device, after compression
 	// over this interval
 	ReadsPerSec  float64 `json:"reads_per_sec"`
 	MissesPerSec float64 `json:"misses_per_sec"`
 	HitPct       float64 `json:"hit_pct"`       // of the interval's reads; an idle interval keeps the last one
 	HitPctKnown  bool    `json:"hit_pct_known"` // false until an interval has had any reads
 	L2HitPct     float64 `json:"l2_hit_pct"`
 	// since boot
 	Hits           uint64  `json:"hits"`
 	Misses         uint64  `json:"misses"`
 	LifetimeHitPct float64 `json:"lifetime_hit_pct"`
 	// pressure
 	NoGrow    bool   `json:"no_grow"`   // arc_no_grow: memory pressure is stopping it growing
 	Throttles uint64 `json:"throttles"` // memory_throttle_count since boot
 	Prunes    uint64 `json:"prunes"`    // arc_prune: metadata evictions asked of the filesystems
 	EvictSkip uint64 `json:"evict_skip"`
 }
 type arcPrev struct {
 	hits, misses, l2hits, l2misses uint64
 	hitPct                         float64
 	known                          bool
 	at                             time.Time
 }
 // readKstat parses a kstat table ("name type data" rows after two header lines).
 func readKstat(p string) map[string]uint64 {
 	s := readStr(p)
 	if s == "" {
 		return nil
 	}
 	out := map[string]uint64{}
 	for _, line := range strings.Split(s, "\n") {
 		f := strings.Fields(line)
 		if len(f) != 3 || f[1] != "4" { // type 4 is KSTAT_DATA_UINT64
 			continue
 		}
 		v, err := strconv.ParseUint(f[2], 10, 64)
 		if err == nil {
 			out[f[0]] = v
 		}
 	}
 	return out
 }
 // arc reads the ARC; nil when this machine has no ZFS module loaded.
 func (c *Collector) arc(now time.Time) *ARC {
 	k := readKstat(arcstatsPath)
 	if k == nil {
 		return nil
 	}
 	a := &ARC{Size: k["size"], Target: k["c"], Min: k["c_min"], Max: k["c_max"],
 		MRU: k["mru_size"], MFU: k["mfu_size"], Anon: k["anon_size"], Data: k["data_size"], Metadata: k["metadata_size"],
 		MetaUsed: k["arc_meta_used"], Headers: k["hdr_size"], Compressed: k["compressed_size"], Uncompressed: k["uncompressed_size"],
 		MRUGhost: k["mru_ghost_size"], MFUGhost: k["mfu_ghost_size"], L2Size: k["l2_size"], L2ASize: k["l2_asize"],
 		Hits: k["hits"], Misses: k["misses"], NoGrow: k["arc_no_grow"] != 0, Throttles: k["memory_throttle_count"], Prunes: k["arc_prune"], EvictSkip: k["evict_skip"]}
 	if n := a.Hits + a.Misses; n > 0 {
 		a.LifetimeHitPct = round1(float64(a.Hits) / float64(n) * 100)
 	}
 	p := c.prevARC
 	if !p.at.IsZero() && a.Hits >= p.hits && a.Misses >= p.misses {
 		if dt := now.Sub(p.at).Seconds(); dt > 0 {
 			hits, misses := a.Hits-p.hits, a.Misses-p.misses
 			a.ReadsPerSec = round1(float64(hits+misses) / dt)
 			a.MissesPerSec = round1(float64(misses) / dt)
 			if hits+misses > 0 {
 				p.hitPct, p.known = round1(float64(hits)/float64(hits+misses)*100), true
 			}
 			if l2 := (k["l2_hits"] - p.l2hits) + (k["l2_misses"] - p.l2misses); l2 > 0 && k["l2_hits"] >= p.l2hits {
 				a.L2HitPct = round1(float64(k["l2_hits"]-p.l2hits) / float64(l2) * 100)
 			}
 		}
 	}
 	a.HitPct, a.HitPctKnown = p.hitPct, p.known
 	c.prevARC = arcPrev{hits: a.Hits, misses: a.Misses, l2hits: k["l2_hits"], l2misses: k["l2_misses"], hitPct: p.hitPct, known: p.known, at: now}
 	return a
 }
@@ -0,0 +1,680 @@
 package metrics
 // GPU telemetry, the way btop and intel_gpu_top get it without a vendor library: the
 // device from sysfs, and who is using it from DRM fdinfo.
 //
 //   - The device: /sys/class/drm/card<n> with a driver we know to be a GPU. Clock from the
 //     driver's frequency files, temperature/power/fan from the hwmon under the device, and
 //     how busy it is from the one counter every Intel GT exports without root — RC6 idle
 //     residency, so busy = 100 − idle. amdgpu says it outright (gpu_busy_percent).
 //   - The clients: every process holding /dev/dri/* open has a drm-engine-<name> running
 //     total (nanoseconds) in /proc/<pid>/fdinfo/<fd>; the delta between two reads is that
 //     process's share of each engine, and the sum across clients is the engine's load. Other
 //     users' fdinfo is unreadable, so when the daemon isn't root it keeps one root child
 //     (`hamlab drm-clients`) that does the walk, asked over a pipe — one sudo per daemon
 //     start, not one per second.
 //
 // A GPU with no counter of either kind (Asahi, today) is still listed: a name is better
 // than silence about hardware the machine has.
 import (
 	"bufio"
 	"encoding/json"
 	"io"
 	"log"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"regexp"
 	"sort"
 	"strconv"
 	"strings"
 	"sync"
 	"time"
 	"hamlab/internal/priv"
 )
 // GPU is one graphics device as a Sample reports it.
 type GPU struct {
 	Name       string             `json:"name"`
 	Driver     string             `json:"driver"`
 	Bus        string             `json:"bus,omitempty"` // PCI address; empty for a platform device
 	Integrated bool               `json:"integrated,omitempty"`
 	Usage      float64            `json:"usage"`               // % busy
 	UsageSrc   string             `json:"usage_src,omitempty"` // "idle" (100 − idle residency), "busy" (the driver's own %), "clients" (the busiest engine); "" when nothing measures it
 	Engines    map[string]float64 `json:"engines,omitempty"`   // render/video/copy/enhance… % from the clients, normalised by engine count
 	Temp       float64            `json:"temp,omitempty"`
 	TempSrc    string             `json:"temp_src,omitempty"` // "gpu" from the device's own sensor, "package" the CPU die an integrated GPU shares
 	PowerW     float64            `json:"power_w,omitempty"`
 	PowerCapW  float64            `json:"power_cap_w,omitempty"`
 	FanRPM     int                `json:"fan_rpm,omitempty"`
 	MHz        float64            `json:"mhz,omitempty"`     // actual clock (0 when power-gated)
 	ReqMHz     float64            `json:"req_mhz,omitempty"` // what the driver asked for
 	MaxMHz     float64            `json:"max_mhz,omitempty"`
 	MemUsed    uint64             `json:"mem_used,omitempty"` // VRAM used, or the clients' resident bytes where the driver has no VRAM counter
 	MemTotal   uint64             `json:"mem_total,omitempty"`
 	Clients    []GPUClient        `json:"clients"`
 	ClientsOK  bool               `json:"clients_ok"`        // false when fdinfo of other users' processes could not be read
 	Stalled    bool               `json:"stalled,omitempty"` // a sysfs read of this card never returned; its readings are paused until restart
 }
 // GPUClient is one process with the device open.
 type GPUClient struct {
 	PID     int                `json:"pid"`
 	Name    string             `json:"name"`
 	Usage   float64            `json:"usage"` // its busiest engine, %
 	Engines map[string]float64 `json:"engines,omitempty"`
 	Mem     uint64             `json:"mem,omitempty"` // resident bytes
 }
 // gpuDrivers are DRM drivers that drive a GPU, not just a scanout (simpledrm, ast, the
 // Apple display coprocessor) — those have no engines and nothing to say.
 var gpuDrivers = map[string]bool{"i915": true, "xe": true, "amdgpu": true, "radeon": true, "nouveau": true, "nvidia-drm": true,
 	"asahi": true, "panfrost": true, "panthor": true, "v3d": true, "msm": true, "etnaviv": true, "lima": true, "powervr": true, "virtio_gpu": true}
 type gpuDev struct {
 	card, dev, key, bus, driver, name, hwmon string
 	integrated, stalled                      bool
 	prevIdle, prevEnergy                     float64
 	prevAt                                   time.Time
 }
 type gpuState struct {
 	devs     []*gpuDev
 	listedAt time.Time
 	clients  map[string]clientPrev // device key + client id → engine totals at the previous scan
 	last     []drmFD               // the newest client walk, reused between walks
 	scanAt   time.Time
 	helper   *drmHelper
 	ok       bool // the last walk saw every process
 }
 type clientPrev struct {
 	ns map[string]uint64
 	at time.Time
 }
 const gpuRelist = time.Minute    // how often the card list is re-read: hotplug is rare, the readdir is cheap
 const gpuScanEvery = time.Second // engine shares over less than this are noise
 // gpus is the GPU section of a sample. cpuTemp stands in for an integrated GPU's
 // temperature: it is the same die.
 func (c *Collector) gpus(now time.Time, cpuTemp float64) []GPU {
 	g := &c.gpu
 	if g.devs == nil || now.Sub(g.listedAt) >= gpuRelist {
 		g.devs, g.listedAt = listGPUs(g.devs), now
 	}
 	if len(g.devs) == 0 {
 		return []GPU{}
 	}
 	if g.clients == nil {
 		g.clients = map[string]clientPrev{}
 	}
 	if g.scanAt.IsZero() || now.Sub(g.scanAt) >= gpuScanEvery {
 		g.last, g.ok = c.scanClients()
 		g.scanAt = now
 	}
 	out := make([]GPU, 0, len(g.devs))
 	live := map[string]bool{}
 	for _, d := range g.devs {
 		x := GPU{Name: d.name, Driver: d.driver, Bus: d.bus, Integrated: d.integrated, Clients: []GPUClient{}, ClientsOK: g.ok}
 		d.readWithin(&x, now)
 		if x.Temp == 0 && d.integrated && cpuTemp > 0 {
 			x.Temp, x.TempSrc = cpuTemp, "package"
 		}
 		// the clients: one entry per client id, under the first pid that holds it (a child
 		// that inherited the fd shares the client)
 		byID := map[string]*GPUClient{}
 		var order []string
 		engines := map[string]float64{}
 		var resident uint64
 		for _, f := range g.last {
 			fi := parseFdinfo(f.Info)
 			if fi.pdev != "" && fi.pdev != d.bus || fi.pdev == "" && fi.driver != d.driver {
 				continue
 			}
 			key := d.key + "/" + fi.client
 			live[key] = true
 			if byID[fi.client] != nil {
 				continue
 			}
 			cl := &GPUClient{PID: f.PID, Name: f.Comm, Mem: fi.mem}
 			byID[fi.client] = cl
 			order = append(order, fi.client)
 			resident += fi.mem
 			if prev, ok := g.clients[key]; ok {
 				if dt := now.Sub(prev.at).Nanoseconds(); dt > 0 {
 					for e, ns := range fi.engines {
 						p, ok := prev.ns[e]
 						if !ok || ns < p {
 							continue
 						}
 						pct := float64(ns-p) / float64(dt) * 100
 						if n := fi.capacity[e]; n > 1 {
 							pct /= float64(n)
 						}
 						pct = round1(pct)
 						if cl.Engines == nil {
 							cl.Engines = map[string]float64{}
 						}
 						cl.Engines[e] = pct
 						engines[e] += pct
 						if pct > cl.Usage {
 							cl.Usage = pct
 						}
 					}
 				}
 			}
 			g.clients[key] = clientPrev{ns: fi.engines, at: now}
 		}
 		for _, id := range order {
 			x.Clients = append(x.Clients, *byID[id])
 		}
 		sort.SliceStable(x.Clients, func(i, j int) bool { return x.Clients[i].Usage > x.Clients[j].Usage })
 		if len(engines) > 0 {
 			x.Engines = map[string]float64{}
 			for e, v := range engines {
 				if v > 100 {
 					v = 100
 				}
 				x.Engines[e] = round1(v)
 			}
 		}
 		if x.UsageSrc == "" && len(x.Engines) > 0 {
 			for _, v := range x.Engines {
 				if v > x.Usage {
 					x.Usage = v
 				}
 			}
 			x.UsageSrc = "clients"
 		}
 		if x.MemTotal == 0 && x.MemUsed == 0 {
 			x.MemUsed = resident
 		}
 		out = append(out, x)
 	}
 	for k := range g.clients { // a client that closed the device
 		if !live[k] {
 			delete(g.clients, k)
 		}
 	}
 	return out
 }
 // listGPUs enumerates the cards, keeping the counters of those already known.
 func listGPUs(known []*gpuDev) []*gpuDev {
 	ents, _ := os.ReadDir("/sys/class/drm")
 	old := map[string]*gpuDev{}
 	for _, d := range known {
 		old[d.card] = d
 	}
 	out := []*gpuDev{}
 	for _, e := range ents {
 		n := e.Name()
 		if !strings.HasPrefix(n, "card") || strings.Contains(n, "-") {
 			continue
 		}
 		card := "/sys/class/drm/" + n
 		dev, err := filepath.EvalSymlinks(card + "/device")
 		if err != nil {
 			continue
 		}
 		driver := ueventValue(dev+"/uevent", "DRIVER")
 		if driver == "" { // some drivers hang the DRIVER key off the parent
 			driver = filepath.Base(readLink(dev + "/driver"))
 		}
 		if !gpuDrivers[driver] {
 			continue
 		}
 		if d, ok := old[card]; ok && d.dev == dev {
 			out = append(out, d)
 			continue
 		}
 		d := &gpuDev{card: card, dev: dev, driver: driver}
 		if pci := ueventValue(dev+"/uevent", "PCI_SLOT_NAME"); pci != "" {
 			d.bus, d.key = pci, pci
 			// an on-die GPU hangs off root bus 0 (Intel's 00:02.0, an AMD APU's 00:08.0 bridge)
 			d.integrated = strings.HasPrefix(pci, "0000:00:")
 		} else {
 			d.key = "driver:" + driver
 			d.integrated = true
 		}
 		if hw, _ := filepath.Glob(dev + "/hwmon/hwmon*"); len(hw) > 0 {
 			d.hwmon = hw[0]
 		}
 		d.name = gpuName(dev, d.bus, driver)
 		out = append(out, d)
 	}
 	sort.Slice(out, func(i, j int) bool { return out[i].card < out[j].card })
 	return out
 }
 var gpuVendors = map[string]string{"0x8086": "Intel", "0x1002": "AMD", "0x10de": "NVIDIA", "0x1af4": "Virtio", "0x106b": "Apple"}
 var (
 	revRe     = regexp.MustCompile(`\s*\(rev [^)]*\)`)
 	bracketRe = regexp.MustCompile(`\[([^\]]+)\]`)
 )
 // gpuName is "Intel Arc A380": from the embedded PCI ID table first, so the name is the
 // same on every machine; from lspci for a card newer than the table; and failing both,
 // the vendor and the raw device ID.
 func gpuName(dev, bus, driver string) string {
 	if bus != "" {
 		vid := strings.TrimPrefix(strings.TrimSpace(readStr(dev+"/vendor")), "0x")
 		did := strings.TrimPrefix(strings.TrimSpace(readStr(dev+"/device")), "0x")
 		if vn, dn := pciName(vid, did); dn != "" {
 			return lspciName(vn + " " + dn)
 		}
 		if out, ok := runOut("lspci", "-s", bus); ok {
 			if _, desc, found := strings.Cut(out, ": "); found {
 				return lspciName(desc)
 			}
 		}
 	}
 	vendor := gpuVendors[strings.TrimSpace(readStr(dev+"/vendor"))]
 	if compat := readStr(dev + "/of_node/compatible"); compat != "" { // "apple,agx-t8103\0apple,agx-g13g"
 		first := strings.SplitN(compat, "\x00", 2)[0]
 		if v, model, ok := strings.Cut(first, ","); ok {
 			if vendor == "" && v != "" {
 				vendor = strings.ToUpper(v[:1]) + v[1:]
 			}
 			return vendor + " GPU (" + model + ")"
 		}
 	}
 	if vendor == "" && driver != "" {
 		vendor = strings.ToUpper(driver[:1]) + driver[1:]
 	}
 	if id := strings.TrimPrefix(strings.TrimSpace(readStr(dev+"/device")), "0x"); id != "" {
 		return strings.TrimSpace(vendor + " GPU " + id)
 	}
 	return strings.TrimSpace(vendor + " GPU")
 }
 // lspciName turns "Intel Corporation DG2 [Arc A380] (rev 05)" into "Intel Arc A380".
 func lspciName(desc string) string {
 	desc = revRe.ReplaceAllString(strings.TrimSpace(desc), "")
 	vendor := ""
 	for _, v := range []struct{ long, short string }{{"Intel Corporation", "Intel"}, {"Advanced Micro Devices, Inc. [AMD/ATI]", "AMD"}, {"Advanced Micro Devices, Inc. [AMD]", "AMD"}, {"NVIDIA Corporation", "NVIDIA"}, {"Red Hat, Inc.", "Virtio"}} {
 		if strings.HasPrefix(desc, v.long) {
 			vendor = v.short
 			desc = strings.TrimSpace(desc[len(v.long):])
 			break
 		}
 	}
 	if vendor == "" { // vendor not recognised: its first word
 		if i := strings.IndexByte(desc, ' '); i > 0 {
 			vendor, desc = desc[:i], strings.TrimSpace(desc[i+1:])
 		}
 	}
 	if m := bracketRe.FindStringSubmatch(desc); m != nil {
 		desc = m[1]
 	}
 	return strings.TrimSpace(vendor + " " + desc)
 }
 // readWithin is read with a deadline. A sysfs read that does not come back must not take
 // the whole sampler with it — every graph on every page of this machine stops — so the
 // device is marked stalled, its goroutine abandoned, and the rest of the sample goes on.
 const gpuReadWithin = 3 * time.Second
 func (d *gpuDev) readWithin(x *GPU, now time.Time) {
 	if d.stalled {
 		x.Stalled = true
 		return
 	}
 	done := make(chan GPU, 1)
 	go func() {
 		y := *x
 		d.read(&y, now)
 		done <- y
 	}()
 	select {
 	case y := <-done:
 		*x = y
 	case <-time.After(gpuReadWithin):
 		d.stalled, x.Stalled = true, true
 		log.Printf("gpu: %s (%s): a sysfs read did not return within %s; its readings are paused until restart", d.name, d.card, gpuReadWithin)
 	}
 }
 // read fills the device-level numbers: clock, idle/busy, hwmon.
 func (d *gpuDev) read(x *GPU, now time.Time) {
 	dt := now.Sub(d.prevAt).Seconds()
 	switch d.driver {
 	case "i915":
 		x.MHz = readF(d.card + "/gt_act_freq_mhz")
 		x.ReqMHz = readF(d.card + "/gt_cur_freq_mhz")
 		x.MaxMHz = readF(d.card + "/gt_max_freq_mhz")
 		idle, ok := readInt(d.card + "/gt/gt0/rc6_residency_ms")
 		if !ok {
 			idle, ok = readInt(d.card + "/power/rc6_residency_ms")
 		}
 		if ok {
 			d.busyFromIdle(x, float64(idle), dt)
 		}
 	case "xe":
 		gt := d.dev + "/tile0/gt0"
 		x.MHz = readF(gt + "/freq0/act_freq")
 		x.ReqMHz = readF(gt + "/freq0/cur_freq")
 		x.MaxMHz = readF(gt + "/freq0/max_freq")
 		if idle, ok := readInt(gt + "/gtidle/idle_residency_ms"); ok {
 			d.busyFromIdle(x, float64(idle), dt)
 		}
 	case "amdgpu":
 		if v, ok := readInt(d.dev + "/gpu_busy_percent"); ok {
 			x.Usage, x.UsageSrc = float64(v), "busy"
 		}
 		x.MemUsed = uint64(readF(d.dev + "/mem_info_vram_used"))
 		x.MemTotal = uint64(readF(d.dev + "/mem_info_vram_total"))
 		// "1: 1800Mhz *" — the starred level is the current one, the last the highest
 		for _, line := range strings.Split(readStr(d.dev+"/pp_dpm_sclk"), "\n") {
 			f := strings.Fields(line)
 			if len(f) < 2 {
 				continue
 			}
 			mhz := leadingNum(f[1])
 			x.MaxMHz = mhz
 			if strings.HasSuffix(line, "*") {
 				x.MHz = mhz
 			}
 		}
 	}
 	if d.hwmon != "" {
 		if t := readF(d.hwmon + "/temp1_input"); t > 0 {
 			x.Temp, x.TempSrc = t/1000, "gpu"
 		}
 		if p := readF(d.hwmon + "/power1_input"); p > 0 {
 			x.PowerW = round1(p / 1e6)
 		} else if p := readF(d.hwmon + "/power1_average"); p > 0 {
 			x.PowerW = round1(p / 1e6)
 		} else if e := readF(d.hwmon + "/energy1_input"); e > 0 { // µJ, cumulative
 			if d.prevEnergy > 0 && e >= d.prevEnergy && dt > 0 && !d.prevAt.IsZero() {
 				x.PowerW = round1((e - d.prevEnergy) / 1e6 / dt)
 			}
 			d.prevEnergy = e
 		}
 		if p := readF(d.hwmon + "/power1_cap"); p > 0 {
 			x.PowerCapW = p / 1e6
 		} else if p := readF(d.hwmon + "/power1_max"); p > 0 {
 			x.PowerCapW = p / 1e6
 		}
 		x.FanRPM = int(readF(d.hwmon + "/fan1_input"))
 	}
 	d.prevAt = now
 }
 // busyFromIdle: the GT's idle residency is a running ms total; the share of the interval
 // it did not cover is how busy the device was.
 func (d *gpuDev) busyFromIdle(x *GPU, idle, dt float64) {
 	x.UsageSrc = "idle"
 	if d.prevAt.IsZero() || dt <= 0 || idle < d.prevIdle {
 		d.prevIdle = idle
 		return
 	}
 	busy := 100 - (idle-d.prevIdle)/(dt*1000)*100
 	if busy < 0 {
 		busy = 0
 	}
 	if busy > 100 {
 		busy = 100
 	}
 	x.Usage = round1(busy)
 	d.prevIdle = idle
 }
 // ---- DRM fdinfo: who has the device open ----
 // drmFD is one client of a DRM device: the fdinfo text of one open fd, with its owner.
 type drmFD struct {
 	PID  int    `json:"pid"`
 	Comm string `json:"comm"`
 	Info string `json:"info"`
 }
 type fdinfo struct {
 	driver, pdev, client string
 	engines              map[string]uint64 // ns, cumulative
 	capacity             map[string]int
 	mem                  uint64 // bytes
 }
 func parseFdinfo(s string) fdinfo {
 	fi := fdinfo{engines: map[string]uint64{}, capacity: map[string]int{}}
 	var resident, memory uint64
 	for _, line := range strings.Split(s, "\n") {
 		k, v, ok := strings.Cut(line, ":")
 		if !ok || !strings.HasPrefix(k, "drm-") {
 			continue
 		}
 		v = strings.TrimSpace(v)
 		switch {
 		case k == "drm-driver":
 			fi.driver = v
 		case k == "drm-pdev":
 			fi.pdev = v
 		case k == "drm-client-id":
 			fi.client = v
 		case strings.HasPrefix(k, "drm-engine-capacity-"):
 			fi.capacity[k[len("drm-engine-capacity-"):]], _ = strconv.Atoi(v)
 		case strings.HasPrefix(k, "drm-engine-"):
 			n, _ := strconv.ParseUint(strings.TrimSuffix(v, " ns"), 10, 64)
 			fi.engines[k[len("drm-engine-"):]] = n
 		case strings.HasPrefix(k, "drm-resident-"):
 			resident += kib(v)
 		case strings.HasPrefix(k, "drm-memory-"):
 			memory += kib(v)
 		}
 	}
 	fi.mem = resident
 	if fi.mem == 0 {
 		fi.mem = memory
 	}
 	return fi
 }
 // kib parses "52624 KiB" (or MiB, or a bare count of bytes) to bytes.
 func kib(v string) uint64 {
 	f := strings.Fields(v)
 	if len(f) == 0 {
 		return 0
 	}
 	n, _ := strconv.ParseUint(f[0], 10, 64)
 	if len(f) > 1 {
 		switch f[1] {
 		case "KiB":
 			n <<= 10
 		case "MiB":
 			n <<= 20
 		case "GiB":
 			n <<= 30
 		}
 	}
 	return n
 }
 // drmScan walks /proc for every fd open on /dev/dri/*, one entry per client per process.
 // Run as root it sees everyone; as a user, only that user's processes.
 func drmScan() (out []drmFD, complete bool) {
 	ents, err := os.ReadDir("/proc")
 	if err != nil {
 		return []drmFD{}, false
 	}
 	complete = true
 	out = []drmFD{}
 	for _, e := range ents {
 		pid, err := strconv.Atoi(e.Name())
 		if err != nil {
 			continue
 		}
 		fds, err := os.ReadDir("/proc/" + e.Name() + "/fd")
 		if err != nil {
 			if !os.IsNotExist(err) {
 				complete = false
 			}
 			continue
 		}
 		seen := map[string]bool{}
 		comm := ""
 		for _, fd := range fds {
 			target, err := os.Readlink("/proc/" + e.Name() + "/fd/" + fd.Name())
 			if err != nil || !strings.HasPrefix(target, "/dev/dri/") {
 				continue
 			}
 			info := readStr("/proc/" + e.Name() + "/fdinfo/" + fd.Name())
 			_, rest, ok := strings.Cut(info, "drm-client-id:")
 			if !ok {
 				continue
 			}
 			id := strings.TrimSpace(strings.SplitN(rest, "\n", 2)[0])
 			if seen[id] {
 				continue
 			}
 			seen[id] = true
 			if comm == "" {
 				comm = strings.TrimSpace(readStr("/proc/" + e.Name() + "/comm"))
 			}
 			out = append(out, drmFD{PID: pid, Comm: comm, Info: info})
 		}
 	}
 	sort.SliceStable(out, func(i, j int) bool { return out[i].PID < out[j].PID })
 	return out, complete
 }
 // scanClients walks in-process when that sees everything (root), else through the helper.
 func (c *Collector) scanClients() ([]drmFD, bool) {
 	if priv.Mode() == "root" || !priv.Available() {
 		return drmScan()
 	}
 	if c.gpu.helper == nil {
 		c.gpu.helper = &drmHelper{}
 	}
 	if fds, ok := c.gpu.helper.scan(); ok {
 		return fds, true
 	}
 	return drmScan() // the helper is down: our own processes are better than nothing
 }
 // drmHelper is the root child that answers "scan" requests. One line in, one JSON line out.
 type drmHelper struct {
 	mu       sync.Mutex
 	cmd      *exec.Cmd
 	in       io.WriteCloser
 	out      *bufio.Reader
 	failedAt time.Time
 }
 const helperRetry = 30 * time.Second
 func (h *drmHelper) scan() ([]drmFD, bool) {
 	h.mu.Lock()
 	defer h.mu.Unlock()
 	if h.cmd == nil {
 		if !h.failedAt.IsZero() && time.Since(h.failedAt) < helperRetry {
 			return nil, false
 		}
 		if !h.start() {
 			h.failedAt = time.Now()
 			return nil, false
 		}
 	}
 	// a stuck child must not stall the sampler: it is killed, and tried again later
 	cmd := h.cmd
 	stop := time.AfterFunc(5*time.Second, func() { cmd.Process.Kill() })
 	defer stop.Stop()
 	if _, err := io.WriteString(h.in, "\n"); err != nil {
 		h.fail()
 		return nil, false
 	}
 	line, err := h.out.ReadBytes('\n')
 	if err != nil {
 		h.fail()
 		return nil, false
 	}
 	var fds []drmFD
 	if json.Unmarshal(line, &fds) != nil {
 		h.fail()
 		return nil, false
 	}
 	if fds == nil {
 		fds = []drmFD{}
 	}
 	return fds, true
 }
 func (h *drmHelper) start() bool {
 	self, err := os.Executable()
 	if err != nil {
 		return false
 	}
 	full := priv.Wrap(self, "drm-clients")
 	cmd := exec.Command(full[0], full[1:]...)
 	in, err := cmd.StdinPipe()
 	if err != nil {
 		return false
 	}
 	out, err := cmd.StdoutPipe()
 	if err != nil {
 		return false
 	}
 	if err := cmd.Start(); err != nil {
 		return false
 	}
 	h.cmd, h.in, h.out = cmd, in, bufio.NewReaderSize(out, 1<<20)
 	return true
 }
 func (h *drmHelper) fail() {
 	if h.cmd != nil {
 		h.in.Close()
 		h.cmd.Process.Kill()
 		go h.cmd.Wait()
 	}
 	h.cmd, h.in, h.out = nil, nil, nil
 	h.failedAt = time.Now()
 }
 // DRMClientsHelper is `hamlab drm-clients`: for every line read on stdin, one JSON line of
 // drmFD on stdout. Ends when stdin does — when the daemon that started it has gone.
 func DRMClientsHelper(stdin io.Reader, stdout io.Writer) int {
 	sc := bufio.NewScanner(stdin)
 	w := bufio.NewWriter(stdout)
 	for sc.Scan() {
 		fds, _ := drmScan()
 		b, _ := json.Marshal(fds)
 		w.Write(b)
 		w.WriteByte('\n')
 		if w.Flush() != nil {
 			return 1
 		}
 	}
 	return 0
 }
 // ---- small helpers ----
 func readF(p string) float64 {
 	v, err := strconv.ParseFloat(strings.TrimSpace(readStr(p)), 64)
 	if err != nil {
 		return 0
 	}
 	return v
 }
 func readLink(p string) string {
 	t, err := os.Readlink(p)
 	if err != nil {
 		return ""
 	}
 	return t
 }
 func ueventValue(p, key string) string {
 	for _, line := range strings.Split(readStr(p), "\n") {
 		if k, v, ok := strings.Cut(line, "="); ok && k == key {
 			return strings.TrimSpace(v)
 		}
 	}
 	return ""
 }
@@ -0,0 +1,172 @@
 package metrics
 import (
 	"encoding/json"
 	"os"
 	"strings"
 	"testing"
 	"time"
 )
 const fdinfoA = `pos:	0
 flags:	02100002
 drm-driver:	i915
 drm-client-id:	544
 drm-pdev:	0000:00:02.0
 drm-total-system0:	52624 KiB
 drm-resident-system0:	50140 KiB
 drm-resident-stolen-system0:	32 KiB
 drm-engine-render:	1000000000 ns
 drm-engine-copy:	0 ns
 drm-engine-video:	2000000000 ns
 drm-engine-capacity-video:	2
 drm-engine-video-enhance:	0 ns
 `
 func TestParseFdinfo(t *testing.T) {
 	fi := parseFdinfo(fdinfoA)
 	if fi.driver != "i915" || fi.pdev != "0000:00:02.0" || fi.client != "544" {
 		t.Fatalf("identity: %+v", fi)
 	}
 	if fi.engines["video"] != 2000000000 || fi.capacity["video"] != 2 {
 		t.Errorf("video engine: %+v", fi)
 	}
 	if fi.mem != (50140+32)<<10 {
 		t.Errorf("resident = %d", fi.mem)
 	}
 }
 func TestLspciName(t *testing.T) {
 	for in, want := range map[string]string{
 		"Intel Corporation DG2 [Arc A380] (rev 05)":                                              "Intel Arc A380",
 		"Intel Corporation RocketLake-S GT1 [UHD Graphics 750] (rev 04)":                         "Intel UHD Graphics 750",
 		"Intel Corporation Alder Lake-N [UHD Graphics]":                                          "Intel UHD Graphics",
 		"Advanced Micro Devices, Inc. [AMD/ATI] Navi 23 [Radeon RX 6600/6600 XT/6600M] (rev c7)": "AMD Radeon RX 6600/6600 XT/6600M",
 		"NVIDIA Corporation GA106 [GeForce RTX 3060 Lite Hash Rate] (rev a1)":                    "NVIDIA GeForce RTX 3060 Lite Hash Rate",
 		"Moore Threads Technology Co.,Ltd MTT S80":                                               "Moore Threads Technology Co.,Ltd MTT S80",
 	} {
 		if got := lspciName(in); got != want {
 			t.Errorf("%q → %q, want %q", in, got, want)
 		}
 	}
 }
 // Two scans a second apart: a client that ran its render engine for half of it and both
 // video engines flat out reads 50% render, 100% video, and the device's engines follow.
 func TestGPUClientsShare(t *testing.T) {
 	c := NewCollector()
 	d := &gpuDev{card: "/nonexistent/card9", dev: "/nonexistent", key: "0000:00:02.0", bus: "0000:00:02.0", driver: "i915", name: "Test GPU", integrated: true}
 	c.gpu.devs = []*gpuDev{d}
 	c.gpu.listedAt = time.Now().Add(time.Hour) // keep the fake list
 	c.gpu.clients = map[string]clientPrev{}
 	t0 := time.Now()
 	c.gpu.last = []drmFD{{PID: 7, Comm: "ffmpeg", Info: fdinfoA}}
 	c.gpu.scanAt = t0 // no walk: the fake list stands
 	first := c.gpus(t0, 40)
 	if len(first) != 1 || len(first[0].Clients) != 1 || first[0].Clients[0].Usage != 0 {
 		t.Fatalf("first pass: %+v", first)
 	}
 	if first[0].Temp != 40 || first[0].TempSrc != "package" {
 		t.Errorf("an integrated GPU wears the package temperature: %+v", first[0])
 	}
 	later := fdinfoA
 	later = replace(later, "drm-engine-render:	1000000000 ns", "drm-engine-render:	1500000000 ns")
 	later = replace(later, "drm-engine-video:	2000000000 ns", "drm-engine-video:	4000000000 ns")
 	c.gpu.last = []drmFD{{PID: 7, Comm: "ffmpeg", Info: later}}
 	c.gpu.scanAt = t0.Add(time.Second)
 	got := c.gpus(t0.Add(time.Second), 40)
 	cl := got[0].Clients[0]
 	if cl.Engines["render"] != 50 || cl.Engines["video"] != 100 || cl.Usage != 100 {
 		t.Errorf("client share: %+v", cl)
 	}
 	if got[0].Engines["video"] != 100 || got[0].Usage != 100 || got[0].UsageSrc != "clients" {
 		t.Errorf("device from clients: %+v", got[0])
 	}
 	if got[0].MemUsed != cl.Mem {
 		t.Errorf("resident memory should be the clients' sum: %d vs %d", got[0].MemUsed, cl.Mem)
 	}
 	// the Processes table joins the client to its row
 	c.ring = []Sample{{T: 1, GPUs: got}}
 	c.prevProc = nil
 	table := c.Processes()
 	if len(table.GPUs) != 1 {
 		t.Errorf("table carries the GPUs: %+v", table.GPUs)
 	}
 }
 func TestFlattenGPU(t *testing.T) {
 	s := Sample{T: 5, GPUs: []GPU{{Name: "Intel Arc A380", Driver: "i915", Bus: "0000:03:00.0", Usage: 12, UsageSrc: "idle", Engines: map[string]float64{"video": 30}, PowerW: 9.5, PowerCapW: 55, MHz: 2450, MaxMHz: 2450, Clients: []GPUClient{{PID: 1}}}}}
 	got := map[string]any{}
 	for _, x := range flatten(s, nil) {
 		got[x.ID] = x.Value
 	}
 	for id, want := range map[string]any{"gpu.intel_arc_a380.usage": 12.0, "gpu.intel_arc_a380.video": 30.0, "gpu.intel_arc_a380.power": 9.5, "gpu.intel_arc_a380.freq": 2450.0, "gpu.intel_arc_a380.clients": 1.0} {
 		if got[id] != want {
 			t.Errorf("%s = %v, want %v", id, got[id], want)
 		}
 	}
 	if _, ok := got["gpu.intel_arc_a380.temp"]; ok {
 		t.Error("no temperature sensor, no temperature")
 	}
 }
 func replace(s, old, new string) string {
 	for i := 0; i+len(old) <= len(s); i++ {
 		if s[i:i+len(old)] == old {
 			return s[:i] + new + s[i+len(old):]
 		}
 	}
 	return s
 }
 // TestLiveGPU reads this machine's GPUs for real: HAMLAB_LIVE=1 go test -run TestLiveGPU -v ./internal/metrics
 func TestLiveGPU(t *testing.T) {
 	if os.Getenv("HAMLAB_LIVE") == "" {
 		t.Skip("set HAMLAB_LIVE=1 to read the real hardware")
 	}
 	c := NewCollector()
 	c.sample()
 	time.Sleep(1100 * time.Millisecond)
 	c.sample()
 	s, _ := c.Latest()
 	b, _ := json.MarshalIndent(s.GPUs, "", "  ")
 	t.Logf("%s", b)
 	for _, x := range flatten(s, func(id string) bool { return len(id) > 4 && id[:4] == "gpu." }) {
 		t.Logf("%s = %v %s", x.ID, x.Value, x.Unit)
 	}
 }
 // TestLiveARC: HAMLAB_LIVE=1 go test -run TestLiveARC -v ./internal/metrics
 func TestLiveARC(t *testing.T) {
 	if os.Getenv("HAMLAB_LIVE") == "" {
 		t.Skip("set HAMLAB_LIVE=1 to read the real machine")
 	}
 	c := NewCollector()
 	c.sample()
 	time.Sleep(1100 * time.Millisecond)
 	c.sample()
 	s, _ := c.Latest()
 	b, _ := json.MarshalIndent(s.ARC, "", "  ")
 	t.Logf("%s", b)
 	for _, x := range flatten(s, func(id string) bool { return strings.HasPrefix(id, "zfs.arc.") }) {
 		t.Logf("%s = %v %s", x.ID, x.Value, x.Unit)
 	}
 }
 // The embedded table names the lab's four GPUs the way lspci would, so a machine without
 // pciutils (the router) agrees with one that has it.
 func TestPCIName(t *testing.T) {
 	for id, want := range map[string]string{"8086:3185": "Intel UHD Graphics 600", "8086:56a5": "Intel Arc A380", "8086:46d1": "Intel UHD Graphics", "8086:4c8a": "Intel UHD Graphics 750", "10de:2684": "NVIDIA GeForce RTX 4090"} {
 		v, d := pciName(id[:4], id[5:])
 		if d == "" {
 			t.Errorf("%s: not in the table", id)
 			continue
 		}
 		if got := lspciName(v + " " + d); got != want {
 			t.Errorf("%s → %q, want %q", id, got, want)
 		}
 	}
 	if _, d := pciName("8086", "ffff"); d != "" {
 		t.Errorf("an unknown device should be empty, got %q", d)
 	}
 }
@@ -122,9 +122,10 @@ type Proc struct {
 	Name    string  `json:"name"`
 	Cmd     string  `json:"cmd,omitempty"`
 	User    string  `json:"user"`
-	CPU     float64 `json:"cpu"`      // percent of one core
+	CPU     float64 `json:"cpu"`           // percent of one core
-	CPUTime float64 `json:"cpu_time"` // seconds, lifetime
+	GPU     float64 `json:"gpu,omitempty"` // its busiest GPU engine, percent; summed over the GPUs it has open
-	RSS     uint64  `json:"rss"`      // bytes
+	CPUTime float64 `json:"cpu_time"`      // seconds, lifetime
 	RSS     uint64  `json:"rss"`           // bytes
 	Threads int     `json:"threads"`
 	State   string  `json:"state"`
 	Nice    int     `json:"nice"`
@@ -145,6 +146,8 @@ type Sample struct {
 	Disks    []Disk     `json:"disks"`
 	FS       []FS       `json:"fs"`
 	UPS      *UPS       `json:"ups"`
 	GPUs     []GPU      `json:"gpus"`
 	ARC      *ARC       `json:"arc"` // the ZFS read cache; nil without the module
 	Uptime   float64    `json:"uptime"`
 	fsT, upsT int64 // when the slow sections were really read; sensors report these, not T
@@ -173,6 +176,8 @@ type Collector struct {
 	procAt    time.Time
 	lastTable *ProcTable
 	upsProbe  func() *UPS // nil when no UPS daemon is installed
 	gpu       gpuState
 	prevARC   arcPrev
 	subMu sync.Mutex
 	subs  map[chan struct{}]struct{} // told after every sample; see Subscribe
@@ -356,6 +361,8 @@ func (c *Collector) sample() {
 		c.lastUPS, c.upsAt = c.upsProbe(), now
 	}
 	s.UPS, s.upsT = c.lastUPS, c.upsAt.UnixMilli()
 	s.GPUs = c.gpus(now, cpuTemp(s.Temps))
 	s.ARC = c.arc(now)
 	if first {
 		return // deltas are meaningless on the priming pass
@@ -1112,6 +1119,7 @@ type ProcTable struct {
 	Mem       Mem        `json:"mem"`
 	Swap      Mem        `json:"swap"`
 	Net       []Net      `json:"net"`
 	GPUs      []GPU      `json:"gpus"`
 	Uptime    float64    `json:"uptime"`
 	Procs     []Proc     `json:"procs"`
 	Total     int        `json:"total"`
@@ -1274,6 +1282,20 @@ func (c *Collector) Processes() ProcTable {
 		}
 	}
 	c.prevProc, c.procAt = cur, now
 	// what each process is doing on the GPUs, from the newest sample's client lists
 	gpuBy := map[int]float64{}
 	gpus := []GPU{}
 	if n := len(c.ring); n > 0 && c.ring[n-1].GPUs != nil {
 		gpus = c.ring[n-1].GPUs
 		for _, g := range gpus {
 			for _, cl := range g.Clients {
 				gpuBy[cl.PID] += cl.Usage
 			}
 		}
 	}
 	for i := range procs {
 		procs[i].GPU = round1(gpuBy[procs[i].PID])
 	}
 	sort.Slice(procs, func(i, j int) bool {
 		if procs[i].CPU != procs[j].CPU {
 			return procs[i].CPU > procs[j].CPU
@@ -1281,7 +1303,7 @@ func (c *Collector) Processes() ProcTable {
 		return procs[i].RSS > procs[j].RSS
 	})
 	t := ProcTable{T: now.UnixMilli(), Procs: procs, Total: len(procs), Uptime: uptime(), Load: loadavg(),
-		CPUModel: cpuModel, CPUMHz: round1(cpuMHz()), CPUMaxMHz: cpuMaxMHz, CPUHist: []float64{}, Net: []Net{}}
+		CPUModel: cpuModel, CPUMHz: round1(cpuMHz()), CPUMaxMHz: cpuMaxMHz, CPUHist: []float64{}, Net: []Net{}, GPUs: gpus}
 	for _, p := range procs {
 		if p.State == "R" {
 			t.Running++
@@ -1330,12 +1352,32 @@ func Signal(pid int, sig syscall.Signal) error {
 // ---- small helpers ----
 // readStr reads a /proc or /sys file whole, with plain blocking syscalls and never through
 // the runtime poller. os.ReadFile would hand a sysfs file to the poller — kernfs accepts
 // epoll — and a driver that answers a non-blocking read with EAGAIN (i915's fan attribute
 // did, on grid, 2026-10-02) then leaves the reader waiting for a readiness kernfs only
 // signals on sysfs_notify: never. The sampler sat in that wait for 42 minutes.
 func readStr(p string) string {
-	b, err := os.ReadFile(p)
+	fd, err := syscall.Open(p, syscall.O_RDONLY|syscall.O_CLOEXEC, 0)
 	if err != nil {
 		return ""
 	}
-	return string(b)
+	defer syscall.Close(fd)
 	var out []byte
 	buf := make([]byte, 4096)
 	for {
 		n, err := syscall.Read(fd, buf)
 		if n > 0 {
 			out = append(out, buf[:n]...)
 		}
 		if err == syscall.EINTR {
 			continue
 		}
 		if err != nil || n == 0 {
 			break
 		}
 	}
 	return string(out)
 }
 func readInt(p string) (int64, bool) {
@@ -0,0 +1,47 @@
 package metrics
 // A trimmed PCI ID database, embedded, so a GPU is named the same on every machine —
 // "Intel UHD Graphics 600" — whether or not that machine has pciutils. OpenWrt has neither
 // lspci nor pci.ids, and the router's GPU came up as "Intel GPU 3185" beside its twin's
 // "Intel UHD Graphics 600" until this. scripts/pci-ids-gpu.sh regenerates the file.
 import (
 	_ "embed"
 	"strings"
 	"sync"
 )
 //go:embed pciids_gpu.txt
 var pciIDsGPU string
 var pciTable struct {
 	once    sync.Once
 	vendors map[string]string // "8086" → "Intel Corporation"
 	devices map[string]string // "8086:3185" → "GeminiLake [UHD Graphics 600]"
 }
 // pciName looks a vendor and device ID up (hex, lower case, no 0x): the vendor's name and
 // the device's; empty where unknown.
 func pciName(vendor, device string) (string, string) {
 	pciTable.once.Do(func() {
 		pciTable.vendors, pciTable.devices = map[string]string{}, map[string]string{}
 		cur := ""
 		for _, line := range strings.Split(pciIDsGPU, "\n") {
 			switch {
 			case line == "" || line[0] == '#':
 			case line[0] == '\t':
 				if cur == "" || len(line) < 7 || line[1] == '\t' {
 					continue // a subsystem, or a device before any vendor
 				}
 				pciTable.devices[cur+":"+line[1:5]] = strings.TrimSpace(line[5:])
 			default:
 				if len(line) >= 6 {
 					cur = line[:4]
 					pciTable.vendors[cur] = strings.TrimSpace(line[4:])
 				}
 			}
 		}
 	})
 	vendor, device = strings.ToLower(vendor), strings.ToLower(device)
 	return pciTable.vendors[vendor], pciTable.devices[vendor+":"+device]
 }
@@ -131,6 +131,75 @@ func flatten(s Sample, want func(id string) bool) []sensors.Sensor {
 		f.add(p+".used_pct", x.Mount+" used", "usage", "%", ratio(x.Used, x.Size), s.fsT, pct, lbl...)
 	}
 	for _, g := range s.GPUs {
 		p := "gpu." + sensors.Seg(g.Name)
 		lbl := []string{"name", g.Name, "driver", g.Driver}
 		if g.Bus != "" {
 			lbl = append(lbl, "bus", g.Bus)
 		}
 		if g.Integrated {
 			lbl = append(lbl, "integrated", "1")
 		}
 		if g.UsageSrc != "" {
 			f.add(p+".usage", g.Name, "usage", "%", g.Usage, t, pct, lbl...)
 		}
 		for _, e := range []string{"render", "video", "copy", "video-enhance", "compute"} {
 			if v, ok := g.Engines[e]; ok {
 				f.add(p+"."+sensors.Seg(e), g.Name+" "+e, "usage", "%", v, t, pct, lbl...)
 			}
 		}
 		if g.Temp > 0 {
 			f.add(p+".temp", g.Name+" temperature", "temperature", "°C", g.Temp, t, none, append(lbl, "source", g.TempSrc)...)
 		}
 		if g.PowerW > 0 || g.PowerCapW > 0 {
 			b := none
 			if g.PowerCapW > 0 {
 				b = upTo(g.PowerCapW)
 			}
 			f.add(p+".power", g.Name+" power", "power", "W", g.PowerW, t, b, lbl...)
 		}
 		if g.FanRPM > 0 {
 			f.add(p+".fan", g.Name+" fan", "fan", "rpm", float64(g.FanRPM), t, none, lbl...)
 		}
 		if g.MaxMHz > 0 {
 			f.add(p+".freq", g.Name+" clock", "frequency", "MHz", g.MHz, t, upTo(g.MaxMHz), lbl...)
 		}
 		if g.MemTotal > 0 {
 			f.add(p+".mem_used", g.Name+" memory used", "bytes", "B", float64(g.MemUsed), t, upTo(float64(g.MemTotal)), lbl...)
 			f.add(p+".mem_total", g.Name+" memory", "bytes", "B", float64(g.MemTotal), t, none, lbl...)
 		} else if g.MemUsed > 0 {
 			f.add(p+".mem_used", g.Name+" memory used", "bytes", "B", float64(g.MemUsed), t, none, lbl...)
 		}
 		f.add(p+".clients", g.Name+" clients", "count", "", float64(len(g.Clients)), t, none, lbl...)
 	}
 	if a := s.ARC; a != nil {
 		p := "zfs.arc."
 		cap := upTo(float64(a.Max))
 		f.add(p+"size", "ARC size", "bytes", "B", float64(a.Size), t, cap)
 		f.add(p+"target", "ARC target", "bytes", "B", float64(a.Target), t, cap)
 		f.add(p+"min", "ARC floor", "bytes", "B", float64(a.Min), t, none)
 		f.add(p+"max", "ARC ceiling", "bytes", "B", float64(a.Max), t, none)
 		f.add(p+"mru", "ARC recently used", "bytes", "B", float64(a.MRU), t, cap)
 		f.add(p+"mfu", "ARC frequently used", "bytes", "B", float64(a.MFU), t, cap)
 		f.add(p+"metadata", "ARC metadata", "bytes", "B", float64(a.Metadata), t, cap)
 		f.add(p+"data", "ARC data", "bytes", "B", float64(a.Data), t, cap)
 		if a.HitPctKnown {
 			f.add(p+"hit_pct", "ARC hit rate", "usage", "%", a.HitPct, t, pct)
 		}
 		f.add(p+"reads", "ARC reads", "rate", "/s", a.ReadsPerSec, t, none)
 		f.add(p+"misses", "ARC misses", "rate", "/s", a.MissesPerSec, t, none)
 		f.add(p+"lifetime_hit_pct", "ARC hit rate since boot", "usage", "%", a.LifetimeHitPct, t, pct)
 		f.add(p+"no_grow", "ARC held by memory pressure", "state", "", a.NoGrow, t, none)
 		if a.Compressed > 0 {
 			f.add(p+"compress_ratio", "ARC compression", "ratio", "x", round1(float64(a.Uncompressed)/float64(a.Compressed)*10)/10, t, none)
 		}
 		if a.L2Size > 0 {
 			f.add(p+"l2_size", "L2ARC size", "bytes", "B", float64(a.L2Size), t, none)
 			f.add(p+"l2_hit_pct", "L2ARC hit rate", "usage", "%", a.L2HitPct, t, pct)
 		}
 	}
 	if u := s.UPS; u != nil {
 		ut := s.upsT
 		f.add("ups.battery", "UPS battery", "usage", "%", u.BatteryPct, ut, pct, "model", u.Model)
@@ -0,0 +1,360 @@
 // Package network is what the Network tab shows: the machine's interfaces as the kernel has
 // them, how the internet looks from out there, and a speed test that measures the line.
 //
 // Everything here is read from /proc and /sys, or measured against Cloudflare's public speed
 // endpoints — no tool to install, on any of the machines. That matters for the router, where
 // flash is small and `opkg install` is a decision.
 package network
 import (
 	"bufio"
 	"encoding/binary"
 	"fmt"
 	"net"
 	"os"
 	"sort"
 	"strconv"
 	"strings"
 	"sync"
 	"time"
 )
 // Addr is one address on an interface.
 type Addr struct {
 	IP     string `json:"ip"`
 	Prefix int    `json:"prefix"` // network bits, as in /24
 	V6     bool   `json:"v6,omitempty"`
 }
 // Iface is one network interface: what it is, whether it is up, and what it has moved.
 type Iface struct {
 	Name string `json:"name"`
 	Kind string `json:"kind"` // ethernet | wifi | bridge | tunnel | loopback | virtual
 	MAC  string `json:"mac,omitempty"`
 	MTU  int    `json:"mtu"`
 	// Up is the administrative state, Carrier whether a cable (or an association) is actually
 	// there. Up without Carrier is the usual "it's plugged into nothing".
 	Up      bool `json:"up"`
 	Carrier bool `json:"carrier"`
 	// SpeedMbps is what the link negotiated, where the driver says: 0 for wireless, virtual
 	// and tunnel devices, which have no such number.
 	SpeedMbps int      `json:"speed_mbps,omitempty"`
 	Duplex    string   `json:"duplex,omitempty"`
 	Driver    string   `json:"driver,omitempty"`
 	SignalDbm *float64 `json:"signal_dbm,omitempty"` // wireless only, from /proc/net/wireless
 	Addrs     []Addr   `json:"addrs"`
 	RxTotal   uint64   `json:"rx_total"`
 	TxTotal   uint64   `json:"tx_total"`
 	// Default marks the interface the default route leaves by — the one that matters when
 	// the question is about the internet.
 	Default bool `json:"default,omitempty"`
 }
 // Report is the machine's network as it stands.
 type Report struct {
 	At         time.Time `json:"at"`
 	Host       string    `json:"host,omitempty"`
 	Ifaces     []Iface   `json:"ifaces"`
 	Gateway    string    `json:"gateway,omitempty"`
 	GatewayVia string    `json:"gateway_via,omitempty"` // the interface it is reached through
 	DNS        []string  `json:"dns,omitempty"`
 	Search     []string  `json:"search,omitempty"`
 }
 func readStr(path string) string {
 	b, err := os.ReadFile(path)
 	if err != nil {
 		return ""
 	}
 	return strings.TrimSpace(string(b))
 }
 func readInt(path string) int {
 	n, err := strconv.Atoi(readStr(path))
 	if err != nil {
 		return 0
 	}
 	return n
 }
 // kindOf names what sort of device this is, from what the kernel exposes about it rather
 // than from its name — a bridge is a bridge whatever it is called.
 func kindOf(name string, flags net.Flags) string {
 	sys := "/sys/class/net/" + name + "/"
 	switch {
 	case flags&net.FlagLoopback != 0:
 		return "loopback"
 	case dirExists(sys + "wireless"), dirExists(sys + "phy80211"):
 		return "wifi"
 	case dirExists(sys + "bridge"):
 		return "bridge"
 	case dirExists(sys + "bonding"):
 		return "bond"
 	}
 	// DEVTYPE says vlan, bridge, wlan… for the devices that have one
 	for _, line := range strings.Split(readStr(sys+"uevent"), "\n") {
 		if v, ok := strings.CutPrefix(line, "DEVTYPE="); ok && v != "" {
 			return v
 		}
 	}
 	if !dirExists(sys + "device") { // no bus device behind it: veth, tun, wireguard, docker's own
 		return "virtual"
 	}
 	return "ethernet"
 }
 func dirExists(p string) bool {
 	st, err := os.Stat(p)
 	return err == nil && st.IsDir()
 }
 // driverOf is the module behind an interface, which is what names the hardware in one word.
 func driverOf(name string) string {
 	l, err := os.Readlink("/sys/class/net/" + name + "/device/driver")
 	if err != nil {
 		return ""
 	}
 	return l[strings.LastIndexByte(l, '/')+1:]
 }
 // wirelessSignal is the signal level of each wireless interface, in dBm.
 func wirelessSignal() map[string]float64 {
 	out := map[string]float64{}
 	f, err := os.Open("/proc/net/wireless")
 	if err != nil {
 		return out
 	}
 	defer f.Close()
 	sc := bufio.NewScanner(f)
 	for sc.Scan() {
 		name, rest, ok := strings.Cut(sc.Text(), ":")
 		if !ok {
 			continue
 		}
 		fs := strings.Fields(rest)
 		if len(fs) < 3 {
 			continue
 		}
 		if v, err := strconv.ParseFloat(strings.TrimSuffix(fs[2], "."), 64); err == nil {
 			out[strings.TrimSpace(name)] = v
 		}
 	}
 	return out
 }
 // counters is every interface's byte totals, from the same file the sampler reads.
 func counters() map[string][2]uint64 {
 	out := map[string][2]uint64{}
 	f, err := os.Open("/proc/net/dev")
 	if err != nil {
 		return out
 	}
 	defer f.Close()
 	sc := bufio.NewScanner(f)
 	for sc.Scan() {
 		name, rest, ok := strings.Cut(sc.Text(), ":")
 		if !ok {
 			continue
 		}
 		fs := strings.Fields(rest)
 		if len(fs) < 16 {
 			continue
 		}
 		rx, _ := strconv.ParseUint(fs[0], 10, 64)
 		tx, _ := strconv.ParseUint(fs[8], 10, 64)
 		out[strings.TrimSpace(name)] = [2]uint64{rx, tx}
 	}
 	return out
 }
 // defaultRoute is the gateway the machine sends the internet to, and the interface it
 // leaves by. Read from /proc/net/route, whose addresses are little-endian hex.
 func defaultRoute() (gw, via string) {
 	f, err := os.Open("/proc/net/route")
 	if err != nil {
 		return "", ""
 	}
 	defer f.Close()
 	sc := bufio.NewScanner(f)
 	best := -1
 	for sc.Scan() {
 		fs := strings.Fields(sc.Text())
 		if len(fs) < 8 || fs[0] == "Iface" {
 			continue
 		}
 		if fs[1] != "00000000" { // not a default route
 			continue
 		}
 		metric, _ := strconv.Atoi(fs[6])
 		if best >= 0 && metric >= best {
 			continue
 		}
 		ip, err := hexIP(fs[2])
 		if err != nil {
 			continue
 		}
 		best, gw, via = metric, ip, fs[0]
 	}
 	return gw, via
 }
 // hexIP reads one address out of /proc/net/route, where each is printed as a 32-bit word in
 // the host's byte order: 10.0.0.1 appears as "0100000A" on a little-endian machine. So the
 // text is parsed as a plain number, then written back out in native order to recover the four
 // bytes — which is right on either sort of machine, without asking which this is.
 func hexIP(h string) (string, error) {
 	v, err := strconv.ParseUint(h, 16, 32)
 	if err != nil || len(h) != 8 {
 		return "", fmt.Errorf("not an address: %q", h)
 	}
 	var b [4]byte
 	binary.NativeEndian.PutUint32(b[:], uint32(v))
 	return net.IPv4(b[0], b[1], b[2], b[3]).String(), nil
 }
 // resolvers is what /etc/resolv.conf points at.
 func resolvers() (servers, search []string) {
 	f, err := os.Open("/etc/resolv.conf")
 	if err != nil {
 		return nil, nil
 	}
 	defer f.Close()
 	sc := bufio.NewScanner(f)
 	for sc.Scan() {
 		fs := strings.Fields(sc.Text())
 		if len(fs) < 2 || strings.HasPrefix(fs[0], "#") {
 			continue
 		}
 		switch fs[0] {
 		case "nameserver":
 			servers = append(servers, fs[1])
 		case "search", "domain":
 			search = append(search, fs[1:]...)
 		}
 	}
 	return servers, search
 }
 // Interfaces is every interface the kernel has, with its addresses and totals.
 func Interfaces() []Iface {
 	list, err := net.Interfaces()
 	if err != nil {
 		return nil
 	}
 	_, via := defaultRoute()
 	sig, cnt := wirelessSignal(), counters()
 	out := make([]Iface, 0, len(list))
 	for _, in := range list {
 		sys := "/sys/class/net/" + in.Name + "/"
 		f := Iface{
 			Name: in.Name, Kind: kindOf(in.Name, in.Flags), MTU: in.MTU,
 			MAC: in.HardwareAddr.String(), Up: in.Flags&net.FlagUp != 0,
 			Driver: driverOf(in.Name), Default: in.Name == via, Addrs: []Addr{},
 		}
 		f.Carrier = readStr(sys+"carrier") == "1"
 		if f.Kind == "loopback" {
 			f.Carrier = f.Up
 		}
 		if sp := readInt(sys + "speed"); sp > 0 { // -1 when the driver doesn't know
 			f.SpeedMbps = sp
 		}
 		if d := readStr(sys + "duplex"); d != "" && d != "unknown" {
 			f.Duplex = d
 		}
 		if v, ok := sig[in.Name]; ok {
 			f.SignalDbm = &v
 		}
 		if c, ok := cnt[in.Name]; ok {
 			f.RxTotal, f.TxTotal = c[0], c[1]
 		}
 		addrs, _ := in.Addrs()
 		for _, a := range addrs {
 			n, ok := a.(*net.IPNet)
 			if !ok {
 				continue
 			}
 			ones, _ := n.Mask.Size()
 			f.Addrs = append(f.Addrs, Addr{IP: n.IP.String(), Prefix: ones, V6: n.IP.To4() == nil})
 		}
 		out = append(out, f)
 	}
 	// the interface the internet leaves by first, then the ones carrying addresses, then the rest
 	sort.SliceStable(out, func(i, j int) bool {
 		a, b := out[i], out[j]
 		if a.Default != b.Default {
 			return a.Default
 		}
 		ai, bi := len(a.Addrs) > 0 && a.Kind != "loopback", len(b.Addrs) > 0 && b.Kind != "loopback"
 		if ai != bi {
 			return ai
 		}
 		return a.Name < b.Name
 	})
 	return out
 }
 // Look is the machine's network now. Cheap: a handful of small files.
 func Look(host string) Report {
 	gw, via := defaultRoute()
 	dns, search := resolvers()
 	return Report{At: time.Now(), Host: host, Ifaces: Interfaces(), Gateway: gw, GatewayVia: via, DNS: dns, Search: search}
 }
 // ---- how the internet sees this machine ----
 // Outside is what the far end of the line says about this connection: the address the
 // internet sees, and where it came ashore. Refreshed on its own clock, because the public
 // address here is a DHCP lease and moves without warning.
 type Outside struct {
 	At      time.Time `json:"at"`
 	IP      string    `json:"ip,omitempty"`
 	Colo    string    `json:"colo,omitempty"` // the Cloudflare edge that answered — "SEA"
 	Country string    `json:"country,omitempty"`
 	Error   string    `json:"error,omitempty"`
 }
 // Watcher keeps the outside view fresh and holds the speed tester.
 type Watcher struct {
 	mu   sync.RWMutex
 	out  Outside
 	Test *Tester
 }
 const outsideEvery = 15 * time.Minute
 func NewWatcher(onResult func(Result)) *Watcher {
 	w := &Watcher{Test: NewTester(onResult)}
 	go func() {
 		for {
 			w.refresh()
 			time.Sleep(outsideEvery)
 		}
 	}()
 	return w
 }
 func (w *Watcher) refresh() {
 	o := probeOutside()
 	w.mu.Lock()
 	// a failed probe does not throw away a good answer; it is noted beside it
 	if o.IP == "" && w.out.IP != "" {
 		w.out.Error, w.out.At = o.Error, o.At
 	} else {
 		w.out = o
 	}
 	w.mu.Unlock()
 }
 // Outside is the last known view from outside, asking again if it has gone stale.
 func (w *Watcher) Outside() Outside {
 	w.mu.RLock()
 	o := w.out
 	w.mu.RUnlock()
 	if time.Since(o.At) > outsideEvery {
 		w.refresh()
 		w.mu.RLock()
 		o = w.out
 		w.mu.RUnlock()
 	}
 	return o
 }
@@ -0,0 +1,317 @@
 package network
 import (
 	"context"
 	"errors"
 	"io"
 	"math"
 	"net/http"
 	"net/http/httptest"
 	"os"
 	"sync/atomic"
 	"testing"
 	"time"
 )
 func TestHexIP(t *testing.T) {
 	// /proc/net/route prints each address as a host-order word: on a little-endian machine
 	// 10.0.0.1 is written "0100000A". These cases are that machine, which is all of ours.
 	for hex, want := range map[string]string{
 		"0100000A": "10.0.0.1",
 		"00000000": "0.0.0.0",
 		"0101A8C0": "192.168.1.1",
 	} {
 		got, err := hexIP(hex)
 		if err != nil || got != want {
 			t.Errorf("hexIP(%q) = %q, %v; want %q", hex, got, err, want)
 		}
 	}
 	if _, err := hexIP("nonsense"); err == nil {
 		t.Error("hexIP accepted nonsense")
 	}
 }
 func TestServerTimeIsNotLatency(t *testing.T) {
 	h := http.Header{}
 	h.Add("Server-Timing", "cfSpeedEdge;dur=7, cfSpeedWorker;dur=36")
 	h.Add("Server-Timing", `cfL4;desc="?proto=TCP&rtt=3505&min_rtt=2856"`)
 	if got := serverMs(h); got != 43 {
 		t.Errorf("serverMs = %v, want 43 (7+36, and nothing from the cfL4 description)", got)
 	}
 	if got := serverMs(http.Header{}); got != 0 {
 		t.Errorf("serverMs with no header = %v, want 0", got)
 	}
 	h2 := http.Header{"Cf-Ray": {"a41f6720f966a36c-SEA"}, "Cf-Meta-Ip": {"50.47.252.221"}}
 	if m := readMeta(h2); m.Colo != "SEA" || m.IP != "50.47.252.221" {
 		t.Errorf("readMeta = %+v, want the edge SEA and the address", m)
 	}
 }
 func TestRates(t *testing.T) {
 	// a gibibyte in eight seconds is 1073.7 Mbps
 	if got := mbps(1<<30, 8); math.Abs(got-1073.7) > 0.1 {
 		t.Errorf("mbps(1GiB, 8s) = %v, want ~1073.7", got)
 	}
 	if got := best([]float64{12.5, 9.2, 40, 9.9}); got != 9.2 {
 		t.Errorf("best = %v, want 9.2 (the least of them)", got)
 	}
 	if got := jitter([]float64{10, 10, 10}); got != 0 {
 		t.Errorf("jitter of a steady line = %v, want 0", got)
 	}
 	if got := jitter([]float64{10}); got != 0 {
 		t.Errorf("jitter of one sample = %v, want 0", got)
 	}
 	if got := jitter([]float64{5, 50, 5, 50}); got != 45 {
 		t.Errorf("jitter of a line alternating by 45 = %v, want 45", got)
 	}
 	// one slow probe among steady ones moves it a little, not enormously
 	if got := jitter([]float64{10, 10, 10, 60, 10, 10}); got > 25 {
 		t.Errorf("jitter with a single outlier = %v, want it not to run away", got)
 	}
 }
 // The body an upload sends is a fixed size, counts itself, and costs nothing to make.
 func TestZerosBodyCountsItself(t *testing.T) {
 	var n atomic.Int64
 	z := &zeros{left: 100, n: &n}
 	buf := make([]byte, 30)
 	total := 0
 	for {
 		got, err := z.Read(buf)
 		total += got
 		if err != nil {
 			break
 		}
 	}
 	if total != 100 || n.Load() != 100 {
 		t.Errorf("read %d bytes and counted %d, want 100 and 100", total, n.Load())
 	}
 }
 // Interfaces has to work on whatever machine this runs on, and always name a loopback.
 func TestInterfacesHasLoopback(t *testing.T) {
 	ifs := Interfaces()
 	if len(ifs) == 0 {
 		t.Fatal("no interfaces at all")
 	}
 	lo := false
 	for _, f := range ifs {
 		if f.Kind == "loopback" {
 			lo = true
 			if !f.Up || len(f.Addrs) == 0 {
 				t.Errorf("loopback %s looks wrong: up=%v addrs=%v", f.Name, f.Up, f.Addrs)
 			}
 		}
 		if f.Name == "" {
 			t.Error("an interface with no name")
 		}
 	}
 	if !lo {
 		t.Error("no loopback among the interfaces")
 	}
 }
 // The real thing, against the real endpoint: skipped by default, because `make test` must
 // not need the internet. HAMLAB_NET_TEST=1 go test ./internal/network/ -run Live -v
 func TestLiveSpeedTest(t *testing.T) {
 	if os.Getenv("HAMLAB_NET_TEST") == "" {
 		t.Skip("set HAMLAB_NET_TEST=1 to measure the actual line")
 	}
 	tr := NewTester(nil)
 	if _, err := tr.Start(); err != nil {
 		t.Fatal(err)
 	}
 	if _, err := tr.Start(); err == nil {
 		t.Error("a second test started while the first was running")
 	}
 	deadline := time.Now().Add(90 * time.Second)
 	for tr.Running() && time.Now().Before(deadline) {
 		time.Sleep(250 * time.Millisecond)
 	}
 	st := tr.State()
 	if st.Result == nil {
 		t.Fatal("no result")
 	}
 	r := *st.Result
 	t.Logf("down %s  up %s  latency %.1f ms  jitter %.1f ms  via %s  ip %s  %.0f MB moved in %.1fs",
 		Mbits(r.DownMbps), Mbits(r.UpMbps), r.LatencyMs, r.JitterMs, r.Colo, r.IP,
 		float64(r.DownBytes+r.UpBytes)/1e6, r.TookSec)
 	if r.Error != "" {
 		t.Fatalf("test failed: %s", r.Error)
 	}
 	if r.DownMbps <= 0 || r.UpMbps <= 0 || r.LatencyMs <= 0 {
 		t.Errorf("a figure came back zero: %+v", r)
 	}
 	if r.IP == "" {
 		t.Error("no public address reported")
 	}
 	if len(tr.History()) != 1 {
 		t.Errorf("history has %d results, want 1", len(tr.History()))
 	}
 }
 // A phase measured against a local server, at a rate the test controls. This is the part of
 // the speed test with arithmetic in it — the window after the ramp, the byte cap that ends it
 // early, and the rate that comes out — and none of it needs the internet to check.
 func TestPhaseMeasuresTheWindow(t *testing.T) {
 	const perTick = 1 << 20 // a mebibyte every 5 ms ≈ 1678 Mbps, whatever the machine is doing
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		if r.URL.Path == "/up" {
 			_, _ = io.Copy(io.Discard, r.Body)
 			w.WriteHeader(204)
 			return
 		}
 		buf := make([]byte, perTick)
 		fl, _ := w.(http.Flusher)
 		for {
 			if _, err := w.Write(buf); err != nil {
 				return
 			}
 			if fl != nil {
 				fl.Flush()
 			}
 			select {
 			case <-time.After(5 * time.Millisecond):
 			case <-r.Context().Done():
 				return
 			}
 		}
 	}))
 	defer srv.Close()
 	defer func(up string, w, d, mw time.Duration, m int64) {
 		upURL, warmUp, downFor, minWindow, measureBytes = up, w, d, mw, m
 	}(upURL, warmUp, downFor, minWindow, measureBytes)
 	upURL = srv.URL + "/up"
 	warmUp, downFor, minWindow, measureBytes = 100*time.Millisecond, 3*time.Second, 200*time.Millisecond, 8<<20
 	here := Source{"local", func(int) string { return srv.URL + "/down" }}
 	tr := NewTester(nil)
 	began := time.Now()
 	mbps, moved, note, err := tr.phase(context.Background(), srv.Client(), "download", 2, downFor, 0, 1, here)
 	took := time.Since(began)
 	if err != nil {
 		t.Fatalf("phase: %v", err)
 	}
 	if note != "" {
 		t.Errorf("note = %q, want none — nothing went wrong", note)
 	}
 	// The cap must have ended it early rather than the clock: the phase was allowed three
 	// seconds and should be done in well under one, having moved at least the cap.
 	if moved < measureBytes {
 		t.Errorf("moved %d bytes, want at least the %d-byte cap", moved, measureBytes)
 	}
 	if took > time.Second {
 		t.Errorf("phase took %v of the %v it was allowed; the byte cap should have ended it", took.Round(time.Millisecond), downFor)
 	}
 	// The server hands out a mebibyte per 5 ms per stream, so the rate is in the hundreds of
 	// Mbps. The point is that it is a real quotient of what moved over how long, not that it
 	// hits a particular figure on a loaded test machine.
 	if mbps < 50 || mbps > 20000 {
 		t.Errorf("mbps = %v, which is not a believable quotient", mbps)
 	}
 	t.Logf("measured %.0f Mbps over %.1f MiB", mbps, float64(moved)/(1<<20))
 	// A phase whose server refuses outright fails, and says what it was told.
 	bad := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) { w.WriteHeader(429) }))
 	defer bad.Close()
 	refuser := Source{"refuser", func(int) string { return bad.URL + "/down" }}
 	_, _, _, err = tr.phase(context.Background(), bad.Client(), "download", 2, 300*time.Millisecond, 0, 1, refuser)
 	if err == nil {
 		t.Error("a phase against a server that refuses everything should fail")
 	} else if !refused(err) {
 		t.Errorf("error = %q, want the refusal that makes the fallback kick in", err)
 	}
 }
 // Stopping a test mid-flight is reported as cancelled, not as whatever the streams said on
 // their way out.
 func TestCancelIsReportedAsCancelled(t *testing.T) {
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		buf := make([]byte, 1<<16)
 		for {
 			if _, err := w.Write(buf); err != nil {
 				return
 			}
 			select {
 			case <-time.After(time.Millisecond):
 			case <-r.Context().Done():
 				return
 			}
 		}
 	}))
 	defer srv.Close()
 	defer func(w, d time.Duration) { warmUp, downFor = w, d }(warmUp, downFor)
 	warmUp, downFor = 10*time.Millisecond, 10*time.Second
 	here := Source{"local", func(int) string { return srv.URL + "/down" }}
 	tr := NewTester(nil)
 	ctx, cancel := context.WithCancel(context.Background())
 	go func() { time.Sleep(200 * time.Millisecond); cancel() }()
 	_, _, _, err := tr.phase(ctx, srv.Client(), "download", 1, downFor, 0, 1, here)
 	// Cancelled after bytes moved: the figure stands, and the phase does not claim an error.
 	if err != nil && !errors.Is(err, context.Canceled) {
 		t.Errorf("err = %v, want nil or a cancellation", err)
 	}
 }
 // When the first source refuses, the download is taken from the next one and the result says
 // so. This is the whole reason there is more than one.
 func TestDownloadFallsBackWhenRefused(t *testing.T) {
 	serving := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		if r.URL.Path == "/up" {
 			_, _ = io.Copy(io.Discard, r.Body)
 			w.WriteHeader(204)
 			return
 		}
 		// paced, so the phase is governed by its window and not by a loopback moving a
 		// gibibyte before the first tick
 		buf := make([]byte, 1<<20)
 		fl, _ := w.(http.Flusher)
 		for i := 0; i < 64; i++ {
 			if _, err := w.Write(buf); err != nil {
 				return
 			}
 			if fl != nil {
 				fl.Flush()
 			}
 			select {
 			case <-time.After(2 * time.Millisecond):
 			case <-r.Context().Done():
 				return
 			}
 		}
 	}))
 	defer serving.Close()
 	refusing := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) { w.WriteHeader(429) }))
 	defer refusing.Close()
 	defer func(srcs []Source, up string, w, d, mw time.Duration, m int64) {
 		downSources, upURL, warmUp, downFor, minWindow, measureBytes = srcs, up, w, d, mw, m
 	}(downSources, upURL, warmUp, downFor, minWindow, measureBytes)
 	downSources = []Source{
 		{"refuser", func(int) string { return refusing.URL + "/down" }},
 		{"the-other-one", func(int) string { return serving.URL + "/down" }},
 	}
 	upURL = serving.URL + "/up"
 	warmUp, downFor, upFor, minWindow, measureBytes = 50*time.Millisecond, time.Second, 300*time.Millisecond, 100*time.Millisecond, 4<<20
 	tr := NewTester(nil)
 	if _, err := tr.Start(); err != nil {
 		t.Fatal(err)
 	}
 	deadline := time.Now().Add(30 * time.Second)
 	for tr.Running() && time.Now().Before(deadline) {
 		time.Sleep(50 * time.Millisecond)
 	}
 	res := tr.State().Result
 	if res == nil {
 		t.Fatal("no result")
 	}
 	if res.Source != "the-other-one" {
 		t.Errorf("source = %q, want the second one — the first refused", res.Source)
 	}
 	if res.DownMbps <= 0 {
 		t.Errorf("download = %v, want a figure from the source that did serve: %+v", res.DownMbps, res)
 	}
 }
@@ -0,0 +1,687 @@
 package network
 // The speed test. It measures against Cloudflare's public speed endpoints, the same ones
 // speed.cloudflare.com uses: a plain GET of N bytes down, a plain POST of N bytes up, and a
 // zero-byte GET for the round trip. No account, no API key, nothing to install, and an edge
 // close to every one of these machines.
 //
 // The method, since a number without one means little — it is stated in the UI too:
 //
 //   - Latency is the round trip of a zero-byte response on a connection that is already open,
 //     less the time the far end says it spent on the request (its Server-Timing header). That
 //     subtraction is the difference between measuring the line and measuring Cloudflare's
 //     thinking time: a zero-byte response is almost entirely the latter — tens of
 //     milliseconds of worker, against a few of actual wire. The figure reported is the best of
 //     several probes; jitter is the average change from one to the next.
 //   - Download and upload each run several streams at once for a fixed few seconds. One
 //     stream measures one TCP connection's congestion window as much as the line, which is
 //     why every speed test ever made uses more than one.
 //   - The first second of each phase is thrown away. That is the ramp: counting it reports a
 //     number nobody's connection ever actually did.
 //   - Each phase is capped in bytes as well as seconds, so a very fast line cannot run away
 //     with the meter.
 //
 // The one thing to know about the far end: it rate-limits by volume, and a test moves a few
 // hundred megabytes. Several in a row will start earning 429s, which is why there is a
 // cooldown between tests and why a phase that gets refused says so instead of reporting the
 // decayed figure it would otherwise compute.
 //
 // HTTP/2 is deliberately off: over one HTTP/2 connection the streams would share a single
 // congestion window, which measures the window, not the line.
 import (
 	"context"
 	"crypto/tls"
 	"errors"
 	"fmt"
 	"io"
 	"log"
 	"math"
 	"net"
 	"net/http"
 	"strconv"
 	"strings"
 	"sync"
 	"sync/atomic"
 	"time"
 )
 // The endpoints, and the shape of a phase. Variables rather than constants only so the tests
 // can point them at a local server and shrink the windows; nothing else rebinds them.
 // Source is somewhere to pull bytes from. Cloudflare's endpoint is the first choice because
 // it is anycast — near wherever Ham Lab is running — and because the same host answers for
 // upload and for the trace. The fallback exists because that endpoint rate-limits by volume
 // and will refuse a connection that has been tested hard: a plain 100 MB file on a public
 // mirror has no such opinion. It is in Seattle, so on a machine far from there it will read
 // low, which is exactly why every result names the source it used.
 type Source struct {
 	Name string
 	// URL takes the number of bytes wanted, where the source can be asked; a fixed file
 	// ignores it and is fetched again each time round.
 	URL func(bytes int) string
 }
 var downSources = []Source{
 	{"cloudflare", func(b int) string { return fmt.Sprintf(cfDownURL, b) }},
 	{"linode-seattle", func(int) string { return "https://speedtest.seattle.linode.com/100MB-seattle.bin" }},
 }
 var (
 	cfDownURL = "https://speed.cloudflare.com/__down?bytes=%d"
 	upURL     = "https://speed.cloudflare.com/__up"
 	// traceURL is Cloudflare's standard trace: a few lines of key=value naming the address it
 	// sees, the edge that answered and the country. A few hundred bytes.
 	traceURL = "https://speed.cloudflare.com/cdn-cgi/trace"
 	warmUp  = time.Second
 	downFor = 4 * time.Second
 	upFor   = 4 * time.Second
 	// measureBytes ends a phase early once this much has moved inside the window. A fast line
 	// says what it is in well under the time allowed, and stopping there is what keeps a test
 	// down to a couple of hundred megabytes — which is the difference between the far end
 	// serving this happily and rate-limiting it. A slow line never reaches it and is timed out
 	// by downFor/upFor instead, which is the honest way round.
 	measureBytes = int64(64) << 20
 	// minWindow is how long the window must last whatever the byte cap says. Without it a very
 	// fast line fills the cap in a few hundredths of a second and the phase would be over
 	// before there was anything to divide by.
 	minWindow = time.Second
 )
 const (
 	latencyProbes = 8
 	downStreams   = 4
 	upStreams     = 3
 	downChunk     = 25 << 20 // bytes per download request; the far end refuses much more
 	upChunk       = 25 << 20 // and per upload request
 	// phaseBudget is the runaway guard, counting the ramp too: nothing measures forever.
 	phaseBudget = int64(512) << 20
 	httpTimeout = 90 * time.Second
 	// Cooldown is how long the line is left alone between tests. The far end rate-limits by
 	// volume, and two back-to-back tests are the surest way to be refused by it.
 	Cooldown = 20 * time.Second
 	// Method is what the UI tells the reader about the numbers it is showing.
 	Method = "Cloudflare's speed endpoints · 4 streams down, 3 up · a 1 s ramp, then up to 4 s or 64 MiB each way"
 )
 // Result is one finished speed test.
 type Result struct {
 	At        time.Time `json:"at"`
 	DownMbps  float64   `json:"down_mbps"`
 	UpMbps    float64   `json:"up_mbps"`
 	LatencyMs float64   `json:"latency_ms"`
 	JitterMs  float64   `json:"jitter_ms"`
 	DownBytes int64     `json:"down_bytes"`
 	UpBytes   int64     `json:"up_bytes"`
 	TookSec   float64   `json:"took_sec"`
 	// Colo is the Cloudflare edge that answered — "SEA" — and IP is this connection's public
 	// address as that edge saw it.
 	Colo string `json:"colo,omitempty"`
 	IP   string `json:"ip,omitempty"`
 	// Source is where the download came from, which matters because the fallback is not in the
 	// same place as the anycast one. Empty means the usual.
 	Source string `json:"source,omitempty"`
 	// Note is set when the figures are real but something went wrong along the way — a stream
 	// that was refused part-way, most often. The number is what was actually seen; the note is
 	// why it may be low.
 	Note  string `json:"note,omitempty"`
 	Error string `json:"error,omitempty"`
 }
 // Run is a test in progress, or the last one to have finished.
 type Run struct {
 	Started time.Time `json:"started"`
 	Running bool      `json:"running"`
 	// Phase is where it has got to: latency, download, upload, done.
 	Phase string `json:"phase"`
 	// Done is how far through the whole test, 0–1, so one bar can show all of it.
 	Done float64 `json:"done"`
 	// Mbps is the live figure for the phase under way — what the needle shows.
 	Mbps   float64 `json:"mbps"`
 	Bytes  int64   `json:"bytes"`
 	Result *Result `json:"result,omitempty"`
 }
 // Tester runs one speed test at a time and remembers what the last ones said.
 type Tester struct {
 	mu   sync.RWMutex
 	run  Run
 	hist []Result
 	// lastTry is when a test was last *started*, which is what the cooldown counts from — a
 	// test that was refused still used the line, and hammering retries at an endpoint that has
 	// just rate-limited us is the one thing guaranteed not to help.
 	lastTry time.Time
 	cancel  context.CancelFunc
 	onDone  func(Result)
 	maxHist int
 }
 func NewTester(onDone func(Result)) *Tester {
 	return &Tester{onDone: onDone, maxHist: 50}
 }
 // Load takes the results kept from before a restart, oldest first.
 func (t *Tester) Load(rs []Result) {
 	t.mu.Lock()
 	defer t.mu.Unlock()
 	t.hist = append(t.hist[:0], rs...)
 	if n := len(t.hist); n > 0 {
 		last := t.hist[n-1]
 		t.run = Run{Started: last.At, Phase: "done", Done: 1, Result: &last}
 		t.lastTry = last.At
 	}
 }
 func (t *Tester) State() Run {
 	t.mu.RLock()
 	defer t.mu.RUnlock()
 	return t.run
 }
 func (t *Tester) History() []Result {
 	t.mu.RLock()
 	defer t.mu.RUnlock()
 	return append([]Result{}, t.hist...)
 }
 func (t *Tester) Running() bool {
 	t.mu.RLock()
 	defer t.mu.RUnlock()
 	return t.run.Running
 }
 // Cancel stops a test in progress. Reports whether there was one.
 func (t *Tester) Cancel() bool {
 	t.mu.Lock()
 	c := t.cancel
 	t.mu.Unlock()
 	if c == nil {
 		return false
 	}
 	c()
 	return true
 }
 var ErrBusy = errors.New("a speed test is already running")
 // Start begins a test in the background and returns its opening state.
 func (t *Tester) Start() (Run, error) {
 	t.mu.Lock()
 	if t.run.Running {
 		r := t.run
 		t.mu.Unlock()
 		return r, ErrBusy
 	}
 	if since := time.Since(t.lastTry); since < Cooldown && !t.lastTry.IsZero() {
 		r := t.run
 		t.mu.Unlock()
 		return r, fmt.Errorf("give the line %d more seconds — testing back to back is what gets this connection rate-limited", int((Cooldown-since).Seconds())+1)
 	}
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	t.cancel = cancel
 	t.lastTry = time.Now()
 	t.run = Run{Started: t.lastTry, Running: true, Phase: "latency"}
 	r := t.run
 	t.mu.Unlock()
 	go t.measure(ctx, cancel)
 	return r, nil
 }
 func (t *Tester) set(fn func(*Run)) {
 	t.mu.Lock()
 	fn(&t.run)
 	t.mu.Unlock()
 }
 func (t *Tester) measure(ctx context.Context, cancel context.CancelFunc) {
 	defer cancel()
 	started := time.Now()
 	res := Result{At: started}
 	tr := newTransport(downStreams)
 	defer tr.CloseIdleConnections()
 	client := &http.Client{Transport: tr, Timeout: httpTimeout}
 	// latency, and the meta the same responses carry
 	lat, meta, err := t.latency(ctx, client)
 	if err == nil {
 		res.LatencyMs, res.JitterMs = best(lat), jitter(lat)
 		res.Colo, res.IP = meta.Colo, meta.IP
 	}
 	var note string
 	if err == nil {
 		t.set(func(r *Run) { r.Phase, r.Done, r.Mbps = "download", 0.1, 0 })
 		// Each source in turn: a refusal from the first is exactly what the second is for, and
 		// anything else (no route, no DNS) would fail the same way twice, so it stops there.
 		for i, src := range downSources {
 			res.Source = src.Name
 			res.DownMbps, res.DownBytes, note, err = t.phase(ctx, client, "download", downStreams, downFor, 0.1, 0.6, src)
 			if err == nil || !refused(err) || i == len(downSources)-1 {
 				break
 			}
 			log.Printf("speed test: %s refused the download, trying %s", src.Name, downSources[i+1].Name)
 		}
 		res.Note = join(res.Note, note)
 	}
 	if err == nil {
 		t.set(func(r *Run) { r.Phase, r.Done, r.Mbps = "upload", 0.7, 0 })
 		res.UpMbps, res.UpBytes, note, err = t.phase(ctx, client, "upload", upStreams, upFor, 0.7, 1.0, Source{Name: "cloudflare"})
 		res.Note = join(res.Note, note)
 	}
 	res.TookSec = round(time.Since(started).Seconds(), 1)
 	if err != nil {
 		if ctx.Err() != nil {
 			res.Error = "cancelled"
 		} else {
 			res.Error = err.Error()
 		}
 	}
 	t.mu.Lock()
 	t.cancel = nil
 	t.run = Run{Started: started, Phase: "done", Done: 1, Result: &res, Bytes: res.DownBytes + res.UpBytes}
 	if res.Error == "" || res.DownMbps > 0 { // a cancelled test that got a number is still a reading
 		t.hist = append(t.hist, res)
 		if len(t.hist) > t.maxHist {
 			t.hist = t.hist[len(t.hist)-t.maxHist:]
 		}
 	}
 	done, out := t.onDone, res
 	t.mu.Unlock()
 	if done != nil {
 		done(out)
 	}
 }
 // meta is what the edge says about the connection. The address comes from a header of the
 // speed endpoint's own; the edge's name is the tail of the ray ID, which every Cloudflare
 // response carries ("a41f6720f966a36c-SEA").
 type meta struct{ IP, Colo string }
 func readMeta(h http.Header) meta {
 	m := meta{IP: h.Get("cf-meta-ip")}
 	if ray := h.Get("cf-ray"); ray != "" {
 		if i := strings.LastIndexByte(ray, '-'); i >= 0 {
 			m.Colo = ray[i+1:]
 		}
 	}
 	return m
 }
 // serverMs is how long the far end says it spent on this request, from Server-Timing. The
 // cfL4 entry carries no dur= and contributes nothing, which is what we want — it is a
 // description of the connection, not time spent.
 func serverMs(h http.Header) float64 {
 	var total float64
 	for _, v := range h.Values("Server-Timing") {
 		for _, part := range strings.Split(v, ",") {
 			_, after, ok := strings.Cut(part, "dur=")
 			if !ok {
 				continue
 			}
 			end := 0
 			for end < len(after) && (after[end] == '.' || (after[end] >= '0' && after[end] <= '9')) {
 				end++
 			}
 			if f, err := strconv.ParseFloat(after[:end], 64); err == nil {
 				total += f
 			}
 		}
 	}
 	return total
 }
 // latency times several zero-byte responses on an already-open connection.
 func (t *Tester) latency(ctx context.Context, client *http.Client) ([]float64, meta, error) {
 	var m meta
 	var out []float64
 	for i := 0; i <= latencyProbes; i++ { // the first is the handshake, and is not counted
 		start := time.Now()
 		req, err := http.NewRequestWithContext(ctx, "GET", fmt.Sprintf(cfDownURL, 0), nil)
 		if err != nil {
 			return nil, m, err
 		}
 		resp, err := client.Do(req)
 		if err != nil {
 			return nil, m, err
 		}
 		_, _ = io.Copy(io.Discard, resp.Body)
 		resp.Body.Close()
 		if resp.StatusCode != 200 {
 			return nil, m, fmt.Errorf("speed endpoint answered %s", resp.Status)
 		}
 		if i == 0 {
 			m = readMeta(resp.Header)
 			continue
 		}
 		rtt := float64(time.Since(start).Microseconds())/1000 - serverMs(resp.Header)
 		if rtt < 0.01 { // the far end claimed more time than the whole request took
 			rtt = 0.01
 		}
 		out = append(out, rtt)
 		t.set(func(r *Run) { r.Done = 0.02 + 0.08*float64(i)/latencyProbes })
 	}
 	if len(out) == 0 {
 		return nil, m, errors.New("no reply from the speed endpoint")
 	}
 	return out, m, nil
 }
 // phase pushes traffic with several streams for a fixed spell and reports the rate over the
 // window after the ramp. from and to are where this phase sits in the whole test's progress.
 //
 // The window ends when the traffic does, not when the clock runs out: if the far end starts
 // refusing part-way, the figure is what was actually moved while it was moving, and the note
 // says what happened. Coasting to the end of the clock with no bytes arriving would report a
 // number a third of the truth.
 func (t *Tester) phase(parent context.Context, client *http.Client, dir string, streams int, forHow time.Duration, from, to float64, src Source) (float64, int64, string, error) {
 	// parent is the test's own context, which is cancelled when somebody presses Stop. ctx is
 	// this phase's, which *this function* cancels to shut the streams down — so the two are
 	// kept apart: asking ctx whether the test was cancelled would always say yes by then, and
 	// every failure would be reported as "context canceled" whatever really happened.
 	ctx, stop := context.WithCancel(parent)
 	defer stop()
 	var moved atomic.Int64
 	var firstErr error
 	var errOnce sync.Once
 	var alive atomic.Int32
 	alive.Store(int32(streams))
 	dead := make(chan struct{})
 	var deadOnce sync.Once
 	for i := 0; i < streams; i++ {
 		go func() {
 			defer func() {
 				if alive.Add(-1) == 0 {
 					deadOnce.Do(func() { close(dead) })
 				}
 			}()
 			tries := 0
 			for ctx.Err() == nil && moved.Load() < phaseBudget {
 				var err error
 				if dir == "download" {
 					err = stream(ctx, client, src, &moved)
 				} else {
 					err = upstream(ctx, client, &moved)
 				}
 				if err == nil {
 					tries = 0
 					continue
 				}
 				if ctx.Err() != nil {
 					return
 				}
 				errOnce.Do(func() { firstErr = err })
 				if tries++; tries > 2 { // refused three times: this stream is finished
 					return
 				}
 				select {
 				case <-time.After(400 * time.Millisecond):
 				case <-ctx.Done():
 					return
 				}
 			}
 		}()
 	}
 	// the ramp, measured and thrown away
 	select {
 	case <-time.After(warmUp):
 	case <-dead:
 	case <-ctx.Done():
 	}
 	mark, at := moved.Load(), time.Now()
 	last, lastAt := mark, at // the last time bytes were still arriving
 	deadline := at.Add(forHow)
 	tick := time.NewTicker(200 * time.Millisecond)
 	defer tick.Stop()
 	for time.Now().Before(deadline) {
 		select {
 		case <-tick.C:
 		case <-dead:
 		case <-ctx.Done():
 		}
 		if parent.Err() != nil {
 			break
 		}
 		now := moved.Load()
 		if now > last {
 			last, lastAt = now, time.Now()
 		}
 		// enough has moved to know, and it has been long enough to mean it
 		if (now-mark >= measureBytes && time.Since(at) >= minWindow) || now >= phaseBudget {
 			break
 		}
 		if secs := time.Since(at).Seconds(); secs > 0 {
 			live := mbps(now-mark, secs)
 			part := math.Min(1, secs/forHow.Seconds())
 			t.set(func(r *Run) { r.Mbps, r.Bytes, r.Done = live, now, from+(to-from)*part })
 		}
 		// every stream has given up and nothing more is coming
 		if alive.Load() == 0 {
 			break
 		}
 	}
 	total, secs := last-mark, lastAt.Sub(at).Seconds()
 	stop()
 	if total <= 0 {
 		switch {
 		case firstErr != nil: // what the far end said explains it better than anything else
 			return 0, moved.Load(), "", firstErr
 		case parent.Err() != nil:
 			return 0, moved.Load(), "", parent.Err()
 		}
 		return 0, moved.Load(), "", fmt.Errorf("nothing moved in the %s phase", dir)
 	}
 	if secs < 0.02 { // too short a window to divide by; the guard is against dividing by zero,
 		// not a judgement about how long a good measurement takes
 		return 0, moved.Load(), "", fmt.Errorf("the %s phase was over before it could be measured", dir)
 	}
 	note := ""
 	if firstErr != nil {
 		note = fmt.Sprintf("the %s phase was cut short: %s", dir, firstErr)
 	}
 	return mbps(total, secs), moved.Load(), note, nil
 }
 // join puts two notes together without leaving an empty one dangling.
 func join(a, b string) string {
 	switch {
 	case a == "":
 		return b
 	case b == "":
 		return a
 	}
 	return a + "; " + b
 }
 // stream fetches one chunk from src, counting bytes as they arrive.
 func stream(ctx context.Context, client *http.Client, src Source, n *atomic.Int64) error {
 	req, err := http.NewRequestWithContext(ctx, "GET", src.URL(downChunk), nil)
 	if err != nil {
 		return err
 	}
 	resp, err := client.Do(req)
 	if err != nil {
 		return err
 	}
 	defer resp.Body.Close()
 	if resp.StatusCode != 200 {
 		return answered(resp.StatusCode, resp.Status)
 	}
 	buf := make([]byte, 64<<10)
 	for {
 		got, err := resp.Body.Read(buf)
 		if got > 0 {
 			n.Add(int64(got))
 		}
 		if err != nil {
 			if err == io.EOF {
 				return nil
 			}
 			return err
 		}
 	}
 }
 // upstream posts one chunk of nothing in particular, counting bytes as they go out.
 func upstream(ctx context.Context, client *http.Client, n *atomic.Int64) error {
 	body := &zeros{left: upChunk, n: n}
 	req, err := http.NewRequestWithContext(ctx, "POST", upURL, body)
 	if err != nil {
 		return err
 	}
 	req.ContentLength = upChunk
 	req.Header.Set("Content-Type", "application/octet-stream")
 	resp, err := client.Do(req)
 	if err != nil {
 		return err
 	}
 	_, _ = io.Copy(io.Discard, resp.Body)
 	resp.Body.Close()
 	if resp.StatusCode != 200 && resp.StatusCode != 204 {
 		return answered(resp.StatusCode, resp.Status)
 	}
 	return nil
 }
 // answered turns the far end's refusal into something worth reading. It rate-limits by
 // volume: enough testing in a short while and it stops serving the big transfers, which is
 // the only failure of this test anyone here is likely to meet.
 func answered(code int, status string) error {
 	if code == 429 || code == 403 {
 		return errRefused
 	}
 	return fmt.Errorf("the speed endpoint answered %s", status)
 }
 // errRefused is the far end rate-limiting this connection, which is the only failure of this
 // test anyone here is likely to meet: it limits by volume, and a test moves a few hundred
 // megabytes. It is recognisable so the download phase can try its other source.
 var errRefused = errors.New("the speed endpoint is rate-limiting this connection — it does that after a lot of testing; leave it a few minutes")
 func refused(err error) bool { return errors.Is(err, errRefused) }
 // zeros is a request body of a fixed size that costs nothing to produce, counting itself as
 // it is read out onto the wire.
 type zeros struct {
 	left int
 	n    *atomic.Int64
 }
 func (z *zeros) Read(p []byte) (int, error) {
 	if z.left <= 0 {
 		return 0, io.EOF
 	}
 	if len(p) > z.left {
 		p = p[:z.left]
 	}
 	for i := range p {
 		p[i] = 0
 	}
 	z.left -= len(p)
 	z.n.Add(int64(len(p)))
 	return len(p), nil
 }
 func newTransport(streams int) *http.Transport {
 	return &http.Transport{
 		Proxy:               http.ProxyFromEnvironment,
 		DialContext:         (&net.Dialer{Timeout: 10 * time.Second}).DialContext,
 		TLSHandshakeTimeout: 10 * time.Second,
 		MaxIdleConns:        streams * 2,
 		MaxIdleConnsPerHost: streams * 2,
 		// An empty map turns HTTP/2 off: see the note at the top of this file.
 		TLSNextProto: map[string]func(string, *tls.Conn) http.RoundTripper{},
 	}
 }
 // probeOutside asks the far end who it thinks we are. One small request.
 func probeOutside() Outside {
 	tr := newTransport(1)
 	defer tr.CloseIdleConnections()
 	client := &http.Client{Transport: tr, Timeout: 15 * time.Second}
 	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second)
 	defer cancel()
 	req, err := http.NewRequestWithContext(ctx, "GET", traceURL, nil)
 	if err != nil {
 		return Outside{At: time.Now(), Error: err.Error()}
 	}
 	resp, err := client.Do(req)
 	if err != nil {
 		return Outside{At: time.Now(), Error: shortErr(err)}
 	}
 	defer resp.Body.Close()
 	body, _ := io.ReadAll(io.LimitReader(resp.Body, 4<<10))
 	if resp.StatusCode != 200 {
 		return Outside{At: time.Now(), Error: "the trace endpoint answered " + resp.Status}
 	}
 	o := Outside{At: time.Now()}
 	for _, line := range strings.Split(string(body), "\n") {
 		k, v, ok := strings.Cut(line, "=")
 		if !ok {
 			continue
 		}
 		switch k {
 		case "ip":
 			o.IP = v
 		case "colo":
 			o.Colo = v
 		case "loc":
 			o.Country = v
 		}
 	}
 	if o.IP == "" {
 		o.Error = "the trace endpoint said nothing about this connection"
 	}
 	return o
 }
 // shortErr keeps the part of a network error a person can act on.
 func shortErr(err error) string {
 	s := err.Error()
 	if i := strings.LastIndex(s, ": "); i > 0 && len(s)-i < 60 {
 		return s[i+2:]
 	}
 	return s
 }
 func mbps(bytes int64, secs float64) float64 {
 	return round(float64(bytes)*8/secs/1e6, 1)
 }
 func best(v []float64) float64 {
 	m := v[0]
 	for _, x := range v {
 		if x < m {
 			m = x
 		}
 	}
 	return round(m, 1)
 }
 // jitter is how much the line wobbles: the average change from one round trip to the next,
 // which is the usual definition and, unlike a standard deviation, is not run away with by a
 // single slow probe.
 func jitter(v []float64) float64 {
 	if len(v) < 2 {
 		return 0
 	}
 	var sum float64
 	for i := 1; i < len(v); i++ {
 		sum += math.Abs(v[i] - v[i-1])
 	}
 	return round(sum/float64(len(v)-1), 1)
 }
 func round(v float64, places int) float64 {
 	p := math.Pow(10, float64(places))
 	return math.Round(v*p) / p
 }
 // Mbits renders a rate the way the line is sold, for a log line.
 func Mbits(v float64) string { return strconv.FormatFloat(v, 'f', 1, 64) + " Mbps" }
@@ -0,0 +1,25 @@
 #!/bin/sh /etc/rc.common
 # Ham Lab web console (written by hamlab install)
 USE_PROCD=1
 # after network and firewall, before the tail of the boot
 START=95
 STOP=10
 start_service() {
 	procd_open_instance hamlab
 	procd_set_param command "@BIN@" serve
 	procd_set_param env HAMLAB_CONFIG="@CONFIG@" HOME="@HOME@" USER="@USER@"
 	procd_set_param user "@USER@"
 	procd_set_param group "@GROUP@"
 	procd_set_param pidfile /var/run/hamlab.pid
 	# respawn: settle 3s, 5 retries inside that window, then give up counting
 	procd_set_param respawn 3600 3 0
 	procd_set_param stdout 1
 	procd_set_param stderr 1
 	procd_close_instance
 }
 service_triggers() {
 	procd_add_reload_trigger network
 }
@@ -37,9 +37,14 @@ var systemdUnit string
 //go:embed hamlab.openrc
 var openrcScript string
 //go:embed hamlab.procd
 var procdScript string
 const (
-	unitPath    = "/etc/systemd/system/hamlab.service"
+	unitPath   = "/etc/systemd/system/hamlab.service"
-	openrcPath  = "/etc/init.d/hamlab"
+	openrcPath = "/etc/init.d/hamlab"
 	// procd (OpenWrt) uses the same path as OpenRC; the two are told apart by content.
 	procdPath   = "/etc/init.d/hamlab"
 	sudoersPath = "/etc/sudoers.d/hamlab-quiet"
 )
@@ -228,6 +233,19 @@ func Install(args []string, version string) error {
 				return err
 			}
 		}
 	case exists("/sbin/procd") && exists("/etc/rc.common"):
 		initSys = "procd"
 		if err := os.WriteFile(procdPath, []byte(fill(procdScript, vars)), 0o755); err != nil {
 			return err
 		}
 		if err := run(procdPath, "enable"); err != nil {
 			say("warning: %s enable: %v", procdPath, err)
 		}
 		if !o.noStart {
 			if err := run(procdPath, "restart"); err != nil {
 				return err
 			}
 		}
 	case exists("/sbin/openrc-run") || exists("/usr/sbin/openrc-run") || have("rc-service"):
 		initSys = "OpenRC"
 		if err := os.WriteFile(openrcPath, []byte(fill(openrcScript, vars)), 0o755); err != nil {
@@ -258,7 +276,7 @@ func Install(args []string, version string) error {
 	url := scheme + "://" + browseAddr(cfg.Listen)
 	switch {
 	case initSys == "":
-		say("\nNo systemd or OpenRC here, so nothing starts it at boot.\nRun it under your own supervisor, as %s:\n  HAMLAB_CONFIG=%s %s serve", acct.name, cfgPath, dest)
+		say("\nNo systemd, procd or OpenRC here, so nothing starts it at boot.\nRun it under your own supervisor, as %s:\n  HAMLAB_CONFIG=%s %s serve", acct.name, cfgPath, dest)
 	case o.noStart:
 	case healthy(cfg.Listen, cfg.TLS == "auto"):
 		say("\nHam Lab is up: %s", url)
@@ -294,9 +312,15 @@ func Uninstall(args []string) error {
 		_ = run("systemctl", "daemon-reload")
 	}
 	if b, err := os.ReadFile(openrcPath); err == nil {
-		dataDir = configDirFrom(string(b), "export HAMLAB_CONFIG=")
+		if strings.Contains(string(b), "USE_PROCD=1") {
-		_ = run("rc-service", "hamlab", "stop")
+			dataDir = configDirFrom(string(b), "HAMLAB_CONFIG=")
-		_ = run("rc-update", "del", "hamlab", "default")
+			_ = run(procdPath, "stop")
 			_ = run(procdPath, "disable")
 		} else {
 			dataDir = configDirFrom(string(b), "export HAMLAB_CONFIG=")
 			_ = run("rc-service", "hamlab", "stop")
 			_ = run("rc-update", "del", "hamlab", "default")
 		}
 		os.Remove(openrcPath)
 	}
 	os.Remove(sudoersPath)
@@ -0,0 +1,292 @@
 package storage
 // The ZFS caching tiers, for the ZFS tab: what the ARC has done since boot and how it is
 // tuned, and for each pool which cache devices it has — L2ARC, a separate log (SLOG), a
 // special allocation class, a dedup table — or, as is common, none of them, which the tab
 // should say plainly rather than leave a blank.
 //
 // The live numbers (size, the hit rate over the last interval) come from the sampler in
 // internal/metrics, which also publishes them as zfs.arc.* sensors; this is the slower,
 // fuller picture, read when the tab asks.
 import (
 	"context"
 	"os"
 	"os/exec"
 	"strconv"
 	"strings"
 	"sync"
 	"time"
 )
 // CacheDev is one member of a pool's cache, log, special, dedup or spare class.
 type CacheDev struct {
 	Name   string `json:"name"`
 	Size   uint64 `json:"size,omitempty"`
 	Alloc  uint64 `json:"alloc,omitempty"`
 	Health string `json:"health,omitempty"`
 }
 // PoolCaching is what one pool has beyond its data vdevs, and the properties that steer
 // what is cached.
 type PoolCaching struct {
 	Pool    string     `json:"pool"`
 	Cache   []CacheDev `json:"cache"`   // L2ARC
 	Log     []CacheDev `json:"log"`     // SLOG
 	Special []CacheDev `json:"special"` // metadata / small blocks
 	Dedup   []CacheDev `json:"dedup"`   // dedup table vdevs
 	Spare   []CacheDev `json:"spare"`
 	// dataset properties at the pool root
 	PrimaryCache   string `json:"primarycache"`
 	SecondaryCache string `json:"secondarycache"`
 	Sync           string `json:"sync"`
 	LogBias        string `json:"logbias"`
 	DedupProp      string `json:"dedup_prop"`
 	Compression    string `json:"compression"`
 	RecordSize     string `json:"recordsize"`
 	DedupRatio     string `json:"dedup_ratio"` // zpool's, e.g. "1.00x"
 }
 // ARCDetail is the full picture.
 type ARCDetail struct {
 	Available bool              `json:"available"`
 	MemTotal  uint64            `json:"mem_total"`
 	Tunables  map[string]string `json:"tunables"` // zfs module parameters that shape the cache; "" when unset
 	Lifetime  struct {
 		Hits, Misses                                   uint64 `json:",omitempty"`
 		DemandDataHits, DemandDataMisses               uint64
 		DemandMetaHits, DemandMetaMisses               uint64
 		PrefetchDataHits, PrefetchDataMisses           uint64
 		PrefetchMetaHits, PrefetchMetaMisses           uint64
 		L2Hits, L2Misses, L2ReadBytes, L2WriteBytes    uint64
 		L2Feeds, L2RebuildSize                         uint64
 		EvictL2Eligible                                uint64 // bytes evicted from RAM that an L2ARC could have caught
 		Deleted, EvictSkip, MutexMiss, Prune, Throttle uint64
 	} `json:"lifetime"`
 	Zfetch struct{ Hits, Misses, IOIssued uint64 } `json:"zfetch"` // the prefetcher
 	ZIL    struct {
 		Commits                  uint64 `json:"commits"`
 		NormalBytes, NormalWrite uint64 // written to the pool's own vdevs
 		SlogBytes, SlogWrite     uint64 // written to a separate log device
 	} `json:"zil"`
 	DBuf  struct{ Size, Target, Hits, Misses uint64 } `json:"dbuf"` // the dbuf cache in front of the ARC
 	Pools []PoolCaching                               `json:"pools"`
 }
 const kstatDir = "/proc/spl/kstat/zfs/"
 func readKstat(p string) map[string]uint64 {
 	b, err := os.ReadFile(p)
 	if err != nil {
 		return nil
 	}
 	out := map[string]uint64{}
 	for _, line := range strings.Split(string(b), "\n") {
 		f := strings.Fields(line)
 		if len(f) != 3 || f[1] != "4" {
 			continue
 		}
 		if v, err := strconv.ParseUint(f[2], 10, 64); err == nil {
 			out[f[0]] = v
 		}
 	}
 	return out
 }
 var arcTunables = []string{"zfs_arc_max", "zfs_arc_min", "zfs_arc_sys_free", "zfs_arc_meta_balance", "zfs_prefetch_disable", "l2arc_write_max", "l2arc_write_boost", "l2arc_noprefetch", "l2arc_rebuild_enabled", "zfs_dirty_data_max"}
 const poolCachingEvery = 30 * time.Second // zpool list -v and zfs get are an exec each
 // ARC reads the kstats now and the pool layout from a short cache.
 func (m *Manager) ARC(ctx context.Context) ARCDetail {
 	d := ARCDetail{Tunables: map[string]string{}, Pools: []PoolCaching{}}
 	k := readKstat(kstatDir + "arcstats")
 	if k == nil {
 		return d
 	}
 	d.Available = true
 	l := &d.Lifetime
 	l.Hits, l.Misses = k["hits"], k["misses"]
 	l.DemandDataHits, l.DemandDataMisses = k["demand_data_hits"], k["demand_data_misses"]
 	l.DemandMetaHits, l.DemandMetaMisses = k["demand_metadata_hits"], k["demand_metadata_misses"]
 	l.PrefetchDataHits, l.PrefetchDataMisses = k["prefetch_data_hits"], k["prefetch_data_misses"]
 	l.PrefetchMetaHits, l.PrefetchMetaMisses = k["prefetch_metadata_hits"], k["prefetch_metadata_misses"]
 	l.L2Hits, l.L2Misses, l.L2ReadBytes, l.L2WriteBytes = k["l2_hits"], k["l2_misses"], k["l2_read_bytes"], k["l2_write_bytes"]
 	l.L2Feeds, l.L2RebuildSize, l.EvictL2Eligible = k["l2_feeds"], k["l2_rebuild_size"], k["evict_l2_eligible"]
 	l.Deleted, l.EvictSkip, l.MutexMiss, l.Prune, l.Throttle = k["deleted"], k["evict_skip"], k["mutex_miss"], k["arc_prune"], k["memory_throttle_count"]
 	if z := readKstat(kstatDir + "zfetchstats"); z != nil {
 		d.Zfetch.Hits, d.Zfetch.Misses, d.Zfetch.IOIssued = z["hits"], z["misses"], z["io_issued"]
 	}
 	if z := readKstat(kstatDir + "zil"); z != nil {
 		d.ZIL.Commits = z["zil_commit_count"]
 		d.ZIL.NormalBytes, d.ZIL.NormalWrite = z["zil_itx_metaslab_normal_bytes"], z["zil_itx_metaslab_normal_write"]
 		d.ZIL.SlogBytes, d.ZIL.SlogWrite = z["zil_itx_metaslab_slog_bytes"], z["zil_itx_metaslab_slog_write"]
 	}
 	if z := readKstat(kstatDir + "dbufstats"); z != nil {
 		d.DBuf.Size, d.DBuf.Target, d.DBuf.Hits, d.DBuf.Misses = z["cache_size_bytes"], z["cache_target_bytes"], z["hash_hits"], z["hash_misses"]
 	}
 	for _, t := range arcTunables {
 		if b, err := os.ReadFile("/sys/module/zfs/parameters/" + t); err == nil {
 			d.Tunables[t] = strings.TrimSpace(string(b))
 		}
 	}
 	if b, err := os.ReadFile("/proc/meminfo"); err == nil {
 		for _, line := range strings.Split(string(b), "\n") {
 			if strings.HasPrefix(line, "MemTotal:") {
 				f := strings.Fields(line)
 				if len(f) >= 2 {
 					d.MemTotal = u64(f[1]) << 10
 				}
 				break
 			}
 		}
 	}
 	d.Pools = m.poolCaching(ctx)
 	return d
 }
 var cachingMu sync.Mutex
 var cachingAt time.Time
 var cachingLast []PoolCaching
 func (m *Manager) poolCaching(ctx context.Context) []PoolCaching {
 	cachingMu.Lock()
 	defer cachingMu.Unlock()
 	if cachingLast != nil && time.Since(cachingAt) < poolCachingEvery {
 		return cachingLast
 	}
 	out := []PoolCaching{}
 	bin, ok := lookPath("zpool")
 	if !ok {
 		return out
 	}
 	// zpool list -v needs no root; the vdev tree under each pool, with the class headers
 	raw, code, err := m.runPlain(ctx, bin, "list", "-Hpv", "-o", "name,size,alloc,free,health")
 	if err != nil || code != 0 {
 		return out
 	}
 	out = parsePoolCaching(string(raw))
 	if zfs, ok := lookPath("zfs"); ok {
 		for i := range out {
 			p := &out[i]
 			if props, code, err := m.runPlain(ctx, zfs, "get", "-Hp", "-o", "property,value", "primarycache,secondarycache,sync,logbias,dedup,compression,recordsize", p.Pool); err == nil && code == 0 {
 				for _, line := range strings.Split(strings.TrimSpace(string(props)), "\n") {
 					f := strings.Split(line, "\t")
 					if len(f) < 2 {
 						continue
 					}
 					switch f[0] {
 					case "primarycache":
 						p.PrimaryCache = f[1]
 					case "secondarycache":
 						p.SecondaryCache = f[1]
 					case "sync":
 						p.Sync = f[1]
 					case "logbias":
 						p.LogBias = f[1]
 					case "dedup":
 						p.DedupProp = f[1]
 					case "compression":
 						p.Compression = f[1]
 					case "recordsize":
 						p.RecordSize = humanSize(u64(f[1]))
 					}
 				}
 			}
 		}
 	}
 	for _, pool := range parsePools(func() []byte {
 		b, _, _ := m.runPlain(ctx, bin, "list", "-Hp", "-o", "name,size,alloc,free,frag,cap,health,dedup")
 		return b
 	}()) {
 		for i := range out {
 			if out[i].Pool == pool.Name {
 				out[i].DedupRatio = pool.Dedup
 			}
 		}
 	}
 	cachingLast, cachingAt = out, time.Now()
 	return out
 }
 // runPlain runs a command as ourselves, never through sudo: these are read-only zpool/zfs
 // queries that need no root on Linux, and a poll must not write to the auth log.
 func (m *Manager) runPlain(ctx context.Context, bin string, args ...string) ([]byte, int, error) {
 	ctx, cancel := context.WithTimeout(ctx, 30*time.Second)
 	defer cancel()
 	out, err := exec.CommandContext(ctx, bin, args...).Output()
 	if ee, ok := err.(*exec.ExitError); ok {
 		return out, ee.ExitCode(), nil
 	}
 	return out, 0, err
 }
 var cacheClasses = map[string]bool{"logs": true, "log": true, "cache": true, "special": true, "dedup": true, "spares": true, "spare": true}
 // parsePoolCaching reads `zpool list -Hpv -o name,size,alloc,free,health`: a pool line, then
 // its vdevs indented, with a bare class name ("logs", "cache", "special", "dedup", "spares")
 // introducing each extra class. Only those classes' members are kept; the data vdevs are the
 // pool and are shown elsewhere.
 func parsePoolCaching(text string) []PoolCaching {
 	var out []PoolCaching
 	var cur *PoolCaching
 	class := ""
 	for _, raw := range strings.Split(text, "\n") {
 		if strings.TrimSpace(raw) == "" {
 			continue
 		}
 		indented := strings.HasPrefix(raw, " ") || strings.HasPrefix(raw, "\t")
 		f := strings.Split(strings.TrimSpace(raw), "\t")
 		name := strings.TrimSpace(f[0])
 		if !indented {
 			out = append(out, PoolCaching{Pool: name, Cache: []CacheDev{}, Log: []CacheDev{}, Special: []CacheDev{}, Dedup: []CacheDev{}, Spare: []CacheDev{}})
 			cur, class = &out[len(out)-1], ""
 			continue
 		}
 		if cur == nil {
 			continue
 		}
 		if cacheClasses[name] && (len(f) < 2 || f[1] == "-" || f[1] == "") {
 			class = name
 			continue
 		}
 		if class == "" {
 			continue // a data vdev or one of its disks
 		}
 		// a mirror or raidz inside a class lists its members below it; keep them all, the
 		// grouping is visible in the vdev tree on the pool card
 		dev := CacheDev{Name: name}
 		if len(f) >= 5 {
 			dev.Size, dev.Alloc, dev.Health = u64(f[1]), u64(f[2]), strings.TrimSpace(f[4])
 			if f[1] == "-" {
 				dev.Size = 0
 			}
 		}
 		switch class {
 		case "logs", "log":
 			cur.Log = append(cur.Log, dev)
 		case "cache":
 			cur.Cache = append(cur.Cache, dev)
 		case "special":
 			cur.Special = append(cur.Special, dev)
 		case "dedup":
 			cur.Dedup = append(cur.Dedup, dev)
 		case "spares", "spare":
 			cur.Spare = append(cur.Spare, dev)
 		}
 	}
 	if out == nil {
 		out = []PoolCaching{}
 	}
 	return out
 }
 func humanSize(b uint64) string {
 	switch {
 	case b >= 1<<20 && b%(1<<20) == 0:
 		return strconv.FormatUint(b>>20, 10) + "M"
 	case b >= 1<<10 && b%(1<<10) == 0:
 		return strconv.FormatUint(b>>10, 10) + "K"
 	}
 	return strconv.FormatUint(b, 10)
 }
@@ -0,0 +1,71 @@
 package storage
 import (
 	"context"
 	"encoding/json"
 	"os"
 	"testing"
 )
 // zpool list -Hpv on a pool with every extra class; the data raidz and its disks are skipped
 const listV = "tank\t32000000000000\t23000000000000\t9000000000000\tONLINE\n" +
 	"\traidz2-0\t32000000000000\t23000000000000\t9000000000000\tONLINE\n" +
 	"\t\twwn-0x1\t4000000000000\t-\t-\tONLINE\n" +
 	"\t\twwn-0x2\t4000000000000\t-\t-\tONLINE\n" +
 	"\tspecial\t-\t-\t-\t-\n" +
 	"\t\tmirror-1\t500000000000\t120000000000\t380000000000\tONLINE\n" +
 	"\t\t\tnvme0n1p1\t500000000000\t-\t-\tONLINE\n" +
 	"\t\t\tnvme1n1p1\t500000000000\t-\t-\tONLINE\n" +
 	"\tlogs\t-\t-\t-\t-\n" +
 	"\t\tnvme2n1p1\t16000000000\t4000000\t15996000000\tONLINE\n" +
 	"\tcache\t-\t-\t-\t-\n" +
 	"\t\tnvme2n1p2\t400000000000\t210000000000\t190000000000\tONLINE\n" +
 	"\tspares\t-\t-\t-\t-\n" +
 	"\t\twwn-0x9\t4000000000000\t-\t-\tAVAIL\n" +
 	"scratch\t1000000000000\t1000000000\t999000000000\tONLINE\n" +
 	"\tsda\t1000000000000\t1000000000\t999000000000\tONLINE\n"
 func TestParsePoolCaching(t *testing.T) {
 	got := parsePoolCaching(listV)
 	if len(got) != 2 || got[0].Pool != "tank" || got[1].Pool != "scratch" {
 		t.Fatalf("pools: %+v", got)
 	}
 	tank := got[0]
 	if len(tank.Cache) != 1 || tank.Cache[0].Name != "nvme2n1p2" || tank.Cache[0].Alloc != 210000000000 {
 		t.Errorf("cache: %+v", tank.Cache)
 	}
 	if len(tank.Log) != 1 || tank.Log[0].Name != "nvme2n1p1" {
 		t.Errorf("log: %+v", tank.Log)
 	}
 	if len(tank.Special) != 3 || tank.Special[0].Name != "mirror-1" || tank.Special[0].Alloc != 120000000000 {
 		t.Errorf("special: %+v", tank.Special)
 	}
 	if len(tank.Spare) != 1 || tank.Spare[0].Health != "AVAIL" {
 		t.Errorf("spares: %+v", tank.Spare)
 	}
 	if len(tank.Dedup) != 0 {
 		t.Errorf("no dedup class: %+v", tank.Dedup)
 	}
 	s := got[1]
 	if len(s.Cache)+len(s.Log)+len(s.Special)+len(s.Spare) != 0 {
 		t.Errorf("a plain pool has no extra classes: %+v", s)
 	}
 }
 func TestHumanSize(t *testing.T) {
 	for in, want := range map[uint64]string{131072: "128K", 1048576: "1M", 16384: "16K", 1000: "1000"} {
 		if got := humanSize(in); got != want {
 			t.Errorf("%d → %s, want %s", in, got, want)
 		}
 	}
 }
 // TestLiveARC reads this machine's ZFS for real: HAMLAB_LIVE=1 go test -run TestLiveARC -v ./internal/storage
 func TestLiveARC(t *testing.T) {
 	if os.Getenv("HAMLAB_LIVE") == "" {
 		t.Skip("set HAMLAB_LIVE=1 to read the real machine")
 	}
 	d := New(false).ARC(context.Background())
 	b, _ := json.MarshalIndent(d, "", "  ")
 	t.Logf("%s", b)
 }
@@ -452,8 +452,25 @@ func (m *Manager) WriteCronFile(ctx context.Context, policies []Policy, user, ex
 	return m.writeRoot(ctx, CronFile, text)
 }
-// writeRoot puts text at a root-owned path, through the same privilege escalation every
+// WriteRootFile puts text at a root-owned path, through the same privilege escalation every
-// other write here uses.
+// other write here uses. Shared by the snapshot schedules and the image-update schedules,
 // which both own a file in /etc/cron.d.
 func (m *Manager) WriteRootFile(ctx context.Context, path, text string) error {
 	return m.writeRoot(ctx, path, text)
 }
 // RemoveRootFile deletes one, for when the last schedule of a kind goes away.
 func (m *Manager) RemoveRootFile(ctx context.Context, path string) error {
 	out, code, err := m.run(ctx, "rm", "-f", path)
 	if err != nil {
 		return err
 	}
 	if code != 0 {
 		return errors.New(strings.TrimSpace(string(out)))
 	}
 	return nil
 }
 func (m *Manager) writeRoot(ctx context.Context, path, text string) error {
 	args := []string{"tee", path}
 	if m.sudo {
@@ -44,6 +44,16 @@ type UserState struct {
 	UI json.RawMessage `json:"ui,omitempty"` // theme, accents; opaque to the server
 }
 // RememberedMount is one bind mount of a container that was running when Ham Lab last
 // looked. A compose stack records the path its file had *inside* whatever deployed it, so
 // when that container is gone the path is the only way back to the file on disk.
 type RememberedMount struct {
 	Container string    `json:"container"`
 	Dest      string    `json:"dest"` // the path inside the container
 	Src       string    `json:"src"`  // the path on the machine
 	Seen      time.Time `json:"seen"`
 }
 type State struct {
 	Terminals []Terminal            `json:"terminals"`
 	Sessions  map[string]Session    `json:"sessions"` // token -> session
@@ -51,10 +61,20 @@ type State struct {
 	// UI is the pre-users machine-wide look; on first boot with users it moves to the first
 	// account. Kept so an old state file still reads.
 	UI         json.RawMessage `json:"ui,omitempty"`
-	Alerts     json.RawMessage `json:"alerts,omitempty"`
+	Alerts     json.RawMessage `json:"alerts,omitempty"`      // rules, channels and recent events (internal/api/alerts.go)
-	ZFS        json.RawMessage `json:"zfs,omitempty"`  // snapshot schedules (internal/api/zfs.go)      // rules, channels and recent events (internal/api/alerts.go)
+	ZFS        json.RawMessage `json:"zfs,omitempty"`         // snapshot schedules (internal/api/zfs.go)
-	SamplerMs  int             `json:"sampler_ms,omitempty"`  // how often this machine samples; the hub keeps every machine on the console's refresh rate
+	DockerJobs json.RawMessage `json:"docker_jobs,omitempty"` // image-update schedules (internal/api/dockersched.go)
-	HistorySec int             `json:"history_sec,omitempty"` // how far back the graphs reach; the hub keeps every machine on the same one
+	Network    json.RawMessage `json:"network,omitempty"`     // speed test results (internal/api/network.go)
 	// Mounts remembers where a container's bind mounts pointed, so a stack deployed by an
 	// orchestrator that has since been removed can still be found on disk. See
 	// internal/api/stacks.go.
 	Mounts []RememberedMount `json:"mounts,omitempty"`
 	// KVMCookies is each KVM's own login, as the cookies its page set in a browser: kept so
 	// the next browser — or this one after it has forgotten them — is signed into the device
 	// already. See internal/api/kvm.go.
 	KVMCookies map[string]map[string]string `json:"kvm_cookies,omitempty"`
 	SamplerMs  int                          `json:"sampler_ms,omitempty"`  // how often this machine samples; the hub keeps every machine on the console's refresh rate
 	HistorySec int                          `json:"history_sec,omitempty"` // how far back the graphs reach; the hub keeps every machine on the same one
 }
 // UserState returns the prefs of an account, making an empty one if needed.
@@ -84,10 +84,21 @@ func newID() string {
 // Create starts a new login shell. id may be empty.
 func (m *Manager) Create(id, name string) (*Session, error) {
 	return m.CreateCmd(id, name, nil)
 }
 // CreateCmd is Create with something other than a login shell: argv nil is the shell, and
 // anything else runs that instead — `docker exec -it …` for a container's console. Every
 // session is the same thing afterwards, so attach, resize, the scrollback ring and the tab
 // strip work without knowing which it is.
 func (m *Manager) CreateCmd(id, name string, argv []string) (*Session, error) {
 	if id == "" {
 		id = newID()
 	}
 	cmd := exec.Command(m.shell, "-l")
 	if len(argv) > 0 {
 		cmd = exec.Command(argv[0], argv[1:]...)
 	}
 	cmd.Dir = m.home
 	cmd.Env = append(os.Environ(), "TERM=xterm-256color", "COLORTERM=truecolor", "HAMLAB=1")
 	ptmx, err := pty.StartWithSize(cmd, &pty.Winsize{Cols: defaultCol, Rows: defaultRow})
@@ -0,0 +1,575 @@
 package vpn
 // Adding, renaming and removing the devices that dial in.
 //
 // Only OpenWrt's UCI is written, because that is where the server here is configured and
 // because a half-supported second path is worse than an honest refusal. A wg-quick host is
 // reported as unmanaged and the buttons do not appear.
 //
 // Two things happen for every change, in the order that fails safest:
 //
 //   - Adding: the configuration is written first and committed, then the peer is applied to
 //     the running tunnel. If applying fails the configuration is rolled back, so the two never
 //     disagree silently — a peer that works only after the next reboot is the worst outcome
 //     of the three.
 //   - Removing: the running tunnel first, because that is the moment access actually ends,
 //     then the configuration. If the second half fails the caller is told plainly that the
 //     device is off now but would return on a reboot.
 //
 // Nothing here restarts the network. `wg set` changes one peer on a live interface and
 // touches nothing else, which matters rather a lot when the interface belongs to the router
 // the house is behind.
 import (
 	"context"
 	"errors"
 	"fmt"
 	"net"
 	"net/netip"
 	"os"
 	"os/exec"
 	"regexp"
 	"sort"
 	"strconv"
 	"strings"
 )
 // NameRe is what a device may be called: a short slug, because it also names the config
 // section it is stored in.
 var NameRe = regexp.MustCompile(`^[a-zA-Z0-9][a-zA-Z0-9 ._-]{0,30}$`)
 // sectionRe guards the identifier that goes into a uci path.
 var sectionRe = regexp.MustCompile(`^[a-zA-Z0-9_]{1,48}$`)
 const sectionPrefix = "wgpeer_"
 var (
 	ErrUnmanaged = errors.New("this tunnel is not configured through UCI, so Ham Lab will not edit it here")
 	ErrNoRoom    = errors.New("no free address is left in this tunnel's range")
 )
 // NewPeer is what the caller asks for.
 type NewPeer struct {
 	Iface string `json:"-"`
 	Name  string `json:"name"`
 	// Address is the tunnel address to give the device, with or without a prefix. Empty means
 	// the next free one.
 	Address string `json:"address"`
 	// Endpoint is what the device will dial: the house's name or address, and the port.
 	Endpoint string `json:"endpoint"`
 	// DNS is what the device should use while connected. Empty leaves the line out.
 	DNS string `json:"dns"`
 	// Split routes only the home networks through the tunnel instead of everything.
 	Split bool `json:"split"`
 }
 // Created is a new device, and the one chance to read its configuration.
 type Created struct {
 	Section   string `json:"section"`
 	Name      string `json:"name"`
 	PublicKey string `json:"public_key"`
 	Address   string `json:"address"`
 	// Config is the complete client profile. It holds the private key, which exists nowhere
 	// else: it is generated here, handed over once and never stored.
 	Config string `json:"config"`
 }
 // uciPeer is one peer section as UCI holds it.
 type uciPeer struct {
 	Section    string
 	Iface      string // the wireguard_<iface> type it belongs to
 	Desc       string
 	PublicKey  string
 	AllowedIPs []string
 }
 // uciPeers reads every peer section. The type of a peer section names its interface
 // ("wireguard_wg0"), which is how a peer is tied to a tunnel.
 func uciPeers(ctx context.Context) ([]uciPeer, error) {
 	if _, err := exec.LookPath("uci"); err != nil {
 		return nil, ErrUnmanaged
 	}
 	b, err := exec.CommandContext(ctx, "uci", "-q", "show", "network").Output()
 	if err != nil {
 		return nil, err
 	}
 	by := map[string]*uciPeer{}
 	var order []string
 	for _, line := range strings.Split(string(b), "\n") {
 		lhs, rhs, ok := strings.Cut(line, "=")
 		if !ok || !strings.HasPrefix(lhs, "network.") {
 			continue
 		}
 		parts := strings.Split(strings.TrimPrefix(lhs, "network."), ".")
 		v := strings.Trim(rhs, "'")
 		section := parts[0]
 		if len(parts) == 1 { // the section's own type
 			if !strings.HasPrefix(v, "wireguard_") {
 				continue
 			}
 			by[section] = &uciPeer{Section: section, Iface: strings.TrimPrefix(v, "wireguard_")}
 			order = append(order, section)
 			continue
 		}
 		p := by[section]
 		if p == nil {
 			continue
 		}
 		switch parts[1] {
 		case "description":
 			p.Desc = v
 		case "public_key":
 			p.PublicKey = v
 		case "allowed_ips":
 			p.AllowedIPs = strings.Fields(strings.ReplaceAll(v, ",", " "))
 		}
 	}
 	out := make([]uciPeer, 0, len(order))
 	for _, s := range order {
 		out = append(out, *by[s])
 	}
 	return out, nil
 }
 // Managed reports whether this tunnel's peers can be edited here, and labels each peer with
 // the section that holds it so the caller has a safe identifier to act on.
 func Managed(ctx context.Context, t *Tunnel) bool {
 	peers, err := uciPeers(ctx)
 	if err != nil {
 		return false
 	}
 	found := false
 	byKey := map[string]string{}
 	for _, p := range peers {
 		if p.Iface == t.Name {
 			byKey[p.PublicKey] = p.Section
 			found = true
 		}
 	}
 	for i := range t.Peers {
 		t.Peers[i].Section = byKey[t.Peers[i].PublicKey]
 	}
 	// A tunnel with no peers at all is still manageable if UCI is the thing configuring it.
 	if !found {
 		for _, line := range uciInterfaces(ctx) {
 			if line == t.Name {
 				return true
 			}
 		}
 	}
 	return found
 }
 // uciInterfaces is every interface UCI configures as wireguard.
 func uciInterfaces(ctx context.Context) []string {
 	var out []string
 	if _, err := exec.LookPath("uci"); err != nil {
 		return out
 	}
 	b, err := exec.CommandContext(ctx, "uci", "-q", "show", "network").Output()
 	if err != nil {
 		return out
 	}
 	for _, line := range strings.Split(string(b), "\n") {
 		lhs, rhs, ok := strings.Cut(line, "=")
 		if !ok || strings.Trim(rhs, "'") != "wireguard" {
 			continue
 		}
 		parts := strings.Split(strings.TrimPrefix(lhs, "network."), ".")
 		if len(parts) == 2 && parts[1] == "proto" {
 			out = append(out, parts[0])
 		}
 	}
 	return out
 }
 // NextAddress is the lowest address in the tunnel's range that nothing is using. The
 // interface's own address is taken, as is every address any peer is allowed.
 func NextAddress(t Tunnel) (string, error) {
 	var base netip.Prefix
 	for _, a := range t.Addrs {
 		p, err := netip.ParsePrefix(a)
 		if err != nil || !p.Addr().Is4() {
 			continue
 		}
 		base = p
 		break
 	}
 	if !base.IsValid() {
 		return "", errors.New("this tunnel has no IPv4 range to allocate from")
 	}
 	taken := map[netip.Addr]bool{base.Addr(): true}
 	for _, p := range t.Peers {
 		for _, a := range p.AllowedIPs {
 			if ip, err := netip.ParsePrefix(strings.TrimSpace(a)); err == nil {
 				taken[ip.Addr()] = true
 			} else if ip, err := netip.ParseAddr(strings.TrimSpace(a)); err == nil {
 				taken[ip] = true
 			}
 		}
 	}
 	ip := base.Masked().Addr().Next() // .0 is the network
 	for base.Contains(ip) {
 		next := ip.Next()
 		if !base.Contains(next) { // the last address in the range is the broadcast
 			break
 		}
 		if !taken[ip] {
 			return ip.String(), nil
 		}
 		ip = next
 	}
 	return "", ErrNoRoom
 }
 // slug turns a device name into a section name, which may hold rather less.
 func slug(name string) string {
 	var b strings.Builder
 	for _, r := range strings.ToLower(name) {
 		switch {
 		case r >= 'a' && r <= 'z', r >= '0' && r <= '9':
 			b.WriteRune(r)
 		case r == ' ', r == '-', r == '.', r == '_':
 			b.WriteByte('_')
 		}
 	}
 	s := strings.Trim(b.String(), "_")
 	if s == "" {
 		s = "peer"
 	}
 	if len(s) > 24 {
 		s = s[:24]
 	}
 	return s
 }
 // genKeys makes a private key, its public half and a pre-shared key. The private key is
 // returned to be handed over once and is never written anywhere by this daemon.
 func genKeys(ctx context.Context) (priv, pub, psk string, err error) {
 	run := func(name string, stdin string, args ...string) (string, error) {
 		cmd := exec.CommandContext(ctx, name, args...)
 		if stdin != "" {
 			cmd.Stdin = strings.NewReader(stdin)
 		}
 		out, err := cmd.Output()
 		return strings.TrimSpace(string(out)), err
 	}
 	if priv, err = run("wg", "", "genkey"); err != nil {
 		return "", "", "", fmt.Errorf("generating a key: %w", err)
 	}
 	if pub, err = run("wg", priv+"\n", "pubkey"); err != nil {
 		return "", "", "", fmt.Errorf("deriving the public key: %w", err)
 	}
 	if psk, err = run("wg", "", "genpsk"); err != nil {
 		return "", "", "", fmt.Errorf("generating a pre-shared key: %w", err)
 	}
 	return priv, pub, psk, nil
 }
 // AddPeer creates a device: it writes the peer into the configuration, applies it to the
 // running tunnel, and returns the client profile once.
 func AddPeer(ctx context.Context, t Tunnel, n NewPeer) (*Created, error) {
 	if !NameRe.MatchString(strings.TrimSpace(n.Name)) {
 		return nil, errors.New("a device name is 1–31 letters, digits, spaces, dots, dashes or underscores")
 	}
 	name := strings.TrimSpace(n.Name)
 	if strings.TrimSpace(n.Endpoint) == "" {
 		return nil, errors.New("the device needs an endpoint to dial: the house's address or name, and the port")
 	}
 	endpoint, err := checkEndpoint(n.Endpoint)
 	if err != nil {
 		return nil, err
 	}
 	addr := strings.TrimSpace(n.Address)
 	if addr == "" {
 		if addr, err = NextAddress(t); err != nil {
 			return nil, err
 		}
 	}
 	host, err := hostOf(addr)
 	if err != nil {
 		return nil, err
 	}
 	for _, p := range t.Peers {
 		for _, a := range p.AllowedIPs {
 			if strings.HasPrefix(strings.TrimSpace(a), host+"/") {
 				return nil, fmt.Errorf("%s is already used by %s", host, firstNonEmpty(p.Name, "another device"))
 			}
 		}
 	}
 	section := sectionPrefix + slug(name)
 	existing, err := uciPeers(ctx)
 	if err != nil {
 		return nil, ErrUnmanaged
 	}
 	taken := map[string]bool{}
 	for _, p := range existing {
 		taken[p.Section] = true
 	}
 	for i := 2; taken[section]; i++ {
 		section = fmt.Sprintf("%s%s%d", sectionPrefix, slug(name), i)
 	}
 	if !sectionRe.MatchString(section) {
 		return nil, errors.New("that name cannot be turned into a configuration section")
 	}
 	priv, pub, psk, err := genKeys(ctx)
 	if err != nil {
 		return nil, err
 	}
 	allowed := host + "/32"
 	// the configuration first, so a peer can never be live but unrecorded
 	set := [][]string{
 		{"set", "network." + section + "=wireguard_" + t.Name},
 		{"set", "network." + section + ".description=" + name},
 		{"set", "network." + section + ".public_key=" + pub},
 		{"set", "network." + section + ".preshared_key=" + psk},
 		{"set", "network." + section + ".allowed_ips=" + allowed},
 	}
 	for _, args := range set {
 		if err := uci(ctx, args...); err != nil {
 			_ = uci(ctx, "revert", "network")
 			return nil, fmt.Errorf("writing the configuration: %w", err)
 		}
 	}
 	if err := uci(ctx, "commit", "network"); err != nil {
 		_ = uci(ctx, "revert", "network")
 		return nil, fmt.Errorf("saving the configuration: %w", err)
 	}
 	// then the running tunnel; if this fails the configuration goes back as it was
 	if err := wgSet(ctx, t.Name, pub, psk, allowed); err != nil {
 		_ = uci(ctx, "delete", "network."+section)
 		_ = uci(ctx, "commit", "network")
 		return nil, fmt.Errorf("applying it to the running tunnel: %w", err)
 	}
 	return &Created{
 		Section: section, Name: name, PublicKey: pub, Address: host,
 		Config: profile(priv, psk, t, host, endpoint, n.DNS, n.Split),
 	}, nil
 }
 // EditPeer changes a device's label or the address it is allowed. The section keeps the name
 // it was created with; the description is what is shown.
 func EditPeer(ctx context.Context, t Tunnel, section, name, address string) error {
 	if !sectionRe.MatchString(section) {
 		return errors.New("that is not a device this can edit")
 	}
 	peers, err := uciPeers(ctx)
 	if err != nil {
 		return ErrUnmanaged
 	}
 	var me *uciPeer
 	for i := range peers {
 		if peers[i].Section == section && peers[i].Iface == t.Name {
 			me = &peers[i]
 		}
 	}
 	if me == nil {
 		return errors.New("no such device on this tunnel")
 	}
 	if name = strings.TrimSpace(name); name != "" {
 		if !NameRe.MatchString(name) {
 			return errors.New("a device name is 1–31 letters, digits, spaces, dots, dashes or underscores")
 		}
 		if err := uci(ctx, "set", "network."+section+".description="+name); err != nil {
 			_ = uci(ctx, "revert", "network")
 			return err
 		}
 	}
 	allowed := ""
 	if address = strings.TrimSpace(address); address != "" {
 		host, err := hostOf(address)
 		if err != nil {
 			_ = uci(ctx, "revert", "network")
 			return err
 		}
 		for _, p := range t.Peers {
 			if p.Section == section {
 				continue
 			}
 			for _, a := range p.AllowedIPs {
 				if strings.HasPrefix(strings.TrimSpace(a), host+"/") {
 					_ = uci(ctx, "revert", "network")
 					return fmt.Errorf("%s is already used by %s", host, firstNonEmpty(p.Name, "another device"))
 				}
 			}
 		}
 		allowed = host + "/32"
 		if err := uci(ctx, "set", "network."+section+".allowed_ips="+allowed); err != nil {
 			_ = uci(ctx, "revert", "network")
 			return err
 		}
 	}
 	if err := uci(ctx, "commit", "network"); err != nil {
 		_ = uci(ctx, "revert", "network")
 		return err
 	}
 	if allowed != "" { // the running tunnel has to be told, or the device keeps the old address
 		if err := wgSet(ctx, t.Name, me.PublicKey, "", allowed); err != nil {
 			return fmt.Errorf("saved, but the running tunnel refused the new address: %w", err)
 		}
 	}
 	return nil
 }
 // DeletePeer removes a device. The running tunnel goes first, because that is the moment the
 // device actually loses its way in.
 func DeletePeer(ctx context.Context, t Tunnel, section string) error {
 	if !sectionRe.MatchString(section) {
 		return errors.New("that is not a device this can remove")
 	}
 	peers, err := uciPeers(ctx)
 	if err != nil {
 		return ErrUnmanaged
 	}
 	var me *uciPeer
 	for i := range peers {
 		if peers[i].Section == section && peers[i].Iface == t.Name {
 			me = &peers[i]
 		}
 	}
 	if me == nil {
 		return errors.New("no such device on this tunnel")
 	}
 	if err := wgRemove(ctx, t.Name, me.PublicKey); err != nil {
 		return fmt.Errorf("removing it from the running tunnel: %w", err)
 	}
 	if err := uci(ctx, "delete", "network."+section); err != nil {
 		return fmt.Errorf("the device is disconnected, but removing it from the configuration failed, so it would come back on a reboot: %w", err)
 	}
 	if err := uci(ctx, "commit", "network"); err != nil {
 		return fmt.Errorf("the device is disconnected, but saving the configuration failed, so it would come back on a reboot: %w", err)
 	}
 	return nil
 }
 // ---- the small dangerous bits, each in one place ----
 func uci(ctx context.Context, args ...string) error {
 	out, err := exec.CommandContext(ctx, "uci", args...).CombinedOutput()
 	if err != nil {
 		return fmt.Errorf("uci %s: %s", strings.Join(args, " "), strings.TrimSpace(string(out)))
 	}
 	return nil
 }
 // wgSet applies a peer to the running interface. The pre-shared key goes in through a file
 // because that is the only way wg takes one, so the file is made readable by nobody else and
 // removed immediately.
 func wgSet(ctx context.Context, iface, pub, psk, allowed string) error {
 	args := []string{"set", iface, "peer", pub}
 	if psk != "" {
 		f, err := os.CreateTemp("", "hamlab-psk")
 		if err != nil {
 			return err
 		}
 		defer os.Remove(f.Name())
 		if err := f.Chmod(0o600); err != nil {
 			f.Close()
 			return err
 		}
 		if _, err := f.WriteString(psk + "\n"); err != nil {
 			f.Close()
 			return err
 		}
 		f.Close()
 		args = append(args, "preshared-key", f.Name())
 	}
 	args = append(args, "allowed-ips", allowed)
 	out, err := exec.CommandContext(ctx, "wg", args...).CombinedOutput()
 	if err != nil {
 		return fmt.Errorf("wg set: %s", strings.TrimSpace(string(out)))
 	}
 	return nil
 }
 func wgRemove(ctx context.Context, iface, pub string) error {
 	out, err := exec.CommandContext(ctx, "wg", "set", iface, "peer", pub, "remove").CombinedOutput()
 	if err != nil {
 		return fmt.Errorf("wg set: %s", strings.TrimSpace(string(out)))
 	}
 	return nil
 }
 // profile is the client's own configuration file, which is the only place its private key
 // ever appears.
 func profile(priv, psk string, t Tunnel, addr, endpoint, dns string, split bool) string {
 	prefix := 24
 	for _, a := range t.Addrs {
 		if p, err := netip.ParsePrefix(a); err == nil && p.Addr().Is4() {
 			prefix = p.Bits()
 			break
 		}
 	}
 	allowed := "0.0.0.0/0"
 	if split {
 		nets := []string{}
 		for _, a := range t.Addrs {
 			if p, err := netip.ParsePrefix(a); err == nil && p.Addr().Is4() {
 				nets = append(nets, p.Masked().String())
 			}
 		}
 		nets = append(nets, "10.0.0.0/8", "192.168.0.0/16", "172.16.0.0/12")
 		allowed = strings.Join(dedupe(nets), ", ")
 	}
 	var b strings.Builder
 	fmt.Fprintf(&b, "[Interface]\nPrivateKey = %s\nAddress = %s/%d\n", priv, addr, prefix)
 	if dns = strings.TrimSpace(dns); dns != "" {
 		fmt.Fprintf(&b, "DNS = %s\n", dns)
 	}
 	fmt.Fprintf(&b, "\n[Peer]\nPublicKey = %s\n", t.PublicKey)
 	if psk != "" {
 		fmt.Fprintf(&b, "PresharedKey = %s\n", psk)
 	}
 	fmt.Fprintf(&b, "Endpoint = %s\nAllowedIPs = %s\n", endpoint, allowed)
 	return b.String()
 }
 func dedupe(in []string) []string {
 	seen := map[string]bool{}
 	out := in[:0:0]
 	for _, s := range in {
 		if !seen[s] {
 			seen[s] = true
 			out = append(out, s)
 		}
 	}
 	sort.Strings(out)
 	return out
 }
 // hostOf takes an address with or without a prefix and returns the bare address.
 func hostOf(s string) (string, error) {
 	s = strings.TrimSpace(s)
 	if p, err := netip.ParsePrefix(s); err == nil {
 		return p.Addr().String(), nil
 	}
 	a, err := netip.ParseAddr(s)
 	if err != nil {
 		return "", fmt.Errorf("%q is not an address", s)
 	}
 	return a.String(), nil
 }
 // checkEndpoint makes sure what the device will dial looks like a host and a port.
 func checkEndpoint(s string) (string, error) {
 	s = strings.TrimSpace(s)
 	host, port, err := net.SplitHostPort(s)
 	if err != nil || host == "" {
 		return "", fmt.Errorf("%q should be a name or address and a port, like home.example.com:51820", s)
 	}
 	if n, err := strconv.Atoi(port); err != nil || n < 1 || n > 65535 {
 		return "", fmt.Errorf("%q is not a port", port)
 	}
 	return s, nil
 }
 func firstNonEmpty(a, b string) string {
 	if a != "" {
 		return a
 	}
 	return b
 }
@@ -0,0 +1,126 @@
 package vpn
 import (
 	"strings"
 	"testing"
 )
 func tunnel(addrs []string, peers ...string) Tunnel {
 	t := Tunnel{Name: "wg0", PublicKey: "SERVERKEY=", Addrs: addrs}
 	for _, p := range peers {
 		t.Peers = append(t.Peers, Peer{PublicKey: p + "KEY=", AllowedIPs: []string{p}})
 	}
 	return t
 }
 func TestNextAddress(t *testing.T) {
 	// .1 is the interface itself and .2 and .3 are taken, so the next free one is .4
 	got, err := NextAddress(tunnel([]string{"172.21.46.1/24"}, "172.21.46.3/32", "172.21.46.2/32"))
 	if err != nil || got != "172.21.46.4" {
 		t.Errorf("NextAddress = %q, %v; want 172.21.46.4", got, err)
 	}
 	// a gap is filled before the end is reached
 	got, _ = NextAddress(tunnel([]string{"172.21.46.1/24"}, "172.21.46.2/32", "172.21.46.4/32"))
 	if got != "172.21.46.3" {
 		t.Errorf("NextAddress = %q, want the gap at .3", got)
 	}
 	// an empty tunnel starts after the server
 	got, _ = NextAddress(tunnel([]string{"10.9.0.1/24"}))
 	if got != "10.9.0.2" {
 		t.Errorf("NextAddress = %q, want 10.9.0.2", got)
 	}
 	// a full one says so rather than handing out a duplicate
 	full := tunnel([]string{"10.9.0.1/30"}, "10.9.0.2/32")
 	if _, err := NextAddress(full); err != ErrNoRoom {
 		t.Errorf("err = %v, want ErrNoRoom on a full range", err)
 	}
 	// IPv6-only, or nothing at all, is refused rather than guessed
 	if _, err := NextAddress(tunnel([]string{"fc00::1/64"})); err == nil {
 		t.Error("want an error when there is no IPv4 range")
 	}
 }
 func TestSlug(t *testing.T) {
 	for in, want := range map[string]string{
 		"ip12":                  "ip12",
 		"Mitchell's iPad":       "mitchells_ipad",
 		"my-phone":              "my_phone",
 		"  spaced  out  ":       "spaced__out",
 		"!!!":                   "peer",
 		strings.Repeat("a", 40): strings.Repeat("a", 24),
 	} {
 		if got := slug(in); got != want {
 			t.Errorf("slug(%q) = %q, want %q", in, got, want)
 		}
 	}
 }
 func TestProfile(t *testing.T) {
 	tun := tunnel([]string{"172.21.46.1/24"})
 	got := profile("PRIV=", "PSK=", tun, "172.21.46.4", "home.example.com:51820", "1.1.1.1, 1.0.0.1", false)
 	for _, want := range []string{
 		"PrivateKey = PRIV=",
 		"Address = 172.21.46.4/24", // the interface's own prefix, not a bare /32
 		"DNS = 1.1.1.1, 1.0.0.1",
 		"PublicKey = SERVERKEY=",
 		"PresharedKey = PSK=",
 		"Endpoint = home.example.com:51820",
 		"AllowedIPs = 0.0.0.0/0",
 	} {
 		if !strings.Contains(got, want) {
 			t.Errorf("profile is missing %q:\n%s", want, got)
 		}
 	}
 	// no DNS line at all when none was asked for, rather than an empty one
 	if strings.Contains(profile("P=", "", tun, "172.21.46.4", "h:1", "", false), "DNS") {
 		t.Error("an empty DNS should leave the line out")
 	}
 	// split routes the private ranges instead of everything
 	split := profile("P=", "", tun, "172.21.46.4", "h:1", "", true)
 	if strings.Contains(split, "0.0.0.0/0") {
 		t.Errorf("a split tunnel must not route everything:\n%s", split)
 	}
 	if !strings.Contains(split, "172.21.46.0/24") {
 		t.Errorf("a split tunnel should include the tunnel's own range:\n%s", split)
 	}
 }
 func TestCheckEndpoint(t *testing.T) {
 	for _, ok := range []string{"bigham.dev:51820", "1.2.3.4:51820", "[2001:db8::1]:51820"} {
 		if _, err := checkEndpoint(ok); err != nil {
 			t.Errorf("checkEndpoint(%q) = %v, want it accepted", ok, err)
 		}
 	}
 	for _, bad := range []string{"", "bigham.dev", ":51820", "bigham.dev:0", "bigham.dev:notaport", "bigham.dev:99999"} {
 		if _, err := checkEndpoint(bad); err == nil {
 			t.Errorf("checkEndpoint(%q) was accepted, want refused", bad)
 		}
 	}
 }
 func TestHostOf(t *testing.T) {
 	for in, want := range map[string]string{"172.21.46.4": "172.21.46.4", "172.21.46.4/32": "172.21.46.4", " 10.0.0.9/24 ": "10.0.0.9"} {
 		got, err := hostOf(in)
 		if err != nil || got != want {
 			t.Errorf("hostOf(%q) = %q, %v; want %q", in, got, err, want)
 		}
 	}
 	if _, err := hostOf("not-an-address"); err == nil {
 		t.Error("hostOf accepted nonsense")
 	}
 }
 // The name is what becomes a configuration section, so it is the input worth being strict
 // about; everything refused here is refused before anything is written.
 func TestNameRe(t *testing.T) {
 	for _, ok := range []string{"ip12", "My Phone", "a", "work-laptop.2"} {
 		if !NameRe.MatchString(ok) {
 			t.Errorf("%q should be a valid name", ok)
 		}
 	}
 	for _, bad := range []string{"", " leading", "semi;colon", "quote'd", "back`tick", "new\nline", "$(whoami)", strings.Repeat("a", 32)} {
 		if NameRe.MatchString(bad) {
 			t.Errorf("%q should be refused as a name", bad)
 		}
 	}
 }
@@ -0,0 +1,372 @@
 // Package vpn reads this machine's WireGuard tunnels: the interfaces, and the peers on them
 // with names put back on.
 //
 // The live state comes from `wg show all dump`, which is the kernel's own answer and the only
 // place handshakes and byte counts exist. The kernel knows nothing about names, though — a
 // peer there is a public key — so the configuration is read alongside it purely to label
 // them: `# comments` in a wg-quick .conf, or the `description` of a UCI peer section on
 // OpenWrt. Everything here is read-only, and no private or pre-shared key is ever read, let
 // alone returned.
 package vpn
 import (
 	"bufio"
 	"context"
 	"net"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"sort"
 	"strconv"
 	"strings"
 	"time"
 	"hamlab/internal/priv"
 )
 // Peer is one end of a tunnel: a device that dials in, or the server this machine dials out
 // to. Which it is depends on the tunnel; see Tunnel.Kind.
 type Peer struct {
 	// Name is what the configuration calls it, where it says at all. The kernel has no idea.
 	Name      string `json:"name,omitempty"`
 	PublicKey string `json:"public_key"`
 	// Endpoint is where its packets come from, empty until it has been heard from.
 	Endpoint   string   `json:"endpoint,omitempty"`
 	AllowedIPs []string `json:"allowed_ips"`
 	// Handshake is the last completed handshake, absent when there has never been one — a peer
 	// that is configured but has not connected since this tunnel came up. A pointer because
 	// omitempty does nothing for a zero time.Time, and a peer that has never been heard from
 	// would otherwise be reported as having last spoken in the year one.
 	Handshake *time.Time `json:"handshake,omitempty"`
 	// Online is a handshake inside the last three minutes. WireGuard is silent when idle, so
 	// this means "was talking recently", not "is connected" — there is no such thing here.
 	Online    bool  `json:"online"`
 	Rx        int64 `json:"rx"`
 	Tx        int64 `json:"tx"`
 	Keepalive int   `json:"keepalive,omitempty"`
 	// PSK notes that a pre-shared key is in use, without the key itself going anywhere.
 	PSK bool `json:"psk,omitempty"`
 	// Section is the configuration section holding this peer, and the identifier anything that
 	// edits it uses. Empty when the peer is not in a configuration this can edit.
 	Section string `json:"section,omitempty"`
 }
 // Tunnel is one WireGuard interface.
 type Tunnel struct {
 	Name string `json:"name"`
 	// Managed is whether devices on this tunnel can be added and removed from here, which
 	// today means its peers are configured through OpenWrt's UCI.
 	Managed   bool   `json:"managed"`
 	PublicKey string `json:"public_key"`
 	// Kind is "server" when peers dial in to this machine, "client" when this machine dials
 	// out through a peer, and "" when it is too new or too odd to tell. It is worked out from
 	// the shape of the peers, not from anything WireGuard states: a peer that is routed
 	// everything and has an endpoint of its own is an exit, and a tunnel with one of those is
 	// this machine going out through it.
 	Kind       string   `json:"kind"`
 	ListenPort int      `json:"listen_port,omitempty"`
 	FwMark     string   `json:"fwmark,omitempty"`
 	Addrs      []string `json:"addrs"`
 	MTU        int      `json:"mtu,omitempty"`
 	Peers      []Peer   `json:"peers"`
 	// Exit is the peer this machine goes out through, for a client tunnel.
 	Exit string `json:"exit,omitempty"`
 }
 // Report is every tunnel on this machine, and why there are none when there are none.
 type Report struct {
 	At      time.Time `json:"at"`
 	Tunnels []Tunnel  `json:"tunnels"`
 	// Tool is false when wireguard-tools is not installed, Root false when the daemon cannot
 	// reach root — either one means the peers cannot be read, and the tab should say which.
 	Tool bool   `json:"tool"`
 	Root bool   `json:"root"`
 	Note string `json:"note,omitempty"`
 }
 const onlineWithin = 3 * time.Minute
 // Look reads the tunnels on this machine.
 func Look(ctx context.Context) Report {
 	r := Report{At: time.Now(), Tunnels: []Tunnel{}, Root: priv.Available()}
 	if _, err := exec.LookPath("wg"); err != nil {
 		r.Note = "wireguard-tools is not installed here, so there is nothing to read"
 		return r
 	}
 	r.Tool = true
 	args := []string{"wg", "show", "all", "dump"}
 	if priv.Available() {
 		args = priv.Wrap(args...)
 	}
 	cmd := exec.CommandContext(ctx, args[0], args[1:]...)
 	out, err := cmd.Output()
 	if err != nil {
 		if !r.Root {
 			r.Note = "reading WireGuard needs root, and this daemon has no way to it"
 		} else {
 			r.Note = "wg show failed: " + strings.TrimSpace(err.Error())
 		}
 		return r
 	}
 	r.Tunnels = parseDump(string(out))
 	names := peerNames()
 	for i := range r.Tunnels {
 		t := &r.Tunnels[i]
 		t.Addrs, t.MTU = ifaceAddrs(t.Name)
 		for j := range t.Peers {
 			p := &t.Peers[j]
 			p.Name = names[p.PublicKey]
 			p.Online = p.Handshake != nil && time.Since(*p.Handshake) < onlineWithin
 		}
 		t.Kind, t.Exit = classify(*t)
 		t.Managed = Managed(ctx, t)
 		sort.SliceStable(t.Peers, func(a, b int) bool {
 			pa, pb := t.Peers[a], t.Peers[b]
 			if pa.Online != pb.Online {
 				return pa.Online
 			}
 			if ha, hb := shookAt(pa), shookAt(pb); !ha.Equal(hb) {
 				return ha.After(hb)
 			}
 			return pa.Name < pb.Name
 		})
 	}
 	if len(r.Tunnels) == 0 {
 		r.Note = "no WireGuard interfaces are up on this machine"
 	}
 	return r
 }
 // parseDump reads `wg show all dump`. The first line of each interface is the interface
 // itself, the rest are its peers; every line is tab-separated and positional.
 //
 //	interface: name  private-key  public-key  listen-port  fwmark
 //	peer:      name  public-key   psk         endpoint     allowed-ips  handshake  rx  tx  keepalive
 //
 // The private key is in the second column and is read past without being kept.
 func parseDump(out string) []Tunnel {
 	// An empty list, never nil: this is JSON to a browser that will take its length, and a
 	// null there is a blank page rather than "no tunnels".
 	tunnels := []Tunnel{}
 	byName := map[string]int{}
 	sc := bufio.NewScanner(strings.NewReader(out))
 	sc.Buffer(make([]byte, 0, 64<<10), 1<<20)
 	for sc.Scan() {
 		f := strings.Split(sc.Text(), "\t")
 		if len(f) < 5 {
 			continue
 		}
 		name := f[0]
 		if len(f) == 5 { // an interface line
 			port, _ := strconv.Atoi(f[3])
 			byName[name] = len(tunnels)
 			tunnels = append(tunnels, Tunnel{Name: name, PublicKey: f[2], ListenPort: port, FwMark: off(f[4]), Peers: []Peer{}})
 			continue
 		}
 		if len(f) < 9 {
 			continue
 		}
 		i, ok := byName[name]
 		if !ok {
 			continue
 		}
 		hs, _ := strconv.ParseInt(f[5], 10, 64)
 		rx, _ := strconv.ParseInt(f[6], 10, 64)
 		tx, _ := strconv.ParseInt(f[7], 10, 64)
 		ka, _ := strconv.Atoi(f[8])
 		p := Peer{PublicKey: f[1], PSK: f[2] != "(none)", Endpoint: off(f[3]), Rx: rx, Tx: tx, Keepalive: ka}
 		if f[4] != "(none)" {
 			p.AllowedIPs = strings.Split(f[4], ",")
 		}
 		if hs > 0 {
 			at := time.Unix(hs, 0)
 			p.Handshake = &at
 		}
 		tunnels[i].Peers = append(tunnels[i].Peers, p)
 	}
 	return tunnels
 }
 // shookAt is a peer's last handshake, or the zero time when there has never been one.
 func shookAt(p Peer) time.Time {
 	if p.Handshake == nil {
 		return time.Time{}
 	}
 	return *p.Handshake
 }
 // off turns wg's placeholders into an empty string.
 func off(v string) string {
 	if v == "(none)" || v == "off" {
 		return ""
 	}
 	return v
 }
 // classify decides whether this machine is the one being dialled or the one dialling. A peer
 // that is allowed every address and has an endpoint is a way out to the internet; a tunnel
 // with one of those is this machine using it. Otherwise, if anything can reach us here, it is
 // a server.
 func classify(t Tunnel) (kind, exit string) {
 	for _, p := range t.Peers {
 		for _, a := range p.AllowedIPs {
 			if a == "0.0.0.0/0" || a == "::/0" {
 				if p.Endpoint != "" {
 					name := p.Name
 					if name == "" {
 						name = p.Endpoint
 					}
 					return "client", name
 				}
 			}
 		}
 	}
 	if t.ListenPort > 0 && len(t.Peers) > 0 {
 		return "server", ""
 	}
 	return "", ""
 }
 // ifaceAddrs is what the interface carries, and its MTU.
 func ifaceAddrs(name string) ([]string, int) {
 	addrs := []string{}
 	mtu := 0
 	ifi, err := net.InterfaceByName(name)
 	if err != nil {
 		return addrs, mtu
 	}
 	mtu = ifi.MTU
 	got, err := ifi.Addrs()
 	if err != nil {
 		return addrs, mtu
 	}
 	for _, a := range got {
 		addrs = append(addrs, a.String())
 	}
 	return addrs, mtu
 }
 // ---- names ----
 // peerNames maps a public key to what the configuration calls that peer. Two shapes are
 // understood, which between them cover every machine here: wg-quick's .conf files, where the
 // name is a comment, and OpenWrt's UCI, where it is the peer section's description.
 func peerNames() map[string]string {
 	out := map[string]string{}
 	for k, v := range confNames() {
 		out[k] = v
 	}
 	for k, v := range uciNames() { // UCI wins: on OpenWrt it is the live configuration
 		out[k] = v
 	}
 	return out
 }
 // confNames reads /etc/wireguard/*.conf. In a wg-quick file a peer is named by the comment
 // that introduces it, which is the convention every tool that writes these files follows:
 //
 //	# phone
 //	[Peer]
 //	PublicKey = …
 //
 // confDir is where wg-quick keeps its files; a variable so the test can point it elsewhere.
 var confDir = "/etc/wireguard"
 func confNames() map[string]string {
 	out := map[string]string{}
 	files, _ := filepath.Glob(filepath.Join(confDir, "*.conf"))
 	for _, path := range files {
 		text, err := readMaybeRoot(path)
 		if err != nil {
 			continue
 		}
 		sc := bufio.NewScanner(strings.NewReader(text))
 		// last is the most recent comment line; when a [Peer] block opens, that comment becomes
 		// its pending name. Keeping the two apart is what lets each block take the comment that
 		// introduces it and ignore any comment written inside it.
 		last, pending := "", ""
 		inPeer := false
 		for sc.Scan() {
 			line := strings.TrimSpace(sc.Text())
 			switch {
 			case strings.EqualFold(line, "[Peer]"):
 				inPeer, pending, last = true, last, ""
 			case strings.HasPrefix(line, "["):
 				inPeer, pending, last = false, "", ""
 			case strings.HasPrefix(line, "#"):
 				if c := strings.TrimSpace(strings.TrimLeft(line, "#")); c != "" {
 					last = c
 				}
 			case inPeer && strings.HasPrefix(line, "PublicKey"):
 				// The separator is the first '='; the key's own base64 padding is the rest of
 				// the line, so everything after that first one is the key.
 				if _, v, ok := strings.Cut(line, "="); ok && pending != "" {
 					out[strings.TrimSpace(v)] = pending
 				}
 				pending = ""
 			}
 		}
 	}
 	return out
 }
 // readMaybeRoot reads a file, falling back to root when it is not ours to read — which
 // /etc/wireguard always is, since it holds private keys. Only the comments and public keys in
 // it are looked at; nothing else in the file is kept or returned.
 func readMaybeRoot(path string) (string, error) {
 	b, err := os.ReadFile(path)
 	if err == nil {
 		return string(b), nil
 	}
 	if !os.IsPermission(err) || !priv.Available() {
 		return "", err
 	}
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()
 	args := priv.Wrap("cat", path)
 	out, cerr := exec.CommandContext(ctx, args[0], args[1:]...).Output()
 	if cerr != nil {
 		return "", cerr
 	}
 	return string(out), nil
 }
 // uciNames reads OpenWrt's configuration: each peer is a section with a description and a
 // public key. `uci show` is used rather than the file, so includes and overrides are applied.
 func uciNames() map[string]string {
 	out := map[string]string{}
 	if _, err := exec.LookPath("uci"); err != nil {
 		return out
 	}
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()
 	b, err := exec.CommandContext(ctx, "uci", "-q", "show", "network").Output()
 	if err != nil {
 		return out
 	}
 	desc, key := map[string]string{}, map[string]string{}
 	for _, line := range strings.Split(string(b), "\n") {
 		lhs, rhs, ok := strings.Cut(line, "=")
 		if !ok || !strings.HasPrefix(lhs, "network.") {
 			continue
 		}
 		parts := strings.Split(strings.TrimPrefix(lhs, "network."), ".")
 		if len(parts) != 2 {
 			continue
 		}
 		section, option := parts[0], parts[1]
 		v := strings.Trim(rhs, "'")
 		switch option {
 		case "description":
 			desc[section] = v
 		case "public_key":
 			key[section] = v
 		}
 	}
 	for section, k := range key {
 		if d := desc[section]; d != "" {
 			out[k] = d
 		}
 	}
 	return out
 }
@@ -0,0 +1,193 @@
 package vpn
 import (
 	"encoding/json"
 	"os"
 	"path/filepath"
 	"strings"
 	"testing"
 	"time"
 )
 // Real output of `wg show all dump`, from the two machines here: gate, where devices dial in,
 // and hamtv, which dials out through Mullvad. Keys are the real public ones, which are public.
 const gateDump = "wg0\tPRIVATEKEYWOULDBEHERE=\t4VwqgzHyelmVmNY+EANRWQggSSxvYYguJQMvUty9Sns=\t51820\toff\n" +
 	"wg0\tNeWxo9wfqgipBQLuP676WFcMlW5RDrOcanc9y8sT7kU=\t3KWsu0mdVb2l5B/j0qHE03yE9LXi3FJ1WAjWMuxaLAw=\t10.0.0.33:63573\t172.21.46.3/32\t1790565744\t134192164\t415977404\toff\n" +
 	"wg0\twUnRD990c6aglU0TCREtmfuUTPzqSmHrd3qTORpRJBg=\t4Pd5u2t6U9EdHlFlneg+hMr0G2v+IvoE/EqGfFdDPI8=\t(none)\t172.21.46.2/32\t0\t0\t0\toff\n"
 const hamtvDump = "wg0\tPRIVATEKEYWOULDBEHERE=\tF/rgcHi+ZiaWPoSQ9iPcldIUbbMotE1hwSB/74Y9Qi4=\t39281\toff\n" +
 	"wg0\tzqsfGglzJPY657WMRxf/S4omG7+ZkSDIpDq+ggbc9yo=\t(none)\t23.234.72.2:51820\t0.0.0.0/0,::/0\t1790571246\t18206675520\t2687837240\t25\n"
 func TestParseServerDump(t *testing.T) {
 	ts := parseDump(gateDump)
 	if len(ts) != 1 {
 		t.Fatalf("got %d tunnels, want 1", len(ts))
 	}
 	w := ts[0]
 	if w.Name != "wg0" || w.ListenPort != 51820 {
 		t.Errorf("interface = %s:%d, want wg0:51820", w.Name, w.ListenPort)
 	}
 	if w.PublicKey != "4VwqgzHyelmVmNY+EANRWQggSSxvYYguJQMvUty9Sns=" {
 		t.Errorf("public key = %q; the third column is the public one", w.PublicKey)
 	}
 	// The private key is column two of the interface line and must not appear anywhere.
 	if w.PublicKey == "PRIVATEKEYWOULDBEHERE=" || w.FwMark == "PRIVATEKEYWOULDBEHERE=" {
 		t.Error("the private key was kept")
 	}
 	if len(w.Peers) != 2 {
 		t.Fatalf("got %d peers, want 2", len(w.Peers))
 	}
 	p := w.Peers[0]
 	if p.PublicKey != "NeWxo9wfqgipBQLuP676WFcMlW5RDrOcanc9y8sT7kU=" || p.Endpoint != "10.0.0.33:63573" {
 		t.Errorf("first peer = %+v", p)
 	}
 	if !p.PSK {
 		t.Error("first peer has a pre-shared key and should say so")
 	}
 	if got := p.AllowedIPs; len(got) != 1 || got[0] != "172.21.46.3/32" {
 		t.Errorf("allowed ips = %v", got)
 	}
 	if p.Rx != 134192164 || p.Tx != 415977404 {
 		t.Errorf("counters = %d/%d", p.Rx, p.Tx)
 	}
 	if p.Handshake == nil || p.Handshake.Unix() != 1790565744 {
 		t.Errorf("handshake = %v", p.Handshake)
 	}
 	// The second peer has never connected: no endpoint, no handshake, nothing moved.
 	if q := w.Peers[1]; q.Endpoint != "" || q.Handshake != nil || q.Rx != 0 {
 		t.Errorf("second peer = %+v, want one that has never been heard from", q)
 	}
 	if kind, exit := classify(w); kind != "server" || exit != "" {
 		t.Errorf("classify = %q/%q, want server", kind, exit)
 	}
 }
 func TestParseClientDump(t *testing.T) {
 	ts := parseDump(hamtvDump)
 	if len(ts) != 1 || len(ts[0].Peers) != 1 {
 		t.Fatalf("got %d tunnels", len(ts))
 	}
 	w := ts[0]
 	p := w.Peers[0]
 	if len(p.AllowedIPs) != 2 || p.AllowedIPs[0] != "0.0.0.0/0" {
 		t.Errorf("allowed ips = %v, want everything", p.AllowedIPs)
 	}
 	if p.Keepalive != 25 {
 		t.Errorf("keepalive = %d, want 25", p.Keepalive)
 	}
 	if p.PSK {
 		t.Error("this peer has no pre-shared key")
 	}
 	// It routes everything and has an endpoint, so this machine goes out through it.
 	kind, exit := classify(w)
 	if kind != "client" || exit != "23.234.72.2:51820" {
 		t.Errorf("classify = %q/%q, want client through the endpoint", kind, exit)
 	}
 }
 // A tunnel that is up but has no peers at all is neither, and must not be guessed at.
 func TestClassifyUndecided(t *testing.T) {
 	if kind, _ := classify(Tunnel{Name: "wg0", ListenPort: 51820}); kind != "" {
 		t.Errorf("classify with no peers = %q, want no opinion", kind)
 	}
 }
 // Peers are named by the comment above them in a wg-quick file. The public key's own base64
 // padding is an '=' too, which is the thing to get wrong.
 func TestConfNames(t *testing.T) {
 	dir := t.TempDir()
 	conf := `[Interface]
 Address = 172.21.46.1/24
 PrivateKey = SECRETSECRETSECRETSECRETSECRETSECRETSECRET=
 ListenPort = 51820
 # phone
 [Peer]
 PublicKey = NeWxo9wfqgipBQLuP676WFcMlW5RDrOcanc9y8sT7kU=
 AllowedIPs = 172.21.46.3/32
 # the laptop
 [Peer]
 PublicKey = wUnRD990c6aglU0TCREtmfuUTPzqSmHrd3qTORpRJBg=
 AllowedIPs = 172.21.46.2/32
 `
 	if err := os.WriteFile(filepath.Join(dir, "wg0.conf"), []byte(conf), 0o600); err != nil {
 		t.Fatal(err)
 	}
 	old := confDir
 	confDir = dir
 	defer func() { confDir = old }()
 	names := confNames()
 	if got := names["NeWxo9wfqgipBQLuP676WFcMlW5RDrOcanc9y8sT7kU="]; got != "phone" {
 		t.Errorf("first peer named %q, want phone", got)
 	}
 	if got := names["wUnRD990c6aglU0TCREtmfuUTPzqSmHrd3qTORpRJBg="]; got != "the laptop" {
 		t.Errorf("second peer named %q, want the laptop", got)
 	}
 	for k := range names {
 		if k == "SECRETSECRETSECRETSECRETSECRETSECRETSECRET=" {
 			t.Error("the private key was read as if it were a peer")
 		}
 	}
 	if len(names) != 2 {
 		t.Errorf("%d names, want 2: %v", len(names), names)
 	}
 }
 func TestOnlineIsRecentHandshake(t *testing.T) {
 	for _, tc := range []struct {
 		name string
 		hs   time.Time
 		want bool
 	}{
 		{"just now", time.Now().Add(-30 * time.Second), true},
 		{"a while back", time.Now().Add(-10 * time.Minute), false},
 		{"never", time.Time{}, false},
 	} {
 		got := !tc.hs.IsZero() && time.Since(tc.hs) < onlineWithin
 		if got != tc.want {
 			t.Errorf("%s: online = %v, want %v", tc.name, got, tc.want)
 		}
 	}
 }
 // A peer that has never handshaken must carry no time at all, not the zero one. omitempty
 // does nothing for a time.Time, so this is the difference between the page saying "never
 // connected" and saying the device was last seen in the year one.
 func TestNeverConnectedHasNoTimestamp(t *testing.T) {
 	w := parseDump(gateDump)[0]
 	b, err := json.Marshal(w.Peers[1])
 	if err != nil {
 		t.Fatal(err)
 	}
 	if strings.Contains(string(b), "handshake") {
 		t.Errorf("json = %s, want no handshake field at all", b)
 	}
 	if strings.Contains(string(b), "0001-01-01") {
 		t.Errorf("json = %s, want no year-one timestamp", b)
 	}
 	// and the one that has connected still carries its time
 	if b, _ := json.Marshal(w.Peers[0]); !strings.Contains(string(b), "handshake") {
 		t.Errorf("json = %s, want the handshake of a peer that has one", b)
 	}
 }
 // A machine with wireguard-tools installed but no tunnel up produces an empty dump. That has
 // to marshal as [] and not as null: the page takes the length of it, and null blanks the
 // whole console rather than saying "no tunnels".
 func TestEmptyDumpIsAnEmptyList(t *testing.T) {
 	for _, dump := range []string{"", "\n", "   \n\n"} {
 		got := parseDump(dump)
 		if got == nil {
 			t.Fatalf("parseDump(%q) = nil, want an empty list", dump)
 		}
 		b, err := json.Marshal(got)
 		if err != nil {
 			t.Fatal(err)
 		}
 		if string(b) != "[]" {
 			t.Errorf("parseDump(%q) marshals to %s, want []", dump, b)
 		}
 	}
 }
@@ -0,0 +1,35 @@
 #!/bin/sh
 # Trim the PCI ID database to the graphics devices of the GPU vendors, for embedding in
 # the binary (internal/metrics/pciids_gpu.txt): a machine without pciutils — OpenWrt has
 # neither lspci nor pci.ids — then names its GPU exactly as every other machine does.
 #
 #   scripts/pci-ids-gpu.sh [/path/to/pci.ids]     default: /usr/share/misc/pci.ids
 #
 # Kept: the vendor lines and device lines (never subsystems) of Intel 8086, AMD 1002 and
 # NVIDIA 10de whose name is bracketed (how pci.ids writes a marketing name) or mentions
 # graphics, and everything of the three virtual-GPU vendors (virtio 1af4, VMware 15ad,
 # Bochs 1234). A non-GPU device slipping through costs a line; it is only ever consulted
 # for a card the kernel already drives as a GPU.
 set -eu
 src=${1:-/usr/share/misc/pci.ids}
 out=$(dirname "$0")/../internal/metrics/pciids_gpu.txt
 {
  echo "# Trimmed from pci.ids ($(grep -m1 -i 'version:' "$src" | sed 's/^#[[:space:]]*//')) by scripts/pci-ids-gpu.sh."
  echo "# Vendor lines and the graphics device lines of Intel, AMD and NVIDIA, plus the virtual GPUs."
  echo "# pci.ids is distributed under the 3-clause BSD license (https://pci-ids.ucw.cz/)."
  awk '
    /^[0-9a-f][0-9a-f][0-9a-f][0-9a-f]  / {
      vendor = substr($0, 1, 4)
      keep = (vendor == "8086" || vendor == "1002" || vendor == "10de")
      all = (vendor == "1af4" || vendor == "15ad" || vendor == "1234")
      if (keep || all) print
      next
    }
    /^\t[0-9a-f][0-9a-f][0-9a-f][0-9a-f]  / {
      if (all) { print; next }
      if (!keep) next
      if ($0 ~ /\[/ || $0 ~ /Graphics|Radeon|GeForce|Quadro|Tesla|Arc |Iris|Xe |GPU|Display|VGA|FirePro|Instinct|Data Center/) print
    }
  ' "$src"
 } > "$out"
 echo "$out: $(wc -l < "$out") lines, $(wc -c < "$out") bytes"
@@ -11,6 +11,7 @@
        "@xterm/addon-fit": "^0.10.0",
        "@xterm/addon-web-links": "^0.11.0",
        "@xterm/xterm": "^5.5.0",
        "qrcode-generator": "^2.0.4",
        "react": "^18.3.1",
        "react-dom": "^18.3.1"
      },
@@ -1551,6 +1552,12 @@
        "node": "^10 || ^12 || >=14"
      }
    },
    "node_modules/qrcode-generator": {
      "version": "2.0.4",
      "resolved": "https://registry.npmjs.org/qrcode-generator/-/qrcode-generator-2.0.4.tgz",
      "integrity": "sha512-mZSiP6RnbHl4xL2Ap5HfkjLnmxfKcPWpWe/c+5XxCuetEenqmNFf1FH/ftXPCtFG5/TDobjsjz6sSNL0Sr8Z9g==",
      "license": "MIT"
    },
    "node_modules/react": {
      "version": "18.3.1",
      "resolved": "https://registry.npmjs.org/react/-/react-18.3.1.tgz",
@@ -13,6 +13,7 @@
    "@xterm/addon-fit": "^0.10.0",
    "@xterm/addon-web-links": "^0.11.0",
    "@xterm/xterm": "^5.5.0",
    "qrcode-generator": "^2.0.4",
    "react": "^18.3.1",
    "react-dom": "^18.3.1"
  },
@@ -0,0 +1,150 @@
 // A textarea that knows what it is holding: line numbers in a gutter, YAML (or a .env)
 // coloured underneath the text, and the keys a compose file needs — Tab indents by two,
 // Shift+Tab outdents, Enter keeps the indentation of the line above (one step deeper after
 // a key or a list dash), and a phone, whose keyboard has none of those, gets a small bar of
 // them. The text itself stays in a real textarea, so selection, undo and the keyboard are
 // the browser's own; the colour is a <pre> drawn behind it, kept in step line for line.
 import { useEffect, useMemo, useRef, useState, ReactNode } from 'react'
 export type Language = 'yaml' | 'env' | 'plain'
 interface Props {
  value: string; onChange: (v: string) => void; language?: Language; className?: string
  readOnly?: boolean; placeholder?: string; spellCheck?: boolean
 }
 const INDENT = '  '
 const coarse = () => typeof window !== 'undefined' && window.matchMedia('(pointer: coarse)').matches
 // One line of YAML, split into spans by what each part is. Deliberately shallow: a compose
 // file is keys, scalars, lists and comments; anything cleverer is a parser, which this is not.
 function yamlLine(line: string): ReactNode[] {
  const out: ReactNode[] = []
  let i = 0
  const push = (cls: string, text: string) => { if (text) out.push(cls ? <span key={out.length} className={cls}>{text}</span> : text) }
  const m = /^(\s*)(-\s+)?/.exec(line)!
  push('', m[1]); push('y-dash', m[2] ?? ''); i = m[0].length
  const rest = line.slice(i)
  if (rest.startsWith('#')) { push('y-comment', rest); return out }
  // a key: up to the first colon that is followed by a space or the end
  const key = /^("[^"]*"|'[^']*'|[^\s:#][^:#]*?)(:)(?=\s|$)/.exec(rest)
  let tail = rest
  if (key) { push('y-key', key[1]); push('y-colon', ':'); tail = rest.slice(key[0].length) }
  // the value, with a trailing comment cut off
  let value = tail, comment = ''
  const c = / #.*$/.exec(tail) ?? (tail.startsWith('#') ? /#.*$/.exec(tail) : null)
  if (c) { value = tail.slice(0, c.index); comment = tail.slice(c.index) }
  const v = value.trim()
  const lead = value.slice(0, value.length - value.trimStart().length)
  push('', lead)
  if (!v) { /* nothing */ }
  else if (/^(["']).*\1$/.test(v)) push('y-string', v)
  else if (/^(true|false|yes|no|on|off|null|~)$/i.test(v)) push('y-bool', v)
  else if (/^[-+]?(\d[\d_]*(\.\d+)?|\.\d+)([eE][-+]?\d+)?$/.test(v) || /^0x[0-9a-fA-F]+$/.test(v)) push('y-number', v)
  else if (/^[&*!]/.test(v)) push('y-anchor', v)
  else if (/^[|>][-+]?\d*$/.test(v)) push('y-block', v)
  else push(key ? 'y-value' : (m[2] ? 'y-value' : ''), v)
  push('', value.slice(lead.length + v.length))
  push('y-comment', comment)
  return out
 }
 function envLine(line: string): ReactNode[] {
  if (/^\s*#/.test(line)) return [<span key="c" className="y-comment">{line}</span>]
  const m = /^(\s*(?:export\s+)?)([A-Za-z_][A-Za-z0-9_.-]*)(=)(.*)$/.exec(line)
  if (!m) return [line]
  return [m[1], <span key="k" className="y-key">{m[2]}</span>, <span key="e" className="y-colon">=</span>, <span key="v" className="y-string">{m[4]}</span>]
 }
 function highlight(text: string, lang: Language): ReactNode[] {
  const lines = text.split('\n')
  const f = lang === 'yaml' ? yamlLine : lang === 'env' ? envLine : (l: string) => [l] as ReactNode[]
  return lines.map((l, i) => <div key={i} className="code-line">{f(l)}{l === '' ? '' : ''}</div>)
 }
 // setText replaces [start, end) in a textarea with s, through the browser's own edit so Undo
 // still works, and leaves the caret where asked.
 function setText(ta: HTMLTextAreaElement, start: number, end: number, s: string, caret: number | [number, number]) {
  ta.focus()
  ta.setSelectionRange(start, end)
  let done = false
  try { done = document.execCommand('insertText', false, s) } catch { /* not every browser */ }
  if (!done || ta.value.slice(start, start + s.length) !== s) { ta.setRangeText(s, start, end, 'end'); ta.dispatchEvent(new Event('input', { bubbles: true })) }
  if (Array.isArray(caret)) ta.setSelectionRange(caret[0], caret[1]); else ta.setSelectionRange(caret, caret)
 }
 export default function CodeEditor({ value, onChange, language = 'yaml', className = '', readOnly, placeholder, spellCheck = false }: Props) {
  const ta = useRef<HTMLTextAreaElement>(null)
  const pre = useRef<HTMLPreElement>(null)
  const gutter = useRef<HTMLDivElement>(null)
  const [touch] = useState(coarse)
  const lines = useMemo(() => value.split('\n').length, [value])
  const painted = useMemo(() => highlight(value, language), [value, language])
  // the colour and the numbers follow the text's own scrolling
  const sync = () => { const t = ta.current; if (!t) return; if (pre.current) { pre.current.scrollTop = t.scrollTop; pre.current.scrollLeft = t.scrollLeft } if (gutter.current) gutter.current.scrollTop = t.scrollTop }
  useEffect(sync, [value])
  // lines touched by the selection, for indent and outdent
  const lineRange = (t: HTMLTextAreaElement) => {
    const s = t.value.lastIndexOf('\n', t.selectionStart - 1) + 1
    let e = t.value.indexOf('\n', Math.max(t.selectionEnd - (t.selectionEnd > t.selectionStart && t.value[t.selectionEnd - 1] === '\n' ? 1 : 0), t.selectionStart))
    if (e < 0) e = t.value.length
    return [s, e] as const
  }
  const indent = (out: boolean) => {
    const t = ta.current; if (!t || readOnly) return
    const [s, e] = lineRange(t)
    const block = t.value.slice(s, e)
    const sel = t.selectionStart !== t.selectionEnd || block.includes('\n')
    if (!sel && !out) { const at = t.selectionStart; setText(t, at, at, INDENT, at + INDENT.length); return }
    const next = block.split('\n').map((l) => (out ? l.replace(/^ {1,2}/, '') : INDENT + l)).join('\n')
    if (next === block) return
    const d0 = out ? -(block.length - next.length) : INDENT.length
    setText(t, s, e, next, sel ? [s, s + next.length] : [Math.max(s, t.selectionStart + d0), Math.max(s, t.selectionStart + d0)])
  }
  const comment = () => {
    const t = ta.current; if (!t || readOnly) return
    const [s, e] = lineRange(t)
    const block = t.value.slice(s, e), ls = block.split('\n')
    const all = ls.every((l) => /^\s*#/.test(l) || !l.trim())
    const next = ls.map((l) => (!l.trim() ? l : all ? l.replace(/^(\s*)# ?/, '$1') : l.replace(/^(\s*)/, '$1# '))).join('\n')
    setText(t, s, e, next, [s, s + next.length])
  }
  const newline = () => {
    const t = ta.current; if (!t || readOnly) return
    const at = t.selectionStart
    const lineStart = t.value.lastIndexOf('\n', at - 1) + 1
    const line = t.value.slice(lineStart, at)
    let ind = /^\s*/.exec(line)![0]
    // one step deeper after a key or a list item that opens a block
    if (language === 'yaml' && /(:|^\s*-)\s*(#.*)?$/.test(line) && line.trim() && !/^\s*#/.test(line)) ind += INDENT
    setText(t, at, t.selectionEnd, '\n' + ind, at + 1 + ind.length)
  }
  const onKey = (e: React.KeyboardEvent<HTMLTextAreaElement>) => {
    if (e.key === 'Tab' && !e.ctrlKey && !e.metaKey && !e.altKey) { e.preventDefault(); indent(e.shiftKey) }
    else if (e.key === 'Enter' && !e.ctrlKey && !e.metaKey && !e.altKey && !e.shiftKey) { e.preventDefault(); newline() }
    else if (e.key === '/' && (e.ctrlKey || e.metaKey)) { e.preventDefault(); comment() }
  }
  return (
    <div className={'code ' + className + (readOnly ? ' ro' : '')}>
      <div className="code-gutter" ref={gutter} aria-hidden="true">{Array.from({ length: lines }, (_, i) => <div key={i}>{i + 1}</div>)}</div>
      <div className="code-body">
        <pre className="code-paint" ref={pre} aria-hidden="true">{painted}</pre>
        <textarea ref={ta} className="code-text" value={value} onChange={(e) => onChange(e.target.value)} onScroll={sync} onKeyDown={onKey}
          readOnly={readOnly} placeholder={placeholder} spellCheck={spellCheck} autoCapitalize="off" autoCorrect="off" autoComplete="off" wrap="off" />
      </div>
      {touch && !readOnly && (
        <div className="code-keys" onMouseDown={(e) => e.preventDefault()}>
          <button type="button" title="outdent" onClick={() => indent(true)}>⇤</button>
          <button type="button" title="indent" onClick={() => indent(false)}>⇥</button>
          <button type="button" title="comment" onClick={comment}>#</button>
          <button type="button" title="list item" onClick={() => { const t = ta.current; if (t) setText(t, t.selectionStart, t.selectionEnd, '- ', t.selectionStart + 2) }}>-</button>
          <button type="button" title="colon" onClick={() => { const t = ta.current; if (t) setText(t, t.selectionStart, t.selectionEnd, ': ', t.selectionStart + 2) }}>:</button>
          <button type="button" title="quotes" onClick={() => { const t = ta.current; if (t) setText(t, t.selectionStart, t.selectionEnd, '"' + t.value.slice(t.selectionStart, t.selectionEnd) + '"', t.selectionEnd + 1) }}>"</button>
        </div>
      )}
    </div>
  )
 }
@@ -1,67 +1,120 @@
 // Docker: containers grouped by the stack that deployed them, with start / stop / restart,
 // logs that follow, and inspect — then images, volumes and networks. Live CPU and memory
 // come from the same container sensors everything else uses.
-import { useCallback, useEffect, useMemo, useRef, useState } from 'react'
+import { FormEvent, useCallback, useEffect, useMemo, useRef, useState } from 'react'
-import { api, DockerContainer, DockerInfo, ImageCheck, Stack, StackDiff } from './api'
+import { api, DockerContainer, DockerEvent, DockerHealth, DockerInfo, DockerLogLine, ImageCheck, MachineUpdates, Stack, StackDiff, UpdateJob, UpdateJobRow, wsURLPath } from './api'
 import TaskOutput from './TaskOutput'
 import StackNew, { DupFrom } from './StackNew'
-import { Sensor, fmtValue, useSensorStream } from './sensors'
+import CodeEditor from './CodeEditor'
 import { Sensor, fetchHistories, fmtValue, useSensorStream } from './sensors'
 import { useHashSel } from './hashsel'
-import { useRefreshMs } from './prefs'
+import { fmtWindowSec, useHistorySec, useRefreshMs } from './prefs'
-import { askText, askYes, Btn, CloseBtn, Loading, setUnsaved, toast } from './ui'
+import Sparkline from './Sparkline'
 import { fmtBytes } from './fmt'
 import { describeCron } from './crontext'
 import { askText, askYes, Btn, CloseBtn, Loading, setUnsaved, toast, useEscape } from './ui'
-type Tab = 'containers' | 'stacks' | 'images' | 'volumes' | 'networks'
+type Tab = 'containers' | 'stacks' | 'updates' | 'images' | 'volumes' | 'networks'
 const dot = (c: DockerContainer) => (c.state !== 'running' ? (c.state === 'exited' || c.state === 'dead' ? 'bad' : 'warn') : c.health === 'unhealthy' ? 'bad' : c.health === 'starting' ? 'warn' : 'ok')
 const ports = (p: string) => [...new Set(p.split(', ').map((x) => x.replace(/^\[?::\]?:/, '').replace(/^0\.0\.0\.0:/, '').replace(/^[\d.]+:/, '')).filter((x) => x.includes('->')))].join('  ')
 // A container's output, followed over a held connection the way the Services tab follows a
 // journal. A reconnect resumes from the last line's timestamp, so a container that restarts
 // — or a phone that slept — doesn't replay what is already on screen.
 function LogView({ id }: { id: string }) {
-  const [lines, setLines] = useState<{ t: string; m: string }[]>([])
+  const [lines, setLines] = useState<DockerLogLine[]>([])
  const [err, setErr] = useState('')
  const [live, setLive] = useState(false)
  const [follow, setFollow] = useState(true)
  const [paused, setPaused] = useState(false)
  const box = useRef<HTMLDivElement>(null)
  const last = useRef('')
  const pausedRef = useRef(false)
  pausedRef.current = paused
  const pending = useRef<DockerLogLine[]>([])
  // What a resume has already shown: `--since` is inclusive, and two lines can share a
  // timestamp (a container writing to both streams at once), so the boundary is a set of
  // lines rather than a moment. Bounded — only the lines around a reconnect can repeat.
  const seen = useRef<{ set: Set<string>; q: string[] }>({ set: new Set(), q: [] })
  useEffect(() => {
-    let stop = false; setLines([]); last.current = ''
+    setLines([]); setErr(''); setLive(false); last.current = ''; pending.current = []
-    const tick = async () => {
+    seen.current = { set: new Set(), q: [] }
-      if (stop) return
+    let closed = false
-      if (document.visibilityState === 'visible') {
+    let sock: WebSocket | null = null
-        try {
+    let retry: number | undefined
-          const r = await api.dockerLogs(id, last.current)
+    let attempt = 0
-          const fresh = last.current ? r.lines.filter((l) => l.t > last.current) : r.lines
+    const connect = () => {
-          if (fresh.length) { last.current = fresh[fresh.length - 1].t || last.current; setLines((old) => [...old, ...fresh].slice(-2000)) }
+      if (closed) return
-          setErr('')
+      const q = last.current ? `since=${encodeURIComponent(last.current)}` : 'n=300'
-        } catch (e) { setErr(e instanceof Error ? e.message : 'failed') }
+      // fixed when the log opens: a reconnect must not follow the console to another machine
      const s = new WebSocket(wsURLPath(`/api/docker/containers/${encodeURIComponent(id)}/logs/tail?${q}`))
      sock = s
      s.onopen = () => { attempt = 0; setLive(true); setErr('') }
      s.onmessage = (ev) => {
        const l = JSON.parse(ev.data) as DockerLogLine
        if (l.t) {
          const k = l.t + '|' + (l.err ? 'e' : 'o') + '|' + l.m
          const { set, q } = seen.current
          if (set.has(k)) return
          set.add(k); q.push(k)
          if (q.length > 1000) for (const old of q.splice(0, 500)) set.delete(old)
          if (l.t > last.current) last.current = l.t
        }
        if (pausedRef.current) { pending.current.push(l); return }
        setLines((prev) => (prev.length > 2000 ? [...prev.slice(-1500), l] : [...prev, l]))
      }
-      window.setTimeout(tick, 2000)
+      // a container that has stopped — or was never there — ends the stream at once, so
      // back off rather than asking Docker the same question twice a second forever
      s.onclose = () => { if (closed) return; setLive(false); retry = window.setTimeout(connect, Math.min(15000, 1000 * 2 ** attempt++)) }
      s.onerror = () => { setErr('log stream disconnected'); s.close() }
    }
-    tick()
+    connect()
-    return () => { stop = true }
+    return () => { closed = true; clearTimeout(retry); sock?.close() }
  }, [id])
-  useEffect(() => { if (follow && box.current) box.current.scrollTop = box.current.scrollHeight }, [lines, follow])
+  // Resume: flush what arrived while paused.
  useEffect(() => {
    if (paused || pending.current.length === 0) return
    const p = pending.current; pending.current = []
    setLines((prev) => [...prev, ...p].slice(-2000))
  }, [paused])
  useEffect(() => { if (follow && !paused && box.current) box.current.scrollTop = box.current.scrollHeight }, [lines, follow, paused])
  return (
    <div className="logs">
-      <div className="logs-bar"><span className="muted">{lines.length} lines</span><span className="spacer" /><button className={'chip' + (follow ? ' on' : '')} onClick={() => setFollow(!follow)}>follow</button><button className="chip" onClick={() => setLines([])}>Clear</button></div>
+      <div className="logs-bar">
-      {err && <div className="error">{err}</div>}
+        <span className={'dot ' + (live ? 'ok' : 'bad')} /> <span className="muted">{live ? 'following' : 'reconnecting…'} · {lines.length} lines{err && ` · ${err}`}</span>
        <span className="spacer" />
        <button className={'chip' + (follow ? ' on' : '')} onClick={() => setFollow(!follow)}>follow</button>
        <button className="chip" onClick={() => setPaused(!paused)}>{paused ? `Resume${pending.current.length ? ` (${pending.current.length})` : ''}` : 'Pause'}</button>
        <button className="chip" onClick={() => setLines([])}>Clear</button>
      </div>
      <div className="log-lines" ref={box} onScroll={(e) => { const el = e.currentTarget; if (el.scrollHeight - el.scrollTop - el.clientHeight > 40) setFollow(false) }}>
-        {lines.map((l, i) => <div className="log" key={i}><span className="lt">{l.t ? new Date(l.t).toLocaleTimeString() : ''}</span><span className="lm">{l.m}</span></div>)}
+        {lines.map((l, i) => <div className={'log' + (l.err ? ' err' : '')} key={i}><span className="lt">{l.t ? new Date(l.t).toLocaleTimeString() : ''}</span><span className="lm">{l.m}</span></div>)}
        {lines.length === 0 && !err && <div className="muted small" style={{ padding: 8 }}>No output yet.</div>}
      </div>
    </div>
  )
 }
 // What a container is doing, and what can be done to it: its numbers over time, its health
 // checks, what has happened to it lately, a shell inside it, and its limits.
 function Detail({ c, live, onBack, onChanged }: { c: DockerContainer; live: Record<string, Sensor>; onBack: () => void; onChanged: () => void }) {
-  const [tab, setTab] = useState<'logs' | 'inspect'>('logs')
+  const [tab, setTab] = useState<'logs' | 'stats' | 'health' | 'events' | 'inspect'>('logs')
  const [info, setInfo] = useState<Record<string, unknown> | null>(null)
-  useEffect(() => { setInfo(null); if (tab === 'inspect') api.dockerInspect(c.name).then(setInfo).catch((e) => toast(e instanceof Error ? e.message : 'failed', 'err')) }, [tab, c.name])
+  const [limits, setLimits] = useState(false)
  useEffect(() => { setInfo(null); if (tab === 'inspect' || limits) api.dockerInspect(c.name).then(setInfo).catch((e) => toast(e instanceof Error ? e.message : 'failed', 'err')) }, [tab, limits, c.name])
  const act = (a: 'start' | 'stop' | 'restart' | 'remove') => async () => { try { await api.dockerAction(c.name, a); toast(`${c.name}: ${a} ok`); onChanged() } catch (e) { toast(`${c.name}: ${e instanceof Error ? e.message : 'failed'}`, 'err') } }
  const running = c.state === 'running'
-  const s = (k: string) => live[`container.${c.name.toLowerCase().replace(/[^a-z0-9-]+/g, '_').replace(/^_|_$/g, '')}.${k}`]
+  const seg = c.name.toLowerCase().replace(/[^a-z0-9-]+/g, '_').replace(/^_|_$/g, '')
  const s = (k: string) => live[`container.${seg}.${k}`]
  const mounts = (info?.mounts as { Source: string; Destination: string; RW: boolean }[] | undefined) ?? []
  const shell = async () => {
    try { const t = await api.dockerExec(c.name); toast(`shell open as “${t.name}” on the Terminal tab`); window.dispatchEvent(new CustomEvent('hamlab:terminal', { detail: t.id })) }
    catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  return (
    <div className="svc-detail">
      <CloseBtn onClick={onBack} />
-      <div className="svc-head"><button className="btn back" onClick={onBack}>‹ Back</button><span className={'sdot ' + dot(c)} /><span className="dname">{c.name}</span></div>
+      <div className="svc-head"><button className="btn back" onClick={onBack}>‹ Back</button><span className={'sdot ' + dot(c)} /><span className="dname">{c.name}</span>{c.health && c.health !== 'none' && <span className={'badge' + (c.health === 'healthy' ? '' : ' bad')}>{c.health}</span>}</div>
      <div className="props">
        <div><span className="k">Image</span><span className="mono small">{c.image}</span></div>
        <div><span className="k">Status</span><span>{c.status}</span></div>
@@ -72,27 +125,181 @@ function Detail({ c, live, onBack, onChanged }: { c: DockerContainer; live: Reco
      <div className="svc-actions">
        {!running && <Btn className="btn on" onClick={act('start')}>Start</Btn>}
        {!running && <Btn className="btn danger" confirm={`Remove the container ${c.name}? Its image, volumes and bind-mounted data stay; anything written inside the container itself is lost.${c.stack ? ` Deploying the ${c.stack} stack makes it again.` : ''}`} onClick={async () => { await act('remove')(); onBack() }}>Remove</Btn>}
        {running && <Btn className="btn" onClick={shell}>Shell</Btn>}
        {running && <Btn className="btn" confirm={`Restart ${c.name}?`} onClick={act('restart')}>Restart</Btn>}
        <Btn className="btn" onClick={() => setLimits(true)}>Limits</Btn>
        {running && <Btn className="btn danger" confirm={`Stop ${c.name}? It stays stopped until something starts it.`} onClick={act('stop')}>Stop</Btn>}
      </div>
-      <div className="tabs"><button className={tab === 'logs' ? 'on' : ''} onClick={() => setTab('logs')}>Logs</button><button className={tab === 'inspect' ? 'on' : ''} onClick={() => setTab('inspect')}>Inspect</button></div>
+      <div className="tabs">
        <button className={tab === 'logs' ? 'on' : ''} onClick={() => setTab('logs')}>Logs</button>
        <button className={tab === 'stats' ? 'on' : ''} onClick={() => setTab('stats')}>Stats</button>
        <button className={tab === 'health' ? 'on' : ''} onClick={() => setTab('health')}>Health</button>
        <button className={tab === 'events' ? 'on' : ''} onClick={() => setTab('events')}>Events</button>
        <button className={tab === 'inspect' ? 'on' : ''} onClick={() => setTab('inspect')}>Inspect</button>
      </div>
      {tab === 'logs' && <LogView id={c.name} />}
      {tab === 'stats' && <Stats seg={seg} live={live} />}
      {tab === 'health' && <Health id={c.name} />}
      {tab === 'events' && <Events id={c.name} />}
      {tab === 'inspect' && (!info ? <Loading label="Inspecting…" /> : (
-        <div className="props">
+        <div className="props stacked">
          <div><span className="k">Restarts</span><span>{String(info.restart || 'no')} policy · restarted {String(info.restarts ?? 0)}×</span></div>
          <div><span className="k">Network</span><span className="mono small">{Object.keys((info.networks as object) ?? {}).join(', ') || String(info.network_mode)}</span></div>
          {Number(info.memory_limit) > 0 && <div><span className="k">Mem limit</span><span>{fmtValue(Number(info.memory_limit), 'B')}</span></div>}
          {!!info.error && <div><span className="k">Error</span><span className="warn-text">{String(info.error)}</span></div>}
          {!!info.compose_file && <div><span className="k">Compose</span><span className="mono small">{String(info.compose_file)}</span></div>}
-          {mounts.map((m) => <div key={m.Destination}><span className="k">Mount</span><span className="mono small">{m.Source} → {m.Destination}{m.RW ? '' : ' (ro)'}</span></div>)}
+          {mounts.length > 0 && <div><span className="k">Mounts</span><span className="mono small">{mounts.map((m) => <div key={m.Destination}>{m.Source} → {m.Destination}{m.RW ? '' : ' (read only)'}</div>)}</span></div>}
-          <div><span className="k">Env</span><span className="mono small muted">{((info.env_names as string[]) ?? []).join('  ')} <span className="muted">(names only)</span></span></div>
+          <div><span className="k">Env <span className="muted">names only</span></span><span className="mono small muted">{((info.env_names as string[]) ?? []).join('  ') || '—'}</span></div>
        </div>
      ))}
      {limits && <LimitsDialog c={c} info={info} onClose={() => setLimits(false)} onSaved={() => { setLimits(false); onChanged() }} />}
    </div>
  )
 }
 // CPU and memory over time, from the container sensors the sampler already keeps.
 function Stats({ seg, live }: { seg: string; live: Record<string, Sensor> }) {
  const [hist, setHist] = useState<{ id: string; t: number[]; v: number[] }[] | null>(null)
  const historySec = useHistorySec()
  useEffect(() => {
    let alive = true
    fetchHistories(`container.${seg}.cpu,container.${seg}.mem`)
      .then((h) => { if (alive) setHist(h) })
      .catch(() => { if (alive) setHist([]) })
    return () => { alive = false }
  }, [seg])
  const css = (n: string) => getComputedStyle(document.documentElement).getPropertyValue(n).trim() || '#3987e5'
  const of = (k: string) => hist?.find((h) => h.id === `container.${seg}.${k}`)?.v ?? []
  const cpu = of('cpu'), mem = of('mem')
  if (hist === null) return <Loading label="Reading the history…" />
  if (cpu.length < 2 && mem.length < 2) return <div className="notice">No history yet for this container — it is sampled while it runs, so give it a moment.</div>
  return (
    <div className="cards">
      <section className="card">
        <div className="sub-head"><span className="dname">CPU</span><span className="spacer" /><span className="num">{fmtValue(live[`container.${seg}.cpu`]?.value, '%')} of a core</span></div>
        <Sparkline series={[{ name: 'cpu', values: cpu, color: css('--s1') }]} format={(v) => `${v.toFixed(1)}%`} height={64} />
        <div className="muted small">the last {fmtWindowSec(historySec)}</div>
      </section>
      <section className="card">
        <div className="sub-head"><span className="dname">Memory</span><span className="spacer" /><span className="num">{fmtValue(live[`container.${seg}.mem`]?.value, 'B')}</span></div>
        <Sparkline series={[{ name: 'mem', values: mem, color: css('--s2') }]} format={(v) => fmtBytes(v)} height={64} />
        <div className="muted small">the last {fmtWindowSec(historySec)}</div>
      </section>
    </div>
  )
 }
 // The health check's verdict, and what its last few probes actually said.
 function Health({ id }: { id: string }) {
  const [h, setH] = useState<DockerHealth | null>(null)
  const [err, setErr] = useState('')
  useEffect(() => { setH(null); api.dockerHealth(id).then(setH).catch((e) => setErr(e instanceof Error ? e.message : 'failed')) }, [id])
  if (err) return <div className="error">{err}</div>
  if (!h) return <Loading label="Asking…" />
  if (!h.configured) return <div className="notice">This image defines no health check, so Docker only knows whether the process is running — which it is ({h.state}). A health check is a <span className="mono">healthcheck:</span> block in the compose file.</div>
  return (
    <>
      <div className="props">
        <div><span className="k">Health</span><span className={h.status === 'healthy' ? '' : 'warn-text'}>{h.status}</span></div>
        <div><span className="k">State</span><span>{h.state}</span></div>
        {!!h.failing_streak && <div><span className="k">Failing</span><span className="warn-text">{h.failing_streak} probe{h.failing_streak === 1 ? '' : 's'} in a row</span></div>}
      </div>
      <div className="muted small" style={{ marginTop: 8 }}>The last {h.log?.length ?? 0} probes — Docker keeps five.</div>
      {(h.log ?? []).slice().reverse().map((p, i) => (
        <section className="card" key={i} style={{ marginTop: 6 }}>
          <div className="sub-head">
            <span className={'sdot ' + (p.ExitCode === 0 ? 'ok' : 'bad')} />
            <span>{new Date(p.Start).toLocaleString()}</span>
            <span className="spacer" />
            <span className="muted small">exit {p.ExitCode} · {Math.max(0, new Date(p.End).getTime() - new Date(p.Start).getTime())} ms</span>
          </div>
          {p.Output.trim() && <pre className="run-out" style={{ maxHeight: 140 }}>{p.Output.trim()}</pre>}
        </section>
      ))}
    </>
  )
 }
 // What has happened to this container lately, asked of Docker rather than kept here.
 function Events({ id }: { id: string }) {
  const [ev, setEv] = useState<DockerEvent[] | null>(null)
  const [kept, setKept] = useState(0)
  const [since, setSince] = useState('7d')
  const [err, setErr] = useState('')
  useEffect(() => { setEv(null); setErr(''); api.dockerEvents(id, since).then((r) => { setEv(r.events); setKept(r.kept) }).catch((e) => setErr(e instanceof Error ? e.message : 'failed')) }, [id, since])
  const bad = (a: string, h?: string) => /die|kill|oom|destroy/.test(a) || (h !== undefined && h !== 'healthy')
  return (
    <>
      <div className="chips" style={{ marginBottom: 8 }}>
        {['24h', '7d', '30d'].map((x) => <button key={x} className={'chip' + (since === x ? ' on' : '')} onClick={() => setSince(x)}>{x}</button>)}
      </div>
      {err && <div className="error">{err}</div>}
      {!ev && !err && <Loading label="Asking Docker…" />}
      {ev && ev.length === 0 && <div className="notice">Nothing for this container in the last {since}. Ham Lab keeps its own record — Docker's own only reaches back a few minutes — so history starts from when this version was installed. {kept} events kept across every container.</div>}
      {ev && ev.length > 0 && (
        <div className="snap-list">
          {ev.map((e, i) => (
            <div className="snap" key={i}>
              <span className={bad(e.action, e.health) ? 'warn-text' : ''}>{e.action}{e.health ? `: ${e.health}` : ''}</span>
              <span className="muted">{new Date(e.t).toLocaleString()}</span>
              <span className="muted small">{e.exit_code ? `exit ${e.exit_code}` : ''}{e.signal ? ` signal ${e.signal}` : ''}</span>
              <span />
              <span />
            </div>
          ))}
        </div>
      )}
    </>
  )
 }
 // Memory, CPU and restart policy, changed on the running container. Docker does this live,
 // which is the point — and also the catch, said plainly below.
 function LimitsDialog({ c, info, onClose, onSaved }: { c: DockerContainer; info: Record<string, unknown> | null; onClose: () => void; onSaved: () => void }) {
  const curMem = Number(info?.memory_limit ?? 0)
  const [mem, setMem] = useState(curMem > 0 ? String(Math.round(curMem / (1 << 20))) + 'm' : '')
  const [cpus, setCpus] = useState('')
  const [restart, setRestart] = useState(String(info?.restart ?? ''))
  const [err, setErr] = useState('')
  const [busy, setBusy] = useState(false)
  useEscape(onClose)
  const save = async (e: FormEvent) => {
    e.preventDefault(); setErr(''); setBusy(true)
    try { await api.dockerLimits(c.name, { memory: mem, cpus, restart }); toast(`${c.name} updated`); onSaved() }
    catch (e) { setErr(e instanceof Error ? e.message : 'failed'); setBusy(false) }
  }
  return (
    <div className="modal-back" onClick={onClose}>
      <form className="modal" onClick={(e) => e.stopPropagation()} onSubmit={save}>
        <h3>Limits for {c.name}</h3>
        <label className="muted">Memory — 512m, 2g, or 0 for no limit{curMem > 0 ? ` (now ${fmtBytes(curMem)})` : ' (no limit now)'}
          <input className="mono" value={mem} onChange={(e) => setMem(e.target.value)} placeholder="leave blank to keep" spellCheck={false} />
        </label>
        <label className="muted">CPUs — 1.5 means one and a half cores, 0 for no limit
          <input className="mono" value={cpus} onChange={(e) => setCpus(e.target.value)} placeholder="leave blank to keep" spellCheck={false} />
        </label>
        <label className="muted">Restart policy
          <select value={restart} onChange={(e) => setRestart(e.target.value)}>
            <option value="">leave as it is</option>
            {['no', 'on-failure', 'unless-stopped', 'always'].map((p) => <option key={p} value={p}>{p}</option>)}
          </select>
        </label>
        <div className="notice attn">
          This changes the running container now, without recreating it. It does <b>not</b> edit the
          compose file{c.stack ? <> — so deploying <b>{c.stack}</b> puts back whatever the file says</> : ''}.
          Edit the stack to make it stick.
        </div>
        {err && <div className="error">{err}</div>}
        <div className="svc-actions">
          <button type="submit" className="btn on" disabled={busy || (!mem && !cpus && !restart)}>Apply</button>
          <span className="spacer" />
          <button type="button" className="btn" onClick={onClose}>Cancel</button>
        </div>
      </form>
    </div>
  )
 }
 // A plain line diff (longest common subsequence): compose files are short, so the simple way is fine.
 function lineDiff(a: string, b: string): { s: ' ' | '+' | '-'; t: string }[] {
  const x = a.split('\n'), y = b.split('\n'); const n = x.length, m = y.length
  const L = Array.from({ length: n + 1 }, () => new Uint16Array(m + 1))
@@ -120,6 +327,7 @@ const maskEnv = (text: string) => text.split('\n').map((l) => { const i = l.inde
 // One stack: its compose file to edit, and the four things to do with it. Saving validates
 // with compose itself and keeps the previous file beside it; deploying is a separate, asked-for act.
 function StackDetail({ name, startTask, onBack, onChanged, onDuplicate }: { name: string; startTask?: string; onBack: () => void; onChanged: () => void; onDuplicate: (from: DupFrom) => void }) {
  const [st, setSt] = useState<Stack | null>(null)
  const [text, setText] = useState(''); const [orig, setOrig] = useState('')
@@ -140,7 +348,7 @@ function StackDetail({ name, startTask, onBack, onChanged, onDuplicate }: { name
    try { const r = await api.deleteStack(name); toast(`${name} deleted — its files are in ${r.trash}`); setUnsaved(''); onChanged(); onBack() } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  const duplicate = async () => { if (!st) return; try { onDuplicate({ name, text: file === 0 ? orig : (await api.stackFile(name, 0)).text, hasEnv: !!st.has_env, files: st.host_files?.length ?? 1 }) } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') } }
-  const run = (a: 'deploy' | 'update' | 'restart' | 'stop' | 'down') => async () => { try { setTask((await api.stackAction(name, a)).id) } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') } }
+  const run = (a: 'deploy' | 'update' | 'recreate' | 'restart' | 'stop' | 'down') => async () => { try { setTask((await api.stackAction(name, a)).id) } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') } }
  if (err && !st) return <div className="svc-detail"><CloseBtn onClick={onBack} /><button className="btn back" onClick={onBack}>‹ Back</button><div className="notice">{err}</div></div>
  if (!st) return <div className="svc-detail"><Loading label="Reading the compose file…" /></div>
  return (
@@ -150,7 +358,7 @@ function StackDetail({ name, startTask, onBack, onChanged, onDuplicate }: { name
      <div className="muted small mono" style={{ overflowWrap: 'anywhere' }}>{file === 'env' ? (st.host_file.replace(/[^/]+$/, '') + '.env') : (st.host_files?.[file] ?? st.host_file)}{st.held_by && <span className="muted"> · kept on <b>{st.held_by}</b>, deployed to this machine over {st.remote}</span>}{st.via && <span className="muted"> · compose runs inside <b>{st.via}</b>, as it was deployed</span>}{st.has_env && ' · has a .env beside it'}</div>
      {((st.host_files?.length ?? 0) > 1 || st.has_env) && <div className="chips" title="this stack is deployed from several files, in this order; Deploy uses all of them">{st.host_files!.map((f, i) => <button key={f} className={'chip' + (i === file ? ' on' : '')} onClick={async () => { if (dirty && !(await askYes('Discard unsaved changes to this file?', 'Discard'))) return; setFile(i) }}>{f.split('/').pop()}</button>)}{st.has_env && <button className={'chip' + (file === 'env' ? ' on' : '')} title="the .env beside the compose file: variables compose substitutes" onClick={async () => { if (dirty && !(await askYes('Discard unsaved changes to this file?', 'Discard'))) return; setFile('env') }}>.env</button>}</div>}
      {file === 'env' && !reveal ? <div className="env-masked"><pre className="editor stack-editor">{maskEnv(text)}</pre><button className="btn" onClick={() => setReveal(true)}>Reveal and edit</button></div> :
-      <textarea className="editor stack-editor" value={text} onChange={(e) => setText(e.target.value)} spellCheck={false} autoCapitalize="off" autoCorrect="off" />}
+      <CodeEditor className="stack-editor" language={file === 'env' ? 'env' : 'yaml'} value={text} onChange={setText} />}
      {err && <div className="error" style={{ whiteSpace: 'pre-wrap' }}>{err}{/changed on disk/.test(err) && <> <button className="link" onClick={async () => { if (await askYes('Reload the file from disk? Your text in the box is replaced.', 'Reload')) load() }}>reload it</button></>}</div>}
      <div className="svc-actions">
        <Btn className="btn on" disabled={!dirty} onClick={save}>Save</Btn>
@@ -159,6 +367,7 @@ function StackDetail({ name, startTask, onBack, onChanged, onDuplicate }: { name
        <Btn className="btn" disabled={dirty} title={dirty ? 'save first' : 'recreate what changed: docker compose up -d'} confirm={`Deploy ${st.name}? Containers whose definition changed are recreated (a few seconds down each).`} onClick={run('deploy')}>Deploy</Btn>
        {st.containers.length > 0 && <Btn className="btn" disabled={dirty} title="pull newer images, then deploy" confirm={`Pull newer images for ${st.name} and redeploy? Anything with a new image is recreated.`} onClick={run('update')}>Update</Btn>}
        {st.containers.length > 0 && <Btn className="btn" confirm={`Restart every container in ${st.name}?`} onClick={run('restart')}>Restart</Btn>}
        {st.containers.length > 0 && <Btn className="btn" disabled={dirty} title="build the containers again from scratch: docker compose up -d --force-recreate" confirm={`Recreate every container in ${st.name}?\n\nThis is the way out of a container Docker refuses to start — one pinned to a network that no longer exists, say, which answers "network … not found" to an ordinary Deploy. Nothing in a volume or a bind mount is touched.`} onClick={run('recreate')}>Recreate</Btn>}
        {st.running > 0 && <Btn className="btn danger" confirm={`Stop every container in ${st.name}? They stay stopped until deployed or started again.`} onClick={run('stop')}>Stop</Btn>}
      </div>
      <div className="svc-actions">
@@ -203,6 +412,12 @@ export default function Docker() {
  // what a new stack could join, and what it could collide with
  const nets = useMemo(() => (d?.networks ?? []).map((n) => String(n.Name)).filter((n) => !['bridge', 'host', 'none'].includes(n)).sort(), [d])
  const taken = useMemo(() => { const ports: Record<string, string> = {}; for (const c of d?.containers ?? []) for (const m of (c.ports ?? '').matchAll(/:(\d+)->/g)) ports[m[1]] = c.name; return { names: (d?.containers ?? []).map((c) => c.name), ports } }, [d])
  // What the registry last said about each image, so a container can be marked without
  // asking again. Only ever read from the cache here — checking is a deliberate act.
  const [upd, setUpd] = useState<MachineUpdates | null>(null)
  useEffect(() => { api.machineUpdates().then(setUpd).catch(() => setUpd(null)) }, [])
  const imgState = useMemo(() => Object.fromEntries((upd?.images ?? []).map((i) => [i.image, i.state])), [upd])
  const behind = (upd?.images ?? []).filter((i) => i.state === 'update').length
  if (!d) return <div className="dash">{err ? <div className="notice">{err === 'not found' ? "This machine's Ham Lab is older than the Docker tab. Update it from Machines." : err}</div> : <Loading label="Asking Docker…" />}</div>
  if (!d.available) return <div className="dash"><div className="notice">{d.message}</div></div>
  const selStack = sel.startsWith('stack:') ? sel.slice(6) : ''
@@ -210,12 +425,12 @@ export default function Docker() {
  const shownTab: Tab = selStack ? 'stacks' : tab
  const up = d.containers.filter((c) => c.state === 'running').length
  const sid = (c: DockerContainer, k: string) => sensors[`container.${c.name.toLowerCase().replace(/[^a-z0-9-]+/g, '_').replace(/^_|_$/g, '')}.${k}`]
-  const cols: Record<Exclude<Tab, 'containers' | 'stacks'>, [string, string][]> = { images: [['Repository', 'Repository'], ['Tag', 'Tag'], ['Size', 'Size'], ['CreatedSince', 'Created']], volumes: [['Name', 'Name'], ['Driver', 'Driver'], ['Mountpoint', 'Mountpoint']], networks: [['Name', 'Name'], ['Driver', 'Driver'], ['Scope', 'Scope']] }
+  const imageCols: [string, string][] = [['Repository', 'Repository'], ['Tag', 'Tag'], ['Size', 'Size'], ['CreatedSince', 'Created']]
  return (
-    <div className={'svc docker' + (((cur || sel) && shownTab === 'containers') || selStack ? ' has-sel' : '')}>
+    <div className={'svc docker' + (((cur || sel) && shownTab === 'containers') || selStack ? ' has-sel' : '') + (shownTab === 'updates' ? ' wide' : '')}>
      <div className="svc-list">
        <div className="svc-tools">
-          <div className="chips">{(['containers', 'stacks', 'images', 'volumes', 'networks'] as Tab[]).map((t) => <button key={t} className={'chip' + (shownTab === t ? ' on' : '')} onClick={() => { setTab(t); if (sel) setSel('') }}>{t} · {t === 'containers' ? `${up}/${d.containers.length}` : t === 'stacks' ? (stacks?.length ?? '…') : d[t].length}</button>)}</div>
+          <div className="chips">{(['containers', 'stacks', 'updates', 'images', 'volumes', 'networks'] as Tab[]).map((t) => <button key={t} className={'chip' + (shownTab === t ? ' on' : '')} onClick={() => { setTab(t); if (sel) setSel('') }}>{t}{t === 'containers' ? ` · ${up}/${d.containers.length}` : t === 'stacks' ? ` · ${stacks?.length ?? '…'}` : t === 'updates' ? (behind > 0 ? ` · ${behind}` : '') : ` · ${d[t].length}`}</button>)}</div>
          {(shownTab === 'containers' || shownTab === 'stacks') && <div className="svc-actions"><button className="btn" onClick={() => setMaking(shownTab === 'stacks' ? 'stack' : 'container')}>{shownTab === 'stacks' ? '+ New stack' : '+ New container'}</button></div>}
          {shownTab === 'images' && <form className="chan-add" onSubmit={async (e) => { e.preventDefault(); try { setPullTask((await api.pullImage(pull.trim())).id) } catch (er) { toast(er instanceof Error ? er.message : 'failed', 'err') } }}>
            <input className="mono" value={pull} onChange={(e) => setPull(e.target.value)} placeholder="pull an image — nginx:alpine, ghcr.io/owner/name:tag" spellCheck={false} autoCapitalize="off" autoCorrect="off" />
@@ -224,6 +439,7 @@ export default function Docker() {
            <Btn className="btn" title="remove dangling images: layers no tag points at any more" confirm="Remove dangling images? Only layers that no tag points at any more — nothing a container or a tag uses." onClick={async () => { try { toast((await api.pruneImages()).result); load() } catch (er) { toast(er instanceof Error ? er.message : 'failed', 'err') } }}>Prune</Btn>
          </form>}
          {shownTab === 'volumes' && (() => { const n = d.volumes.filter((v) => v.anonymous && !((v.used_by as string[] | undefined) ?? []).length).length; return <div className="svc-actions"><span className="muted small">Deleting a volume deletes its data. Only one nothing uses can be deleted.</span><span className="spacer" />{n > 0 && <Btn className="btn danger" title="the nameless volumes an image's VOLUME line leaves behind each time its container is recreated" confirm={`Delete ${n} unused anonymous volume${n === 1 ? '' : 's'} and what is in them? Named volumes are not touched.`} onClick={async () => { try { toast((await api.pruneVolumes()).result); load() } catch (er) { toast(er instanceof Error ? er.message : 'failed', 'err') } }}>Prune anonymous · {n}</Btn>}</div> })()}
          {shownTab === 'networks' && (() => { const n = d.networks.filter((x) => !x.predefined && !((x.used_by as string[] | undefined) ?? []).length).length; return <div className="svc-actions"><span className="muted small">A network holds no data. Only one nothing is attached to can be removed.</span><span className="spacer" />{n > 0 && <Btn className="btn" title="remove every network no container is attached to" confirm={`Remove ${n} network${n === 1 ? '' : 's'} nothing is attached to? A stack that is only down counts as unattached — its network comes back on the next deploy, with a new subnet.`} onClick={async () => { try { toast((await api.pruneNetworks()).result || 'nothing to remove'); load() } catch (er) { toast(er instanceof Error ? er.message : 'failed', 'err') } }}>Prune · {n}</Btn>}</div> })()}
          {shownTab === 'images' && pullTask && <TaskOutput id={pullTask} onDone={() => load()} />}
          {shownTab === 'containers' && <input value={filter} onChange={(e) => setFilter(e.target.value)} placeholder="filter containers, images, stacks" />}
        </div>
@@ -242,7 +458,7 @@ export default function Docker() {
              {list.map((c) => (
                <div key={c.id} className={'svc-row' + (c.name === sel ? ' active' : '')} onClick={() => setSel(c.name)}>
                  <span className={'sdot ' + dot(c)} />
-                  <div className="svc-name"><div className="n">{c.name}</div><div className="d">{c.image}</div></div>
+                  <div className="svc-name"><div className="n">{c.name}<UpdateDot state={imgState[c.image]} /></div><div className="d">{c.image}</div></div>
                  <div className="svc-state">{c.state === 'running' ? <>{fmtValue(sid(c, 'mem')?.value ?? c.mem, 'B')}<br />{fmtValue(sid(c, 'cpu')?.value, '%')}</> : c.state}</div>
                </div>
              ))}
@@ -260,16 +476,208 @@ export default function Docker() {
                return <tr key={name}><td className="mono small" title={String(row.Mountpoint ?? '')}>{anon ? <>{name.slice(0, 12)}… <span className="muted">anonymous</span></> : name}</td><td className="small">{users.length ? users.join(', ') : <span className="muted">nothing</span>}</td><td className="row-end">{users.length ? <span className="badge">in use</span> : <button className="icon-btn" title="delete this volume and its data" onClick={remove}>✕</button>}</td></tr> })}
                {d.volumes.length === 0 && <tr><td colSpan={3} className="muted small">No volumes.</td></tr>}</tbody></table>
          )}
-          {shownTab !== 'containers' && shownTab !== 'stacks' && shownTab !== 'volumes' && (
+          {shownTab === 'networks' && (
-            <table className="dtable"><thead><tr>{cols[shownTab as Exclude<Tab, 'containers' | 'stacks'>].map(([, h]) => <th key={h}>{h}</th>)}{shownTab === 'images' && <th />}</tr></thead>
+            <table className="dtable"><thead><tr><th>Network</th><th>Driver</th><th>Used by</th><th /></tr></thead>
-              <tbody>{d[shownTab as Exclude<Tab, 'containers' | 'stacks'>].map((row, i) => <tr key={i}>{cols[shownTab as Exclude<Tab, 'containers' | 'stacks'>].map(([k]) => <td key={k} className="mono small">{String(row[k] ?? '')}</td>)}{shownTab === 'images' && <td className="row-end">{row.in_use ? <span className="badge" title="a container uses this image">in use</span> : <Btn className="icon-btn" title="remove this image" confirm={`Remove the image ${row.Repository}:${row.Tag}? Docker refuses if a container still uses it.`} onClick={async () => { const ref = row.Tag && row.Tag !== '<none>' ? `${row.Repository}:${row.Tag}` : String(row.ID); try { await api.removeImage(ref); toast(`${ref} removed`); load() } catch (er) { toast(er instanceof Error ? er.message : 'failed', 'err') } }}>✕</Btn>}</td>}</tr>)}</tbody></table>
+              <tbody>{d.networks.map((row, i) => { const name = String(row.Name), users = (row.used_by as string[] | undefined) ?? [], own = !!row.predefined
                return <tr key={i}><td className="mono small">{name}{own && <span className="muted"> · Docker's own</span>}</td><td className="mono small">{String(row.Driver ?? '')}</td><td className="small">{users.length ? users.join(', ') : <span className="muted">nothing</span>}</td><td className="row-end">{users.length ? <span className="badge" title="a container is attached to this network">in use</span> : own ? null : <Btn className="icon-btn" title="remove this network" confirm={`Remove the network ${name}? Nothing is attached to it. A stack that wants it back gets it recreated on its next deploy, with a new subnet.`} onClick={async () => { try { await api.removeNetwork(name); toast(`${name} removed`); load() } catch (er) { toast(er instanceof Error ? er.message : 'failed', 'err') } }}>✕</Btn>}</td></tr> })}
                {d.networks.length === 0 && <tr><td colSpan={4} className="muted small">No networks.</td></tr>}</tbody></table>
          )}
          {shownTab === 'images' && (
            <table className="dtable"><thead><tr>{imageCols.map(([, h]) => <th key={h}>{h}</th>)}<th /></tr></thead>
              <tbody>{d.images.map((row, i) => <tr key={i}>{imageCols.map(([k]) => <td key={k} className="mono small">{String(row[k] ?? '')}</td>)}<td className="row-end">{row.in_use ? <span className="badge" title="a container uses this image">in use</span> : <Btn className="icon-btn" title="remove this image" confirm={`Remove the image ${row.Repository}:${row.Tag}? Docker refuses if a container still uses it.`} onClick={async () => { const ref = row.Tag && row.Tag !== '<none>' ? `${row.Repository}:${row.Tag}` : String(row.ID); try { await api.removeImage(ref); toast(`${ref} removed`); load() } catch (er) { toast(er instanceof Error ? er.message : 'failed', 'err') } }}>✕</Btn>}</td></tr>)}</tbody></table>
          )}
        </div>
      </div>
      {shownTab === 'containers' && !cur && sel && !selStack && <div className="svc-detail"><CloseBtn onClick={() => setSel('')} /><button className="btn back" onClick={() => setSel('')}>‹ Back</button><div className="notice">There is no container called <b className="mono">{sel}</b> on this machine. It may have been removed, renamed or recreated under another name. <button className="link" onClick={() => setSel('')}>Show them all</button></div></div>}
      {shownTab === 'containers' && (cur ? <Detail c={cur} live={sensors} onBack={() => setSel('')} onChanged={load} /> : (sel && !selStack ? null : <div className="svc-empty muted">Select a container.</div>))}
      {making && <StackNew key={'new:' + (dup?.name ?? making)} mode={making} from={dup ?? undefined} networks={nets} taken={taken} onClose={() => { setMaking(''); setDup(null) }} onCreated={async (name, deploy) => { setMaking(''); setDup(null); setTab('stacks'); if (deploy) { try { setStartTask({ name, id: (await api.stackAction(name, 'deploy')).id }) } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') } } else setStartTask(null); await load(); setSel('stack:' + name) }} />}
      {shownTab === 'updates' && <Updates upd={upd} onChecked={setUpd} />}
      {shownTab === 'stacks' && (selStack ? <StackDetail key={selStack + (startTask?.name === selStack ? startTask.id : '')} name={selStack} startTask={startTask?.name === selStack ? startTask.id : undefined} onBack={() => setSel('')} onChanged={load} onDuplicate={(f) => { setDup(f); setMaking('stack') }} /> : <div className="svc-empty muted">Select a stack to edit its compose file.</div>)}
    </div>
  )
 }
 // A container's image against the registry, as a single mark: nothing when nobody has
 // asked, so an uncoloured list means "not checked" rather than "all current".
 function UpdateDot({ state }: { state?: string }) {
  if (!state || state === 'current') return null
  const known = state === 'update'
  return <span className={'updot' + (known ? ' behind' : ' unknown')} title={known ? 'a newer image has been published' : 'the registry could not say'}>{known ? '↑' : '?'}</span>
 }
 // What is out of date here, and the schedules that keep it from happening again.
 function Updates({ upd, onChecked }: { upd: MachineUpdates | null; onChecked: (u: MachineUpdates) => void }) {
  const [busy, setBusy] = useState(false)
  const [jobs, setJobs] = useState<UpdateJobRow[] | null>(null)
  const [file, setFile] = useState('')
  const [editing, setEditing] = useState<Partial<UpdateJob> | null>(null)
  const [stacks, setStacks] = useState<string[]>([])
  const loadJobs = useCallback(async () => {
    try { const r = await api.updateJobs(); setJobs(r.jobs); setFile(r.file) } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err'); setJobs([]) }
  }, [])
  useEffect(() => { loadJobs(); api.stacks().then((r) => setStacks(r.stacks.map((s) => s.name))).catch(() => {}) }, [loadJobs])
  const check = async () => {
    setBusy(true)
    try { onChecked(await api.machineUpdates(true)); toast('checked') } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') } finally { setBusy(false) }
  }
  const run = async (j: UpdateJob) => {
    if (!await askYes(`Run “${j.name}” now? It pulls ${j.stack}'s images${j.apply ? ' and deploys them, which restarts its containers' : ', without deploying them'}.`, 'Run now')) return
    try { await api.runUpdateJob(j.id); toast(`${j.name}: done`); loadJobs() } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  const remove = async (j: UpdateJob) => {
    if (!await askYes(`Remove the schedule “${j.name}”? Nothing is updated or rolled back — the schedule simply stops.`, 'Remove')) return
    try { await api.deleteUpdateJob(j.id); toast('removed'); loadJobs() } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  const imgs = upd?.images ?? []
  const behind = imgs.filter((i) => i.state === 'update')
  // Updating is per stack — that is what compose can do — so an image row acts on whichever
  // stacks use it, and a queue runs them one at a time rather than recreating half the
  // machine at once.
  const [queue, setQueue] = useState<string[]>([])
  const [running, setRunning] = useState<{ stack: string; id: string } | null>(null)
  const [done, setDone] = useState<string[]>([])
  const stacksFor = useCallback((img: string) => upd?.stacks?.[img] ?? [], [upd])
  const startNext = useCallback(async (rest: string[]) => {
    const next = rest[0]
    if (!next) { setRunning(null); setQueue([]); return }
    setQueue(rest.slice(1))
    try { setRunning({ stack: next, id: (await api.stackAction(next, 'update')).id }) }
    catch (e) { toast(`${next}: ${e instanceof Error ? e.message : 'failed'}`, 'err'); setRunning(null); startNext(rest.slice(1)) }
  }, [])
  const updateStacks = async (names: string[], what: string) => {
    const uniq = [...new Set(names)]
    if (uniq.length === 0) { toast('nothing here names a stack this console can deploy', 'info'); return }
    if (!await askYes(`Update ${what}?\n\nThis pulls the newer image${uniq.length > 1 ? 's' : ''} and recreates ${uniq.length === 1 ? `the ${uniq[0]} stack` : `these stacks, one at a time: ${uniq.join(', ')}`}. Their containers restart.`, 'Update')) return
    setDone([]); startNext(uniq)
  }
  const behindStacks = [...new Set(behind.flatMap((i) => stacksFor(i.image)))]
  return (
    <div className="svc-detail">
      <div className="dash-bar"><h2 className="section">What is out of date</h2><span className="spacer" />
        <span className="muted small">{upd?.checked ? `checked ${new Date(upd.checked).toLocaleString()}` : 'not checked yet'}</span>
        {behindStacks.length > 0 && <Btn className="btn on" disabled={!!running} onClick={() => updateStacks(behindStacks, `everything out of date — ${behind.length} image${behind.length === 1 ? '' : 's'} across ${behindStacks.length} stack${behindStacks.length === 1 ? '' : 's'}`)}>Update all · {behind.length}</Btn>}
        <Btn className="btn" busy={busy} onClick={check}>Check now</Btn>
      </div>
      <div className="muted small">Each check is one request per image, which registries count — so it happens when asked, and the answer is kept for six hours.</div>
      {imgs.length === 0 && <div className="notice" style={{ marginTop: 8 }}>Nothing checked yet on this machine.</div>}
      {imgs.length > 0 && (
        <section className="card" style={{ marginTop: 8 }}>
          {behind.length === 0 && <div className="muted">Every image that could be checked is current.</div>}
          <div className="upd-list">
          {imgs.filter((i) => i.state !== 'current').map((i) => (
            <div className="row" key={i.image}>
              <span className="k mono small">{i.image}</span>
              <span className={i.state === 'update' ? 'warn-text' : 'muted small'}>{done.some((d) => stacksFor(i.image).includes(d)) ? <span className="good">updated just now</span> : i.state === 'update' ? 'a newer image is published' : i.detail || 'unknown'}</span>
              <span className="muted small">{stacksFor(i.image).join(' ')}</span>
              {i.state === 'update' && !done.some((d) => stacksFor(i.image).includes(d)) && (
                <Btn className="btn" disabled={!!running || stacksFor(i.image).length === 0} title={stacksFor(i.image).length ? `pull it and recreate ${stacksFor(i.image).join(', ')}` : 'no stack here uses this image'}
                  onClick={() => updateStacks(stacksFor(i.image), `${i.image}`)}>Update</Btn>
              )}
            </div>
          ))}
          </div>
          {behind.length === 0 && imgs.some((i) => i.state === 'current') && <div className="muted small" style={{ marginTop: 6 }}>{imgs.filter((i) => i.state === 'current').length} current.</div>}
          {running && (
            <div style={{ marginTop: 8 }}>
              <div className="sub-head"><span className="dname">Updating {running.stack}</span><span className="spacer" /><span className="muted small">{queue.length > 0 ? `${queue.length} to follow` : 'last one'}</span></div>
              <TaskOutput id={running.id} onDone={(t) => {
                if (t.exit_code === 0 && !t.error) setDone((d) => [...d, running.stack])
                else toast(`${running.stack}: update failed`, 'err')
                startNext(queue)
              }} />
            </div>
          )}
        </section>
      )}
      <div className="dash-bar" style={{ marginTop: 14 }}><h2 className="section">Schedules</h2><span className="spacer" />
        <Btn className="btn on" onClick={() => setEditing({ cron: '0 3 * * *', apply: true, enabled: true })}>+ New schedule</Btn>
      </div>
      <div className="muted small">Written to <span className="mono">{file || '/etc/cron.d/hamlab-docker'}</span>, which Ham Lab owns — so they keep running whether or not this console is up.</div>
      {jobs === null && <Loading label="Reading…" />}
      {jobs?.length === 0 && <div className="notice" style={{ marginTop: 8 }}>No schedule yet. Nothing on this machine updates itself.</div>}
      <div className="cards" style={{ marginTop: 8 }}>
        {(jobs ?? []).map(({ job, next, last }) => (
          <section className="card" key={job.id}>
            <div className="sub-head">
              <span className={'dot ' + (last?.error ? 'bad' : job.enabled ? 'ok' : '')} />
              <span className="dname">{job.name}</span>
              {!job.enabled && <span className="badge">paused</span>}
              <span className="spacer" />
              <span className="muted small">{job.enabled && next ? new Date(next).toLocaleString() : 'not scheduled'}</span>
            </div>
            <div className="props" style={{ marginTop: 6 }}>
              <div><span className="k">Stack</span><span>{job.stack}{job.services?.length ? ` · ${job.services.join(' ')}` : ''}</span></div>
              <div><span className="k">Runs</span><span>{describeCron(job.cron) || <span className="mono small">{job.cron}</span>}{describeCron(job.cron) && <span className="muted mono small"> · {job.cron}</span>}</span></div>
              <div><span className="k">Does</span><span>{job.apply ? 'pulls and deploys' : 'pulls only — you deploy'}</span></div>
              <div><span className="k">Last run</span><span>{last?.at ? new Date(last.at).toLocaleString() : 'never'}</span></div>
            </div>
            {last?.error && <div className="error" style={{ marginTop: 6 }}>{last.error}</div>}
            <div className="svc-actions">
              <Btn className="btn on" onClick={() => run(job)}>Run now</Btn>
              <button className="btn" onClick={() => setEditing(job)}>Edit</button>
              <Btn className="btn" onClick={async () => { await api.saveUpdateJob({ ...job, enabled: !job.enabled }); loadJobs() }}>{job.enabled ? 'Pause' : 'Resume'}</Btn>
              <Btn className="btn danger" onClick={() => remove(job)}>Remove</Btn>
            </div>
          </section>
        ))}
      </div>
      {editing && <JobDialog j={editing} stacks={stacks} onClose={() => setEditing(null)} onSaved={() => { setEditing(null); loadJobs() }} />}
    </div>
  )
 }
 function JobDialog({ j, stacks, onClose, onSaved }: { j: Partial<UpdateJob>; stacks: string[]; onClose: () => void; onSaved: () => void }) {
  const [f, setF] = useState<Partial<UpdateJob>>(j)
  const [svc, setSvc] = useState((j.services ?? []).join(' '))
  const [err, setErr] = useState('')
  const [busy, setBusy] = useState(false)
  useEscape(onClose)
  const set = (k: keyof UpdateJob, v: unknown) => setF((x) => ({ ...x, [k]: v }))
  const save = async (e: FormEvent) => {
    e.preventDefault(); setErr(''); setBusy(true)
    try { await api.saveUpdateJob({ ...f, services: svc.split(/\s+/).filter(Boolean) }); onSaved() }
    catch (e) { setErr(e instanceof Error ? e.message : 'failed'); setBusy(false) }
  }
  return (
    <div className="modal-back" onClick={onClose}>
      <form className="modal" onClick={(e) => e.stopPropagation()} onSubmit={save}>
        <h3>{f.id ? 'Edit schedule' : 'Keep a stack up to date'}</h3>
        <label className="muted">Stack
          <select value={f.stack ?? ''} onChange={(e) => set('stack', e.target.value)} required>
            <option value="">choose…</option>
            {stacks.map((s) => <option key={s} value={s}>{s}</option>)}
          </select>
        </label>
        <label className="muted">Name — optional
          <input value={f.name ?? ''} onChange={(e) => set('name', e.target.value)} placeholder={f.stack ?? 'nightly'} maxLength={64} />
        </label>
        <label className="muted">Only these services — optional, space separated; blank means the whole stack
          <input className="mono" value={svc} onChange={(e) => setSvc(e.target.value)} placeholder="all of them" spellCheck={false} />
        </label>
        <label className="muted">When — a cron expression
          <input className="mono" value={f.cron ?? ''} onChange={(e) => set('cron', e.target.value)} placeholder="0 3 * * *" spellCheck={false} required />
        </label>
        <label className="muted" style={{ flexDirection: 'row', alignItems: 'center', gap: 8 }}>
          <input type="checkbox" checked={!!f.apply} onChange={(e) => set('apply', e.target.checked)} style={{ width: 'auto' }} />
          Deploy the new image, not just fetch it
        </label>
        <label className="muted" style={{ flexDirection: 'row', alignItems: 'center', gap: 8 }}>
          <input type="checkbox" checked={!!f.enabled} onChange={(e) => set('enabled', e.target.checked)} style={{ width: 'auto' }} />
          Enabled
        </label>
        <div className="notice">
          {f.apply
            ? 'Deploying restarts the stack’s containers. On something that carries DNS or ingress, pick an hour you are asleep.'
            : 'Pulling only means the new image is on the machine and the list shows it as available — nothing restarts until you deploy.'}
        </div>
        {err && <div className="error">{err}</div>}
        <div className="svc-actions">
          <button type="submit" className="btn on" disabled={busy || !f.stack || !f.cron}>Save</button>
          <span className="spacer" />
          <button type="button" className="btn" onClick={onClose}>Cancel</button>
        </div>
      </form>
    </div>
  )
 }
@@ -147,13 +147,13 @@ export default function Files({ onOpenInTerminal }: { onOpenInTerminal?: (cmd: s
    })
  // ---- actions ----
-  const act = async (label: string, fn: () => Promise<unknown>) => {
+  const act = async (label: string, fn: () => Promise<unknown>): Promise<boolean> => {
-    try { await fn(); toast(label); await load() } catch (e) { toast(`${label} — ${e instanceof Error ? e.message : 'failed'}`, 'err') }
+    try { await fn(); toast(label); await load(); return true } catch (e) { toast(`${label} — ${e instanceof Error ? e.message : 'failed'}`, 'err'); return false }
  }
  // A name is one path segment. ("../x" and "a/b" would quietly act somewhere else.)
  const okName = (n?: string | null): n is string => { const t = n?.trim(); if (!t) return false; if (t.includes('/') || t === '.' || t === '..') { toast('a name can\'t contain / or be . or .. — use Move… to put it elsewhere', 'err'); return false } return true }
  const mkdir = async () => { const n = await askText('New folder name'); if (okName(n)) act(`created ${n.trim()}`, () => api.fileMkdir(join(path, n.trim()), sudo)) }
-  const touch = async () => { const n = await askText('New file name'); if (okName(n)) act(`created ${n.trim()}`, () => api.fileTouch(join(path, n.trim()), sudo)).then(() => setOpen(join(path, n.trim()))) }
+  const touch = async () => { const n = await askText('New file name'); if (okName(n) && await act(`created ${n.trim()}`, () => api.fileTouch(join(path, n.trim()), sudo))) setOpen(join(path, n.trim())) } // opened only once it exists
  const rename = async (e: FileEntry) => { const n = await askText('Rename to', e.name); if (okName(n) && n.trim() !== e.name) act(`renamed to ${n.trim()}`, () => api.fileRename(join(path, e.name), join(path, n.trim()), sudo)) }
  const move = async (e: FileEntry) => { const n = await askText('Move to (absolute path)', join(path, e.name)); if (n?.trim() && n.trim() !== join(path, e.name)) act(`moved to ${n.trim()}`, () => api.fileRename(join(path, e.name), n.trim(), sudo)) }
  const chmod = async (e: FileEntry) => { const n = await askText(`Permissions for ${e.name} (octal)`, e.perm); if (n?.trim() && n.trim() !== e.perm) act(`chmod ${n.trim()} ${e.name}`, () => api.fileChmod(join(path, e.name), n.trim(), sudo)) }
@@ -0,0 +1,262 @@
 // Hardware: what this machine is made of — the board, the processor, the memory modules in
 // their slots, the drives and the graphics. None of it moves, so the page is a read rather
 // than a monitor: what is live (temperature, usage, SMART) belongs to Dashboard and Disks.
 import { useCallback, useEffect, useState } from 'react'
 import { api, Hardware as HW, HWBoard, HWDrive, HWGPU, HWModule } from './api'
 import { fmtBytes } from './fmt'
 import { Btn, Loading } from './ui'
 const KIND: Record<string, string> = { nvme: 'NVMe', ssd: 'SSD', hdd: 'HDD', usb: 'USB', emmc: 'eMMC', sd: 'SD', other: 'disk' }
 const ghz = (mhz?: number) => (mhz ? `${(mhz / 1000).toFixed(mhz >= 1000 ? 2 : 1)} GHz` : '')
 // A row of the "key: value" kind, dropped entirely when there is no value — a machine that
 // doesn't report its serial should show one fewer line, not an empty one.
 function P({ k, v, mono }: { k: string; v?: string | number | null | false; mono?: boolean }) {
  if (v === undefined || v === null || v === false || v === '' || v === 0) return null
  return <div><span className="k">{k}</span><span className={mono ? 'mono' : undefined} title={String(v)}>{v}</span></div>
 }
 function Module({ m }: { m: HWModule }) {
  if (m.empty) {
    return (
      <div className="sub-card">
        <div className="sub-head"><span className="dname muted">{m.locator}</span><span className="spacer" /><span className="badge">empty</span></div>
      </div>
    )
  }
  const speed = m.running_mts && m.speed_mts && m.running_mts !== m.speed_mts
    ? `${m.running_mts} MT/s (rated ${m.speed_mts})` // the board is running it slower than it could
    : m.running_mts || m.speed_mts ? `${m.running_mts || m.speed_mts} MT/s` : ''
  return (
    <div className="sub-card">
      <div className="sub-head">
        <span className="dname">{fmtBytes(m.size ?? 0)} {m.type}</span>
        <span className="spacer" />
        <span className="model mono">{m.locator}</span>
      </div>
      <div className="props">
        <P k="Maker" v={m.maker} />
        <P k="Form" v={m.form_factor} />
        <P k="Part" v={m.part} mono />
        <P k="Speed" v={speed} />
        <P k="Rank" v={m.rank ? `${m.rank}R` : ''} />
        <P k="Voltage" v={m.voltage} />
        <P k="Serial" v={m.serial} mono />
      </div>
    </div>
  )
 }
 function Drive({ d }: { d: HWDrive }) {
  return (
    <div className="sub-card">
      <div className="sub-head">
        <span className="dname">{d.model || d.name}</span>
        <span className="spacer" />
        <span className="badge">{KIND[d.kind] ?? d.kind}</span>
      </div>
      <div className="props">
        <P k="Size" v={fmtBytes(d.size)} />
        <P k="Device" v={`/dev/${d.name}`} mono />
        <P k="Bus" v={d.tran ? d.tran.toUpperCase() : ''} />
        <P k="Spins" v={d.rotational ? 'yes — mechanical' : 'no — solid state'} />
        <P k="Firmware" v={d.firmware} mono />
        <P k="Serial" v={d.serial} mono />
        <P k="Vendor" v={d.vendor} />
      </div>
    </div>
  )
 }
 function Gpu({ g }: { g: HWGPU }) {
  return (
    <div className="sub-card">
      <div className="sub-head">
        <span className="dname">{g.model}</span>
        <span className="spacer" />
        {g.class && <span className="badge">{g.class}</span>}
      </div>
      <div className="props">
        <P k="Vendor" v={g.vendor} />
        <P k="Driver" v={g.driver} mono />
        <P k="Memory" v={g.memory ? fmtBytes(g.memory) : ''} />
        <P k="Slot" v={g.slot} mono />
        <P k="Node" v={g.node} mono />
      </div>
    </div>
  )
 }
 // The board and the firmware on it. Everything here is SMBIOS, so a machine without one
 // — a Mac, a Pi — gets a card that says as much instead of a column of blanks.
 function Motherboard({ b }: { b: HWBoard }) {
  const named = b.maker || b.product || b.bios
  // A board that reports every slot "Available" is not telling us anything — mini PCs
  // copy the slot table out of a reference design and list slots the case has no room
  // for. Only worth showing when the firmware actually tracks what is plugged in.
  const slots = (b.slots ?? []).some((s) => s.usage === 'In Use') ? (b.slots ?? []) : []
  return (
    <section className="card">
      <h2>Motherboard</h2>
      {!named && <div className="empty">This machine has no SMBIOS table, so its firmware doesn't say what board it is.</div>}
      {named && (
        <div className="props">
          <P k="Maker" v={b.maker} />
          <P k="Model" v={b.product} />
          <P k="Version" v={b.version} />
          <P k="Kind" v={b.type} />
          <P k="Chassis" v={[b.chassis_maker, b.chassis].filter(Boolean).join(' ')} />
          <P k="In chassis" v={b.location} />
          <P k="Serial" v={b.serial} mono />
          <P k="Asset tag" v={b.asset_tag} mono />
        </div>
      )}
      {(b.bios || b.bios_date) && (
        <div className="props" style={{ marginTop: 8 }}>
          <P k="Firmware" v={b.bios} />
          <P k="Revision" v={b.bios_revision} />
          <P k="Dated" v={b.bios_date} />
          <P k="ROM" v={b.rom_size} />
          <P k="Boots" v={b.uefi ? 'UEFI' : ''} />
        </div>
      )}
      {slots.length > 0 && (
        <div style={{ marginTop: 8 }}>
          <div className="sub-head"><span className="dname">Expansion slots</span><span className="spacer" /><span className="muted small">as the firmware lists them</span></div>
          <table className="dtable">
            <thead><tr><th>Slot</th><th>Type</th><th>What is in it</th></tr></thead>
            <tbody>{slots.map((s, i) => (
              <tr key={s.designation + i}>
                <td className="mono small">{s.designation}</td>
                <td className="mono small">{[s.type, s.length && s.length !== 'Other' ? s.length.toLowerCase() : ''].filter(Boolean).join(' · ')}</td>
                <td className="small" title={s.bus}>{s.usage === 'In Use' ? (s.device ?? s.bus ?? 'in use') : <span className="muted">{(s.usage ?? '').toLowerCase() || '—'}</span>}</td>
              </tr>
            ))}</tbody>
          </table>
        </div>
      )}
    </section>
  )
 }
 export default function Hardware() {
  const [hw, setHw] = useState<HW | null>(null)
  const [err, setErr] = useState('')
  const [busy, setBusy] = useState(false)
  const load = useCallback(async (refresh = false) => {
    if (refresh) setBusy(true)
    try { setHw(await api.hardware(refresh)); setErr('') } catch (e) { setErr(e instanceof Error ? e.message : 'failed') } finally { setBusy(false) }
  }, [])
  useEffect(() => { load() }, [load])
  if (err) return <div className="dash"><div className="notice">{err === 'not found' ? "This machine's Ham Lab is older than the Hardware tab. Update it from Machines." : err}</div></div>
  if (!hw) return <div className="dash"><Loading label="Reading the machine…" /></div>
  const { system: sys, board, cpu, memory: mem } = hw
  const drives = hw.drives ?? []
  const gpus = hw.gpus ?? []
  const capacity = drives.reduce((n, d) => n + d.size, 0)
  const threads = cpu.threads ?? 0
  const cores = cpu.cores ?? 0
  return (
    <div className="dash">
      <div className="dash-bar">
        <span className="status"><b>{sys.model || 'this machine'}</b></span>
        {sys.virt && <span className="badge" title="this is a virtual machine, so the parts below are what the host handed it">{sys.virt}</span>}
        <span className="spacer" style={{ flex: 1 }} />
        <Btn className="btn" busy={busy} onClick={() => load(true)}>Re-read</Btn>
      </div>
      <div className="kpis">
        <div className="tile">
          <div className="label">Processor</div>
          <div className="value" title={cpu.model} style={(cpu.model?.length ?? 0) > 20 ? { fontSize: 16 } : undefined}>{cpu.model}</div>
          <div className="sub">{[cores ? `${cores} core${cores === 1 ? '' : 's'}` : '', threads && threads !== cores ? `${threads} threads` : '', cpu.max_mhz ? `up to ${ghz(cpu.max_mhz)}` : ''].filter(Boolean).join(' · ')}</div>
        </div>
        <div className="tile">
          <div className="label">Memory</div>
          <div className="value">{fmtBytes(mem.installed || mem.total)}</div>
          <div className="sub">{mem.slots ? `${mem.filled ?? 0} of ${mem.slots} slot${mem.slots === 1 ? '' : 's'} filled` : 'slots unknown'}{mem.ecc ? ` · ${mem.ecc}` : ''}</div>
        </div>
        <div className="tile">
          <div className="label">Storage</div>
          <div className="value">{fmtBytes(capacity)}</div>
          <div className="sub">{drives.length} drive{drives.length === 1 ? '' : 's'}</div>
        </div>
        <div className="tile">
          <div className="label">Graphics</div>
          <div className="value" title={gpus[0]?.model} style={(gpus[0]?.model?.length ?? 0) > 20 ? { fontSize: 16 } : undefined}>{gpus[0]?.model ?? 'none found'}</div>
          <div className="sub">{gpus.length > 1 ? `${gpus.length} adapters` : gpus[0]?.driver ? `${gpus[0].driver} driver` : ''}</div>
        </div>
      </div>
      {(hw.notes ?? []).map((n, i) => <div className="notice" key={i}>{n}</div>)}
      <div className="cards">
        <section className="card">
          <h2>Processor</h2>
          <div className="props">
            <P k="Model" v={cpu.model} />
            <P k="Vendor" v={cpu.vendor} />
            <P k="Cores" v={cores ? `${cores}${cpu.threads_per_core && cpu.threads_per_core > 1 ? ` · ${cpu.threads_per_core} threads each` : ''}` : ''} />
            <P k="Clock" v={cpu.min_mhz && cpu.max_mhz ? `${ghz(cpu.min_mhz)} – ${ghz(cpu.max_mhz)}` : ghz(cpu.max_mhz)} />
            <P k="Socket" v={(cpu.socket_names ?? []).join(', ')} />
            <P k="Arch" v={cpu.arch} />
            <P k="Family" v={[cpu.family && `family ${cpu.family}`, cpu.model_id && `model ${cpu.model_id}`, cpu.stepping && `stepping ${cpu.stepping}`].filter(Boolean).join(' · ')} />
            <P k="Microcode" v={cpu.microcode} mono />
            <P k="Virt" v={cpu.virtualization} />
          </div>
          {(cpu.caches ?? []).length > 0 && (
            <div className="props" style={{ marginTop: 8 }}>
              {(cpu.caches ?? []).map((c) => <P key={c.level} k={c.level} v={c.size} />)}
            </div>
          )}
          {(cpu.flags ?? []).length > 0 && (
            <div className="chips" style={{ marginTop: 8 }}>{(cpu.flags ?? []).map((f) => <span className="badge" key={f}>{f}</span>)}</div>
          )}
        </section>
        <section className="card">
          <h2>Memory</h2>
          <div className="props">
            <P k="Installed" v={mem.installed ? fmtBytes(mem.installed) : ''} />
            <P k="Usable" v={fmtBytes(mem.total)} />
            <P k="Board takes" v={mem.max_capacity ? fmtBytes(mem.max_capacity) : ''} />
            <P k="ECC" v={mem.ecc} />
          </div>
          <div style={{ marginTop: 8 }}>
            {(mem.modules ?? []).map((m, i) => <Module key={m.locator + i} m={m} />)}
            {(mem.modules ?? []).length === 0 && <div className="empty">The firmware doesn't say what is in the slots.</div>}
          </div>
        </section>
        <section className="card">
          <h2>Drives</h2>
          {drives.map((d) => <Drive key={d.name} d={d} />)}
          {drives.length === 0 && <div className="empty">No drives found.</div>}
        </section>
        <section className="card">
          <h2>Graphics</h2>
          {gpus.map((g, i) => <Gpu key={(g.slot ?? '') + i} g={g} />)}
          {gpus.length === 0 && <div className="empty">No display adapter found.</div>}
        </section>
        <Motherboard b={board} />
        <section className="card">
          <h2>System</h2>
          <div className="props">
            <P k="Model" v={sys.model} />
            <P k="Family" v={sys.family} />
            <P k="Serial" v={sys.serial} mono />
            <P k="Kernel" v={sys.kernel} mono />
            <P k="Arch" v={sys.arch} />
            <P k="Virtual" v={sys.virt} />
          </div>
        </section>
      </div>
    </div>
  )
 }
@@ -1,4 +1,5 @@
 import { useCallback, useEffect, useRef, useState } from 'react'
 import { describeCron } from './crontext'
 import { api, CronJob, CronOverview, CronRun, CronTimer, LogEntry } from './api'
 import { Logs } from './Services'
 import { askYes, Btn, CloseBtn, Loading, Spinner, toast } from './ui'
@@ -13,22 +14,7 @@ function rel(ms: number): string {
  return s >= 0 ? `in ${txt}` : `${txt} ago`
 }
-// Plain-English reading of the common five-field shapes; anything else shows the raw expression.
+const describe = (expr: string) => describeCron(expr) || expr
 function describe(expr: string): string {
  const alias: Record<string, string> = { '@reboot': 'at boot', '@hourly': 'every hour', '@daily': 'every day at 00:00', '@midnight': 'every day at 00:00', '@weekly': 'every Sunday at 00:00', '@monthly': 'on the 1st at 00:00', '@yearly': 'every 1 Jan', '@annually': 'every 1 Jan' }
  if (alias[expr]) return alias[expr]
  const f = expr.split(/\s+/)
  if (f.length !== 5) return expr
  const [mi, h, dom, mon, dow] = f
  const days = ['Sun', 'Mon', 'Tue', 'Wed', 'Thu', 'Fri', 'Sat', 'Sun']
  const at = /^\d+$/.test(mi) && /^\d+$/.test(h) ? `at ${h.padStart(2, '0')}:${mi.padStart(2, '0')}` : /^\d+$/.test(mi) && h === '*' ? `every hour at :${mi.padStart(2, '0')}` : mi.startsWith('*/') && h === '*' ? `every ${mi.slice(2)} min` : null
  if (!at) return expr
  const dowTxt = dow === '*' ? '' : /^\d$/.test(dow) ? `every ${days[+dow]}` : `on days ${dow}`
  const domTxt = dom === '*' ? '' : /^\d+$/.test(dom) ? `on the ${dom}${['st', 'nd', 'rd'][+dom - 1] ?? 'th'}` : `on days ${dom}`
  const monTxt = mon === '*' ? '' : `in month ${mon}`
  const when = [dowTxt, domTxt, monTxt].filter(Boolean).join(', ')
  return (when ? `${when} ` : dow === '*' && dom === '*' && !at.startsWith('every') ? 'every day ' : '') + at
 }
 const srcLabel = (s: string) => (s.startsWith('crontab:') ? `${s.slice(8)}'s crontab` : s)
@@ -0,0 +1,187 @@
 // The KVM of the machine you are looking at: the device's own web page, shown in the tab.
 //
 // One per machine, and it follows the machine switcher — switch to `grid` and you get
 // grid's KVM. The address is held on the **hub**, not on the machine, because a KVM earns
 // its keep exactly when that machine is down and cannot answer for itself; the hub proxies
 // the device either way.
 //
 // Nothing here knows anything about any particular KVM. It is an iframe pointed at Ham
 // Lab's own proxy, so whatever web UI the device serves is what you get.
 //
 // On the keyboard: a page cannot capture keys and forward them into an iframe, and does not
 // need to. An iframe with focus receives them directly. What the browser itself swallows —
 // Escape, Tab, Ctrl-W, the function keys — is what `navigator.keyboard.lock()` releases,
 // and that needs fullscreen and only exists in Chromium. Everywhere else you get the plain
 // version, which is most keys, and the device's own on-screen buttons for the rest.
 import { FormEvent, useCallback, useEffect, useRef, useState } from 'react'
 import { KVM as Device, api } from './api'
 import { Btn, Loading, askYes, toast, useEscape } from './ui'
 // lock() releases the keys the browser would otherwise keep for itself. Chromium only; a
 // browser without it simply does not get the extra keys.
 type KeyboardLock = { lock?: (keys?: string[]) => Promise<void>; unlock?: () => void }
 const keyboardAPI = (): KeyboardLock | undefined => (navigator as unknown as { keyboard?: KeyboardLock }).keyboard
 // Fullscreen, across the three ways browsers spell it. Safari on a Mac wants the webkit
 // prefix; Safari on iOS has no fullscreen for an ordinary element at all — only for a
 // <video> — so there the page expands inside the viewport instead, which is the same thing
 // minus the browser's own chrome.
 type FSElement = HTMLElement & { webkitRequestFullscreen?: () => Promise<void> | void }
 type FSDocument = Document & { webkitFullscreenElement?: Element | null; webkitExitFullscreen?: () => Promise<void> | void }
 const nativeFullscreen = () => {
  if (typeof document === 'undefined') return false
  const el = document.documentElement as FSElement
  return typeof el.requestFullscreen === 'function' || typeof el.webkitRequestFullscreen === 'function'
 }
 const fullscreenElement = () => (document as FSDocument).fullscreenElement ?? (document as FSDocument).webkitFullscreenElement ?? null
 async function enterFullscreen(el: FSElement) {
  if (typeof el.requestFullscreen === 'function') return el.requestFullscreen()
  if (typeof el.webkitRequestFullscreen === 'function') return el.webkitRequestFullscreen()
  throw new Error('this browser has no fullscreen for a page element')
 }
 function exitFullscreen() {
  const d = document as FSDocument
  if (typeof d.exitFullscreen === 'function') return d.exitFullscreen()
  if (typeof d.webkitExitFullscreen === 'function') return d.webkitExitFullscreen()
 }
 export default function KVM({ machine, title }: { machine: string | null; title: string }) {
  const [kvm, setKvm] = useState<Device | null | undefined>(undefined)
  const [adding, setAdding] = useState(false)
  const [err, setErr] = useState('')
  const [full, setFull] = useState(false)   // native fullscreen is on
  const [expanded, setExpanded] = useState(false) // the fallback: filling the viewport in-page
  const box = useRef<HTMLDivElement>(null)
  const frame = useRef<HTMLIFrameElement>(null)
  const key = machine ?? '' // the hub is a machine too, and its key is the empty one
  const load = useCallback(async () => {
    try { setKvm((await api.kvms()).find((k) => k.machine === key) ?? null) }
    catch (e) { setErr(e instanceof Error ? e.message : 'failed') }
  }, [key])
  useEffect(() => { load() }, [load])
  // Fullscreen plus a keyboard lock is what makes Escape and Tab reach the machine rather
  // than the browser. Both are undone together when fullscreen ends, however it ends.
  useEffect(() => {
    const onChange = () => {
      const on = !!fullscreenElement()
      setFull(on)
      if (!on) keyboardAPI()?.unlock?.()
    }
    document.addEventListener('fullscreenchange', onChange)
    document.addEventListener('webkitfullscreenchange', onChange)
    return () => {
      document.removeEventListener('fullscreenchange', onChange)
      document.removeEventListener('webkitfullscreenchange', onChange)
      keyboardAPI()?.unlock?.()
    }
  }, [])
  // Expanding is in-page, so Escape is ours to handle.
  useEscape(() => setExpanded(false), expanded)
  const goBig = async () => {
    if (!nativeFullscreen()) { setExpanded(true); frame.current?.focus(); return }
    try {
      if (box.current) await enterFullscreen(box.current)
      await keyboardAPI()?.lock?.()
      frame.current?.focus()
    } catch {
      setExpanded(true) // refused (an iframe policy, a permissions policy): fill the page instead
      frame.current?.focus()
    }
  }
  const goSmall = () => { setExpanded(false); if (fullscreenElement()) exitFullscreen() }
  const remove = async () => {
    if (!kvm || !await askYes(`Forget ${title}'s KVM? The device itself is not touched — this only removes its address.`, 'Remove')) return
    try { await api.removeKVM(kvm.id); toast('removed'); await load() } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  if (err && kvm === undefined) return <div className="dash"><div className="error">{err}</div></div>
  if (kvm === undefined) return <div className="dash"><Loading label="Looking…" /></div>
  const locks = !!keyboardAPI()?.lock
  const big = full || expanded
  return (
    <div className="kvm">
      <div className="kvm-bar">
        <span className="dname">{title}</span>
        {kvm
          ? <span className="muted small mono kvm-url" title={kvm.url}>{kvm.name !== kvm.url ? `${kvm.name} · ` : ''}{kvm.url}</span>
          : <span className="muted small">no KVM</span>}
        <span className="spacer" />
        {kvm && <>
          <Btn className="btn" onClick={goBig} title={!nativeFullscreen() ? 'fill the screen — this browser has no true fullscreen for a page element' : locks ? 'fullscreen, and the browser stops keeping Escape, Tab and Ctrl-W for itself' : 'fullscreen — this browser will still keep some keys for itself'}>{nativeFullscreen() ? 'Fullscreen' : 'Expand'}</Btn>
          <a className="btn" href={kvm.url} target="_blank" rel="noreferrer" title="the device directly, outside Ham Lab — only works on the LAN">Open ↗</a>
          <Btn className="btn" onClick={() => setAdding(true)}>Change</Btn>
          <Btn className="btn danger" onClick={remove}>Remove</Btn>
        </>}
        {!kvm && <Btn className="btn on" onClick={() => setAdding(true)}>+ Add a KVM for {title}</Btn>}
      </div>
      {kvm ? (
        <div className={'kvm-frame' + (full ? ' full' : '') + (expanded ? ' expanded' : '')} ref={box}>
          <iframe ref={frame} key={kvm.id} src={`/kvm/${kvm.id}/`} title={kvm.name}
            allow="fullscreen; clipboard-read; clipboard-write" />
          {expanded && <button className="btn kvm-shrink" onClick={goSmall} title="back to the console">Done</button>}
        </div>
      ) : (
        <div className="svc-empty muted kvm-none">
          <div>
            <p>No KVM for <b>{title}</b>.</p>
            <p className="small">Add one with its address and its page appears here. The address is kept on the hub,
              so it still works when {title} does not — which is when you will want it.</p>
          </div>
        </div>
      )}
      {kvm && !big && (
        <div className="kvm-hint muted small">
          Click the picture, then type — the keys go to {title}.
          {locks ? ' Fullscreen also takes Escape, Tab and Ctrl-W back from the browser.' : ' This browser keeps Escape, Tab and Ctrl-W for itself; use the device’s own on-screen keys for those, or a Chromium browser.'}
        </div>
      )}
      {adding && <AddDialog machine={key} title={title} existing={kvm} onClose={() => setAdding(false)} onAdded={() => { setAdding(false); load() }} />}
    </div>
  )
 }
 function AddDialog({ machine, title, existing, onClose, onAdded }: { machine: string; title: string; existing: Device | null; onClose: () => void; onAdded: () => void }) {
  const [url, setUrl] = useState(existing?.url ?? '')
  const [name, setName] = useState(existing?.name ?? '')
  const [err, setErr] = useState('')
  const [busy, setBusy] = useState(false)
  useEscape(onClose)
  const save = async (e: FormEvent) => {
    e.preventDefault(); setErr(''); setBusy(true)
    try { await api.addKVM(name.trim(), url.trim(), machine); onAdded() }
    catch (e) { setErr(e instanceof Error ? e.message : 'failed'); setBusy(false) }
  }
  return (
    <div className="modal-back" onClick={onClose}>
      <form className="modal" onClick={(e) => e.stopPropagation()} onSubmit={save}>
        <h3>{existing ? `Change ${title}'s KVM` : `A KVM for ${title}`}</h3>
        <label className="muted">Address — whatever you would type into a browser
          <input autoFocus className="mono" value={url} onChange={(e) => setUrl(e.target.value)} placeholder="10.0.0.217" spellCheck={false} autoCapitalize="off" required />
        </label>
        <label className="muted">Name — optional; the address is used otherwise
          <input value={name} onChange={(e) => setName(e.target.value)} placeholder={`${title} KVM`} maxLength={64} />
        </label>
        <div className="muted small">
          Ham Lab serves the device from its own address, so a KVM that speaks only plain HTTP
          still works here, and it is behind this login rather than needing a second one. The
          address is kept on the hub, so it still reaches the device when {title} is down. The
          device itself is not changed and is never reachable from outside.
          {existing && <> Saving replaces the one there now — a machine has one KVM.</>}
        </div>
        {err && <div className="error">{err}</div>}
        <div className="svc-actions">
          <button type="submit" className="btn on" disabled={busy || !url.trim()}>{existing ? 'Save' : 'Add'}</button>
          <span className="spacer" />
          <button type="button" className="btn" onClick={onClose}>Cancel</button>
        </div>
      </form>
    </div>
  )
 }
@@ -2,6 +2,19 @@ import { FormEvent, useState } from 'react'
 import { api } from './api'
 import { Spinner } from './ui'
 // Drawn, not typed: 👁 is an emoji on phones, full colour and its own size. Outline strokes
 // match the glyphs in the rail and the power button.
 const EyeIcon = (
  <svg viewBox="0 0 16 16" width="17" height="17" fill="none" stroke="currentColor" strokeWidth="1.4" strokeLinecap="round" strokeLinejoin="round" aria-hidden="true">
    <path d="M1.2 8S3.9 3.6 8 3.6 14.8 8 14.8 8 12.1 12.4 8 12.4 1.2 8 1.2 8z" /><circle cx="8" cy="8" r="2.1" />
  </svg>
 )
 const EyeOffIcon = (
  <svg viewBox="0 0 16 16" width="17" height="17" fill="none" stroke="currentColor" strokeWidth="1.4" strokeLinecap="round" strokeLinejoin="round" aria-hidden="true">
    <path d="M1.2 8S3.9 3.6 8 3.6 14.8 8 14.8 8 12.1 12.4 8 12.4 1.2 8 1.2 8z" /><circle cx="8" cy="8" r="2.1" /><path d="M2.4 13.6 13.6 2.4" />
  </svg>
 )
 const USER_KEY = 'hamlab.user' // the last account signed in on this device, offered again
 export default function Login({ onLogin }: { onLogin: () => void }) {
@@ -9,6 +22,9 @@ export default function Login({ onLogin }: { onLogin: () => void }) {
  const [pw, setPw] = useState('')
  const [err, setErr] = useState('')
  const [busy, setBusy] = useState(false)
  // Shown on request: a generated password is typed by hand off a message, and a wrong
  // character looks exactly like a wrong password from here.
  const [show, setShow] = useState(false)
  async function submit(e: FormEvent) {
    e.preventDefault()
@@ -26,8 +42,17 @@ export default function Login({ onLogin }: { onLogin: () => void }) {
        <img className="login-logo" src="/icons/logo.png" alt="Ham Lab" width={176} height={176} />
        <input type="text" autoFocus={!user} autoComplete="username" placeholder="user" spellCheck={false} autoCapitalize="off" autoCorrect="off"
          value={user} onChange={(e) => setUser(e.target.value)} disabled={busy} />
-        <input type="password" autoFocus={!!user} autoComplete="current-password" placeholder="password"
+        <div className="pwfield">
-          value={pw} onChange={(e) => setPw(e.target.value)} disabled={busy} />
+          {/* revealed, the field is a text field, so the keyboard would otherwise offer to
              capitalise and correct what is typed into it */}
          <input type={show ? 'text' : 'password'} autoFocus={!!user} autoComplete="current-password" placeholder="password"
            spellCheck={false} autoCapitalize="off" autoCorrect="off"
            value={pw} onChange={(e) => setPw(e.target.value)} disabled={busy} />
          <button type="button" className="pweye" onClick={() => setShow(!show)} disabled={busy}
            aria-label={show ? 'hide the password' : 'show the password'} aria-pressed={show} title={show ? 'hide' : 'show'}>
            {show ? EyeOffIcon : EyeIcon}
          </button>
        </div>
        <button disabled={busy || !pw}>{busy && <Spinner />}{busy ? 'Signing in…' : 'Sign in'}</button>
        {err && <div className="error">{err}</div>}
      </form>
@@ -54,7 +54,10 @@ export default function Machines({ hub, onChanged, onOpen }: { hub: Me; onChange
    try {
      let res = await api.updateAllMachines()
      const busy = res.filter((r) => !r.ok && (r.error ?? '').startsWith('busy:'))
-      if (busy.length && await askYes(`${busy.map((r) => r.name).join(', ')}: ${busy[0].error?.slice(6)}\n\nForce those too?`)) res = res.map((r) => r).concat(await api.updateAllMachines(true))
+      if (busy.length && await askYes(`${busy.map((r) => r.name).join(', ')}: ${busy[0].error?.slice(6)}\n\nForce those too?`)) {
        const forced = await api.updateAllMachines(true)
        res = res.filter((r) => !forced.some((f) => f.name === r.name)).concat(forced) // one line per machine, the forced answer winning
      }
      for (const r of res) toast(r.ok ? `${r.name} → ${r.version}` : `${r.name}: ${r.error}`, r.ok ? 'ok' : 'err')
      setTimeout(load, 4000)
    } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
@@ -0,0 +1,201 @@
 // Network: how this machine reaches the internet, and how fast the line actually is.
 //
 // The speed test runs on the machine in the switcher, not in this browser — so it measures
 // that machine's line, and it keeps running if the phone is locked or the tab is closed. What
 // is drawn here is what the daemon reports about a test it is running itself.
 import { useCallback, useEffect, useMemo, useState } from 'react'
 import { api, Iface, NetworkInfo, SpeedResult, SpeedRun } from './api'
 import { Sensor, fmtValue, useSensorStream } from './sensors'
 import { fmtBytes } from './fmt'
 import { Btn, Loading, toast } from './ui'
 import { useRefreshMs } from './prefs'
 import Sparkline from './Sparkline'
 // Docker's per-network bridges, and the veth ends inside them, are real interfaces and
 // nobody wants to read nineteen of them. They fold away unless asked for.
 const scenery = (f: Iface) =>
  !f.default && (/^br-[0-9a-f]{12}$/.test(f.name) || /^(veth|docker)/.test(f.name) || f.kind === 'loopback')
 const KIND: Record<string, string> = { ethernet: 'wired', wifi: 'Wi-Fi', bridge: 'bridge', bond: 'bond', loopback: 'loopback', virtual: 'virtual', vlan: 'VLAN', wlan: 'Wi-Fi', wireguard: 'VPN tunnel', none: 'tunnel' }
 // A WireGuard interface is listed here because it is an interface and carries traffic, but
 // everything about it that is worth knowing — who is on it, when they last spoke — belongs to
 // the VPN tab, and is only there.
 const isTunnel = (f: Iface) => f.kind === 'wireguard' || /^(wg|tun|tap)[0-9]/.test(f.name)
 const goVPN = () => window.dispatchEvent(new CustomEvent('hamlab:view', { detail: 'vpn' }))
 // Where the bytes came from. The usual source is anycast and its edge is the interesting
 // part; a fallback source is somewhere fixed, and then *which* one is the interesting part,
 // because a mirror far away reads low and the figure should not pretend otherwise.
 const where = (r: SpeedResult) =>
  r.source && r.source !== 'cloudflare' ? ` from ${r.source}` : r.colo ? ` via ${r.colo}` : ''
 // A round number of megabits, which is how a line is sold and how it should be read.
 const mbps = (v: number) => (v >= 100 ? v.toFixed(0) : v.toFixed(1))
 const ago = (iso: string) => {
  const s = Math.max(0, (Date.now() - new Date(iso).getTime()) / 1000)
  if (s < 60) return 'just now'
  if (s < 3600) return `${Math.floor(s / 60)} min ago`
  if (s < 86400) return `${Math.floor(s / 3600)} h ago`
  return `${Math.floor(s / 86400)} d ago`
 }
 // The four figures a speed test produces, big enough to read from across the room.
 function Figures({ r, live }: { r?: SpeedResult; live?: SpeedRun }) {
  const running = !!live?.running
  const phase = live?.phase
  const downNow = running && phase === 'download' ? live!.mbps : undefined
  const upNow = running && phase === 'upload' ? live!.mbps : undefined
  return (
    <div className="speedgrid">
      <Figure label="Download" unit="Mbps" value={downNow ?? r?.down_mbps} live={downNow !== undefined} waiting={running && phase === 'latency'} />
      <Figure label="Upload" unit="Mbps" value={upNow ?? r?.up_mbps} live={upNow !== undefined} waiting={running && phase !== 'upload'} />
      <Figure label="Latency" unit="ms" value={r?.latency_ms} waiting={running && phase === 'latency'} />
      <Figure label="Jitter" unit="ms" value={r?.jitter_ms} waiting={running && phase === 'latency'} />
    </div>
  )
 }
 function Figure({ label, unit, value, live, waiting }: { label: string; unit: string; value?: number; live?: boolean; waiting?: boolean }) {
  return (
    <div className={'speedfig' + (live ? ' live' : '')}>
      <div className="label">{label}</div>
      <div className="big">{value === undefined || value === null ? (waiting ? '·' : '—') : mbps(value)}<i>{unit}</i></div>
    </div>
  )
 }
 export default function Network() {
  const [d, setD] = useState<NetworkInfo | null>(null)
  const [err, setErr] = useState('')
  const [all, setAll] = useState(false)
  const [busy, setBusy] = useState(false)
  const rate = useRefreshMs()
  const load = useCallback(async () => {
    try { setD(await api.network()); setErr('') } catch (e) { setErr(e instanceof Error ? e.message : 'failed') }
  }, [])
  useEffect(() => { load() }, [load])
  const run = d?.run
  const running = !!run?.running
  // While a test is running the state is polled hard, because the needle should move; at rest
  // the page is a read and asks again only now and then.
  useEffect(() => {
    const every = running ? Math.max(400, Math.min(rate, 1000)) : Math.max(5000, rate)
    const t = setInterval(() => {
      if (document.visibilityState !== 'visible') return
      if (running) api.speedtest().then((r) => setD((old) => (old ? { ...old, run: r } : old))).catch(() => {})
      else load()
    }, every)
    return () => clearInterval(t)
  }, [running, rate, load])
  // The run that just ended brings a result and a history entry with it.
  const phase = run?.phase
  useEffect(() => { if (phase === 'done') load() }, [phase, load])
  const live = useSensorStream<{ n: Record<string, Sensor> }>({ n: 'net.*' }, { everyMs: Math.max(1000, rate) })
  const sens = live.data?.n ?? {}
  const rateOf = (name: string, dir: 'rx' | 'tx') => sens[`net.${name.toLowerCase().replace(/[^a-z0-9_-]+/g, '_')}.${dir}`]?.value
  const ifaces = d?.report.ifaces ?? []
  const shown = all ? ifaces : ifaces.filter((f) => !scenery(f))
  const hidden = ifaces.length - shown.length
  const hist = useMemo(() => (d?.history ?? []).filter((r) => !r.error), [d])
  const last = run?.result && !run.result.error ? run.result : hist[hist.length - 1]
  const start = async () => {
    setBusy(true)
    try { const r = await api.startSpeedtest(); setD((old) => (old ? { ...old, run: r } : old)) }
    catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
    finally { setBusy(false) }
  }
  const cancel = async () => {
    try { await api.cancelSpeedtest(); load() } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  if (!d) return <div className="dash">{err ? <div className="notice">{err === 'not found' ? "This machine's Ham Lab is older than the Network tab. Update it from Machines." : err}</div> : <Loading label="Reading the network…" />}</div>
  const o = d.outside
  const failed = run?.result?.error
  return (
    <div className="dash net">
      <div className="dash-bar"><h2 className="section">Internet speed</h2><span className="spacer" />
        <span className="muted small">{last ? `measured ${ago(last.at)}` : 'not measured yet'}</span></div>
      <section className="card speedtest">
        <Figures r={last} live={run} />
        {running && (
          <div className="speedbar">
            <div className="meter"><div className="fill" style={{ width: `${Math.round((run!.done || 0) * 100)}%` }} /></div>
            <span className="muted small">{run!.phase === 'latency' ? 'measuring the round trip' : run!.phase === 'download' ? 'pulling down' : 'pushing up'} · {fmtBytes(run!.bytes || 0)} moved</span>
          </div>
        )}
        <div className="svc-actions">
          {running
            ? <><Btn className="btn danger" onClick={cancel}>Stop</Btn><span className="muted small">running on {d.report.host} — it carries on if you close this</span></>
            : <><Btn className="btn on" busy={busy} onClick={start}>Run speed test</Btn>
              <span className="muted small">{last ? `${fmtBytes((last.down_bytes ?? 0) + (last.up_bytes ?? 0))} moved in ${last.took_sec}s${where(last)}` : 'about ten seconds'}</span></>}
        </div>
        {failed && <div className="error">{failed === 'cancelled' ? 'Stopped.' : failed}</div>}
        {!failed && last?.note && <div className="notice">{last.note}</div>}
        <div className="muted small method">{d.method}</div>
      </section>
      {hist.length > 1 && (
        <>
          <div className="dash-bar"><h2 className="section">Past tests</h2><span className="spacer" /><span className="muted small">{hist.length} kept</span></div>
          <section className="card">
            <Sparkline height={60} format={(v) => `${mbps(v)} Mbps`}
              series={[{ name: 'down', values: hist.map((r) => r.down_mbps), color: 'var(--s1)' },
                { name: 'up', values: hist.map((r) => r.up_mbps), color: 'var(--s2)' }]} />
            <div className="props" style={{ marginTop: 8 }}>
              {[...hist].reverse().slice(0, 8).map((r) => (
                <div key={r.at}><span className="k">{new Date(r.at).toLocaleString()}</span>
                  <span className="mono">↓ {mbps(r.down_mbps)} · ↑ {mbps(r.up_mbps)} Mbps · {r.latency_ms} ms</span></div>
              ))}
            </div>
          </section>
        </>
      )}
      <div className="dash-bar"><h2 className="section">Out there</h2></div>
      <section className="card">
        <div className="props">
          <div><span className="k">Public address</span><span className="mono">{o.ip || '—'}{o.error ? ` · ${o.error}` : ''}</span></div>
          {o.colo && <div><span className="k">Reached via</span><span>{o.colo}{o.country ? ` · ${o.country}` : ''}</span></div>}
          {d.report.gateway && <div><span className="k">Gateway</span><span className="mono">{d.report.gateway}{d.report.gateway_via ? ` via ${d.report.gateway_via}` : ''}</span></div>}
          {!!d.report.dns?.length && <div><span className="k">DNS</span><span className="mono">{d.report.dns.join(', ')}</span></div>}
          {!!d.report.search?.length && <div><span className="k">Search</span><span className="mono">{d.report.search.join(' ')}</span></div>}
        </div>
      </section>
      <div className="dash-bar"><h2 className="section">Interfaces</h2><span className="spacer" />
        {hidden > 0 && <button className="link" onClick={() => setAll(true)}>show {hidden} more</button>}
        {all && <button className="link" onClick={() => setAll(false)}>hide the scenery</button>}</div>
      {shown.map((f) => {
        const rx = rateOf(f.name, 'rx'), tx = rateOf(f.name, 'tx')
        return (
          <section className="card iface" key={f.name}>
            <div className="sub-head">
              <span className="dname mono">{f.name}</span>
              {f.default && <span className="badge ok" title="the default route leaves by this one">internet</span>}
              <span className="badge">{KIND[f.kind] ?? f.kind}</span>
              {isTunnel(f) && <button className="link" onClick={goVPN} title="who is on this tunnel">peers →</button>}
              {!f.up ? <span className="badge bad">down</span> : !f.carrier && <span className="badge bad">no link</span>}
              <span className="spacer" />
              <span className="muted small mono">{typeof rx === 'number' ? `↓ ${fmtValue(rx, 'B/s')}` : ''} {typeof tx === 'number' ? `↑ ${fmtValue(tx, 'B/s')}` : ''}</span>
            </div>
            <div className="props">
              {f.addrs.length > 0 && <div><span className="k">Address</span><span className="mono">{f.addrs.filter((a) => !a.v6).map((a) => `${a.ip}/${a.prefix}`).join(', ') || '—'}</span></div>}
              {f.addrs.some((a) => a.v6) && <div><span className="k">IPv6</span><span className="mono" title={f.addrs.filter((a) => a.v6).map((a) => `${a.ip}/${a.prefix}`).join('\n')}>{f.addrs.filter((a) => a.v6).length} addresses</span></div>}
              {!!f.speed_mbps && <div><span className="k">Link</span><span>{f.speed_mbps >= 1000 ? `${f.speed_mbps / 1000} Gbps` : `${f.speed_mbps} Mbps`}{f.duplex ? ` ${f.duplex} duplex` : ''}</span></div>}
              {f.signal_dbm !== undefined && <div><span className="k">Signal</span><span>{f.signal_dbm} dBm</span></div>}
              {f.mac && <div><span className="k">MAC</span><span className="mono">{f.mac}</span></div>}
              <div><span className="k">MTU</span><span>{f.mtu}{f.driver ? ` · ${f.driver}` : ''}</span></div>
              <div><span className="k">Moved</span><span className="mono">↓ {fmtBytes(f.rx_total)} · ↑ {fmtBytes(f.tx_total)}</span></div>
            </div>
          </section>
        )
      })}
    </div>
  )
 }
@@ -16,9 +16,16 @@ function fmtCPU(ns: number): string {
  const s = ns / 1e9
  return s >= 3600 ? `${(s / 3600).toFixed(1)} h` : s >= 60 ? `${(s / 60).toFixed(1)} min` : `${s.toFixed(1)} s`
 }
 // systemd writes "Fri 2026-10-02 23:48:25 PDT", which Chromium parses and Firefox does not; read
 // the fields directly, in the browser's own zone (the machine's, on a LAN), and fall back to Date.parse.
 function parseSystemdTime(ts: string): number {
  const m = /(\d{4})-(\d{2})-(\d{2}) (\d{2}):(\d{2}):(\d{2})/.exec(ts)
  if (m) return new Date(+m[1], +m[2] - 1, +m[3], +m[4], +m[5], +m[6]).getTime()
  return Date.parse(ts.replace(/^[A-Za-z]{3} /, ''))
 }
 function since(ts?: string): string {
  if (!ts) return ''
-  const t = Date.parse(ts.replace(/^[A-Za-z]{3} /, ''))
+  const t = parseSystemdTime(ts)
  if (isNaN(t)) return ts
  const s = Math.max(0, (Date.now() - t) / 1000)
  const d = Math.floor(s / 86400), h = Math.floor((s % 86400) / 3600), m = Math.floor((s % 3600) / 60)
@@ -217,7 +224,9 @@ export default function Services() {
            ))}
          </div>
        </div>
-        {err && <div className="error">{err}</div>}
+        {err && units.length === 0 && loaded
          ? <div className="notice" style={{ margin: 8 }}>{err}</div> // a host with no systemd, or one that is unreachable: not a filter result
          : err && <div className="error">{err}</div>}
        <div className="svc-rows">
          {!loaded && <Loading label="Listing services…" />}
          {shown.map((u) => (
@@ -230,7 +239,7 @@ export default function Services() {
              <span className="svc-state">{u.failed ? 'failed' : u.sub}{u.enabled === 'enabled' ? '' : u.enabled ? ` · ${u.enabled}` : ''}</span>
            </div>
          ))}
-          {loaded && shown.length === 0 && <div className="muted" style={{ padding: 12 }}>Nothing matches.</div>}
+          {loaded && shown.length === 0 && !(err && units.length === 0) && <div className="muted" style={{ padding: 12 }}>Nothing matches.</div>}
        </div>
      </div>
      {sel ? <Detail name={sel} onChanged={load} onBack={() => setSel(null)} /> : <div className="svc-empty muted">Select a service.</div>}
@@ -198,7 +198,7 @@ function Updates({ me }: { me: Me }) {
            {st.notes && <div className="mono small" style={{ whiteSpace: 'pre-wrap', marginTop: 6, maxHeight: 140, overflow: 'auto' }}>{st.notes.trim()}</div>}
          </div>
        )}
-        {st && !st.behind && st.latest && <div className="muted small" style={{ marginTop: 8 }}>This is the newest published build.</div>}
+        {st && !st.behind && st.latest && <div className="muted small" style={{ marginTop: 8 }}>{st.current.replace(/^v/, '') === st.latest.replace(/^v/, '') ? 'This is the newest published build.' : `This build is newer than the last published release (${st.latest}).`}</div>}
        {(err || st?.error) && <div className="error" style={{ marginTop: 8 }}>{err || st?.error}</div>}
        <div className="svc-actions" style={{ marginTop: 10 }}>
          <Btn className="btn" busy={busy === 'check'} disabled={!st?.enabled || !!busy} onClick={check}>Check for updates</Btn>
@@ -11,7 +11,12 @@ import Services from './Services'
 import Jobs from './Jobs'
 import Updates from './Updates'
 import Disks from './Disks'
 import Hardware from './Hardware'
 import Network from './Network'
 import VPN from './VPN'
 import ZFS from './ZFS'
 import Virtualization from './Virtualization'
 import KVM from './KVM'
 import Files from './Files'
 import Top from './Top'
 import Machines from './Machines'
@@ -23,7 +28,8 @@ import { MachineStatus, setMachine } from './api'
 import { LIST_FLOOR, useRefreshMs } from './prefs'
 import Power from './Power'
 import { applyAccent } from './accents'
-import { Btn, Toasts, askText, askYes, hasUnsaved, okToLeave, toast } from './ui'
+import { Btn, PageBoundary, Toasts, askText, askYes, hasUnsaved, okToLeave, toast, useEscape } from './ui'
 import { fmtLeft, useTimeLeft } from './guest'
 // Inline outline icon so it matches the monochrome glyphs in the rail (emoji would be full colour).
 const FolderIcon = (
@@ -50,26 +56,45 @@ const BellIcon = (
  </svg>
 )
 const GlobeIcon = (
  <svg viewBox="0 0 16 16" width="14" height="14" fill="none" stroke="currentColor" strokeWidth="1.4" strokeLinecap="round" aria-hidden="true">
    <circle cx="8" cy="8" r="6.2" /><path d="M1.9 8h12.2M8 1.8c1.7 1.8 2.6 3.9 2.6 6.2S9.7 12.4 8 14.2C6.3 12.4 5.4 10.3 5.4 8S6.3 3.6 8 1.8z" />
  </svg>
 )
 // A padlock-ish shield: the tab is about who may come in, and a keyhole reads at 14px where
 // a padlock's shackle does not.
 const ShieldIcon = (
  <svg viewBox="0 0 16 16" width="14" height="14" fill="none" stroke="currentColor" strokeWidth="1.4" strokeLinejoin="round" aria-hidden="true">
    <path d="M8 1.6 13.2 3.5v4.2c0 3-2.1 5.4-5.2 6.7-3.1-1.3-5.2-3.7-5.2-6.7V3.5z" /><circle cx="8" cy="7.2" r="1.3" /><path d="M8 8.5v2" />
  </svg>
 )
 const GROUPS: Group[] = [
  { id: 'procs', label: 'Processes', icon: '▚' },
  { id: 'terminal', label: 'Terminal', icon: '>_' },
  { id: 'services', label: 'Services', icon: '◈' },
  { id: 'docker', label: 'Docker', icon: DockerIcon },
  { id: 'jobs', label: 'Jobs', icon: ClockIcon },
  { id: 'updates', label: 'Updates', icon: '⇪' },
  { id: 'disks', label: 'Disks', icon: '▤' },
  { id: 'zfs', label: 'ZFS', icon: '⛁' },
  { id: 'usage', label: 'Usage', icon: '▧' },
  { id: 'files', label: 'Files', icon: FolderIcon },
  { id: 'network', label: 'Network', icon: GlobeIcon },
  { id: 'vpn', label: 'VPN', icon: ShieldIcon },
  { id: 'usage', label: 'Usage', icon: '▧' },
  { id: 'zfs', label: 'ZFS', icon: '⛁' },
  { id: 'docker', label: 'Docker', icon: DockerIcon },
  { id: 'virt', label: 'Virtualization', icon: '▣' },
  { id: 'kvm', label: 'KVM', icon: '⌨' },
  { id: 'hardware', label: 'Hardware', icon: '⬡' },
  { id: 'updates', label: 'Updates', icon: '⇪' },
  { id: 'alerts', label: 'Alerts', icon: BellIcon },
  { id: 'machines', label: 'Machines', icon: '⌂' },
  { id: 'users', label: 'Users', icon: '⚇' },
  { id: 'widgets', label: 'Widgets', icon: '▦' },
  { id: 'users', label: 'Users', icon: '⚇' },
  { id: 'settings', label: 'Settings', icon: '⚙' },
 ]
-type View = 'terminal' | 'widgets' | 'docker' | 'alerts' | 'services' | 'jobs' | 'updates' | 'disks' | 'zfs' | 'usage' | 'files' | 'procs' | 'machines' | 'users' | 'settings'
+type View = 'terminal' | 'widgets' | 'docker' | 'virt' | 'alerts' | 'services' | 'jobs' | 'updates' | 'disks' | 'hardware' | 'network' | 'vpn' | 'zfs' | 'usage' | 'files' | 'procs' | 'machines' | 'kvm' | 'users' | 'settings'
-const VIEWS: View[] = ['terminal', 'widgets', 'docker', 'alerts', 'services', 'jobs', 'updates', 'disks', 'zfs', 'usage', 'files', 'procs', 'machines', 'users', 'settings']
+const VIEWS: View[] = ['terminal', 'widgets', 'docker', 'virt', 'alerts', 'services', 'jobs', 'updates', 'disks', 'hardware', 'network', 'vpn', 'zfs', 'usage', 'files', 'procs', 'machines', 'kvm', 'users', 'settings']
 const HUB_ONLY: View[] = ['machines', 'users', 'settings', 'alerts']
 // URL hash is the source of truth for where you are: #/<machine|hub>/<view>. Back and
@@ -100,6 +125,9 @@ export default function Shell({ me: hubMe, onLoggedOut }: { me: Me; onLoggedOut:
  const [machines, setMachines] = useState<MachineStatus[]>([])
  const [remoteMe, setRemoteMe] = useState<Me | null>(null)
  const [remoteErr, setRemoteErr] = useState('')
  // A guest login is an account with a clock on it: nothing is hidden from it, but the header
  // says how long it has, for whoever is holding the phone.
  const leftMs = useTimeLeft(hubMe.guest_until)
  // A machine is titled by its own hostname; the name it was added under stays its identity
  // (the URL, its accent, what the hub's API calls it). See titleOf.
  const title = titleOf(machine, machines, hubMe.host)
@@ -109,6 +137,7 @@ export default function Shell({ me: hubMe, onLoggedOut }: { me: Me; onLoggedOut:
  const [active, setActive] = useState<string | null>(null)
  const [view, setView] = useState<View>(() => parseHash()?.view ?? 'procs')
  const [drawer, setDrawer] = useState(false)
  useEscape(() => setDrawer(false), drawer) // the phone's drawer closes like every other overlay
  const [online, setOnline] = useState(true)
  const [creating, setCreating] = useState(false)
  const handles = useRef<Record<string, TermHandle>>({})
@@ -123,6 +152,32 @@ export default function Shell({ me: hubMe, onLoggedOut }: { me: Me; onLoggedOut:
  useEffect(() => { reload() }, [reload])
  // One page sending you to another — the Network tab handing a tunnel over to the VPN tab.
  // A custom event rather than a hash assignment, because assigning location.hash does not
  // reliably raise popstate, which is what this shell navigates by.
  useEffect(() => {
    const h = (e: Event) => {
      const id = (e as CustomEvent<string>).detail
      if (VIEWS.includes(id as View)) { setView(id as View); setDrawer(false) }
    }
    window.addEventListener('hamlab:view', h)
    return () => window.removeEventListener('hamlab:view', h)
  }, [])
  // A shell opened inside a container from the Docker tab is a terminal like any other:
  // pick it up, show the Terminal tab, and make it the active one.
  useEffect(() => {
    const h = async (e: Event) => {
      const id = (e as CustomEvent<string>).detail
      if (!id) return
      await reload()
      setActive(id)
      setView('terminal')
    }
    window.addEventListener('hamlab:terminal', h)
    return () => window.removeEventListener('hamlab:terminal', h)
  }, [reload])
  // Machine list for the switcher (hub-local), refreshed while the console is open.
  const loadMachines = useCallback(async () => { try { setMachines((await api.machines()).nodes) } catch { /* not a hub or offline */ } }, [])
  useEffect(() => { loadMachines(); const t = setInterval(loadMachines, listRate); return () => clearInterval(t) }, [loadMachines, listRate])
@@ -256,6 +311,7 @@ export default function Shell({ me: hubMe, onLoggedOut }: { me: Me; onLoggedOut:
        <span className="stat" title="load average, 1 · 5 · 15 min">load {me.load}</span>
        <span className={'dot ' + (online && !remoteErr ? 'ok' : 'bad')} title={remoteErr ? `${title} unreachable` : online ? 'connected' : 'disconnected'} />
        <Power host={me.host} label={title} rebootNeeded={me.reboot_needed} />
        {hubMe.guest && <span className="badge guest" title={`this guest login expires ${new Date(hubMe.guest_until ?? '').toLocaleTimeString()}`}>guest · {fmtLeft(leftMs)}</span>}
        {hubMe.account && <span className="host account" title="signed in as">{hubMe.account}</span>}
        <Btn className="link" onClick={logout}>logout</Btn>
      </header>
@@ -267,20 +323,27 @@ export default function Shell({ me: hubMe, onLoggedOut }: { me: Me; onLoggedOut:
        <div className="backdrop" onClick={() => setDrawer(false)} />
        <main className="main" key={machine ?? 'hub'}>
          <PageBoundary page={GROUPS.find((g) => g.id === view)?.label ?? view}>
          {view === 'settings' && <Settings me={hubMe} />}
          {view === 'users' && <Users me={hubMe} />}
          {view === 'kvm' && <KVM key={machine ?? 'hub'} machine={machine} title={title} />}
          {view === 'usage' && <Usage key={machine ?? ''} />}
          {view === 'machines' && <Machines hub={hubMe} onChanged={loadMachines} onOpen={(n) => { switchMachine(n); setView('procs') }} />}
          {view === 'widgets' && <Gallery />}
          {view === 'docker' && <Docker />}
          {view === 'virt' && <Virtualization />}
          {view === 'alerts' && <Alerts machines={machines.map((m) => m.name)} titleOf={(n) => titleOf(n === hubMe.host ? null : n, machines, hubMe.host)} />}
          {view === 'procs' && <Top />}
          {view === 'services' && <Services />}
          {view === 'jobs' && <Jobs />}
          {view === 'updates' && <Updates />}
          {view === 'disks' && <Disks />}
          {view === 'hardware' && <Hardware />}
          {view === 'network' && <Network />}
          {view === 'vpn' && <VPN />}
          {view === 'zfs' && <ZFS />}
          {view === 'files' && <Files onOpenInTerminal={runInTerminal} />}
          </PageBoundary>
          {termView && <TermTabs terms={ordered} active={active} creating={creating}
            onSelect={setActive} onNew={create} onRename={rename} onClose={close} onMove={moveTab}
            onFocusTerm={() => window.setTimeout(() => { const id = activeRef.current; if (id) handles.current[id]?.armCtrl(false) }, 0)} />}
@@ -306,7 +369,7 @@ export default function Shell({ me: hubMe, onLoggedOut }: { me: Me; onLoggedOut:
        <span>{machine ? `${title} · ` : ''}{me.version.startsWith('v') ? me.version : 'v' + me.version}</span>
        <span>up {fmtUptime(me.uptime)}</span>
        <span className="spacer" />
-        <span className="muted">{view === 'widgets' ? 'Widgets' : view === 'docker' ? 'Docker' : view === 'alerts' ? 'Alerts' : view === 'services' ? 'Services' : view === 'jobs' ? 'Jobs' : view === 'updates' ? 'Updates' : view === 'disks' ? 'Disks' : view === 'zfs' ? 'ZFS' : view === 'usage' ? 'Usage' : view === 'files' ? 'Files' : view === 'procs' ? 'Processes' : view === 'machines' ? 'Machines' : view === 'users' ? 'Users' : view === 'settings' ? 'Settings' : activeTerm ? activeTerm.name : ''}</span>
+        <span className="muted">{view === 'widgets' ? 'Widgets' : view === 'docker' ? 'Docker' : view === 'virt' ? 'Virtualization' : view === 'alerts' ? 'Alerts' : view === 'services' ? 'Services' : view === 'jobs' ? 'Jobs' : view === 'updates' ? 'Updates' : view === 'disks' ? 'Disks' : view === 'hardware' ? 'Hardware' : view === 'network' ? 'Network' : view === 'vpn' ? 'VPN' : view === 'zfs' ? 'ZFS' : view === 'usage' ? 'Usage' : view === 'files' ? 'Files' : view === 'procs' ? 'Processes' : view === 'machines' ? 'Machines' : view === 'kvm' ? 'KVM' : view === 'users' ? 'Users' : view === 'settings' ? 'Settings' : activeTerm ? activeTerm.name : ''}</span>
      </footer>
    </div>
  )
@@ -6,6 +6,7 @@
 import { useEffect, useMemo, useState } from 'react'
 import { api, StackHome } from './api'
 import { askYes, Btn, Loading, useEscape } from './ui'
 import CodeEditor from './CodeEditor'
 interface Form { image: string; ports: string; volumes: string; env: string; labels: string; restart: string; command: string; network: string }
 const EMPTY: Form = { image: '', ports: '', volumes: '', env: '', labels: '', restart: 'unless-stopped', command: '', network: '' }
@@ -119,7 +120,7 @@ export default function StackNew({ mode, from, networks = [], taken, onClose, on
            <label className="muted">Command (optional)<input className="mono" value={form.command} onChange={set('command')} spellCheck={false} autoCapitalize="off" /></label>
            <pre className="unit-file stack-preview">{built}</pre>
          </> : <>
-            <textarea className="editor stack-editor" value={text} onChange={(e) => setText(e.target.value)} spellCheck={false} autoCapitalize="off" autoCorrect="off" />
+            <CodeEditor className="stack-editor" language="yaml" value={text} onChange={setText} />
            {from?.hasEnv && <label className="check"><input type="checkbox" checked={copyEnv} onChange={(e) => setCopyEnv(e.target.checked)} /> <span>Copy {from.name}'s .env too <span className="muted">— copied on the machine itself; its values are never shown here</span></span></label>}
            {!(from && copyEnv) && <label className="muted">.env (optional) — variables the compose file reads; kept readable by root only
              <textarea className="mono" rows={2} value={env} onChange={(e) => setEnv(e.target.value)} placeholder={'SECRET=…'} spellCheck={false} />
@@ -4,7 +4,7 @@
 // /api/procs for the table — so this works on any machine the console can see, whether or not
 // btop is installed on it.
 import { createContext, memo, useCallback, useContext, useEffect, useMemo, useRef, useState } from 'react'
-import { api, ProcTable, Proc, Sample } from './api'
+import { api, GPU, ProcTable, Proc, Sample } from './api'
 import { Catalog, MONITOR_SHAPE, Values, mergeValues, sampleOf, stale } from './sampleAdapter'
 import { fetchCatalog, fetchHistories, useSensorStream } from './sensors'
 import { fmtMs, nextPoint, REFRESH_OPTIONS, setRefreshMs, strideMs, useHistorySec, useRefreshMs } from './prefs'
@@ -186,7 +186,7 @@ function braille(values: number[], cells = 7, max = 100): string {
 }
 // ---- the curves this page keeps ----
-const CURVES = 'cpu.usage,cpu.core.*.usage,net.*.rx,net.*.tx,disk.*.util'
+const CURVES = 'cpu.usage,cpu.core.*.usage,net.*.rx,net.*.tx,disk.*.util,gpu.*.usage'
 function useCurves(live: Values | null, t: number, refreshMs: number, historySec: number, resumed: number) {
  const cur = useRef<Map<string, number[]>>(new Map())
@@ -216,7 +216,7 @@ function useCurves(live: Values | null, t: number, refreshMs: number, historySec
    const n = binN.current
    for (const [id, v] of Object.entries(live)) {
      if (typeof v !== 'number') continue
-      if (!/^cpu\.usage$|^cpu\.core\.\d+\.usage$|^net\..+\.(rx|tx)$|^disk\..+\.util$/.test(id)) continue
+      if (!/^cpu\.usage$|^cpu\.core\.\d+\.usage$|^net\..+\.(rx|tx)$|^disk\..+\.util$|^gpu\..+\.usage$/.test(id)) continue
      const arr = cur.current.get(id) ?? []
      if (!cur.current.has(id)) cur.current.set(id, arr)
      if (fresh || arr.length === 0) { arr.push(v); if (arr.length > MAX) arr.shift() }
@@ -280,7 +280,7 @@ export default function Top() {
  // how many points the history window holds at this rate: the width of every graph
  const cap = Math.max(2, Math.min(MAX, Math.round((historySec * 1000) / strideMs(historySec, refreshMs))))
-  const [sort, setSort] = useState<'cpu' | 'mem' | 'pid' | 'name'>('cpu')
+  const [sort, setSort] = useState<'cpu' | 'gpu' | 'mem' | 'pid' | 'name'>('cpu')
  const [rev, setRev] = useState(false)
  const [filter, setFilter] = useState('')
  const [sel, setSel] = useState<number | null>(null)
@@ -303,8 +303,8 @@ export default function Top() {
    if (!pt) return []
    const f = filter.trim().toLowerCase()
    const list = pt.procs.filter((p) => !f || p.name.toLowerCase().includes(f) || (p.cmd ?? '').toLowerCase().includes(f) || p.user.toLowerCase().includes(f) || String(p.pid) === f)
-    const key = (p: Proc) => (sort === 'cpu' ? p.cpu : sort === 'mem' ? p.rss : sort === 'pid' ? p.pid : 0)
+    const key = (p: Proc) => (sort === 'cpu' ? p.cpu : sort === 'gpu' ? (p.gpu ?? 0) : sort === 'mem' ? p.rss : sort === 'pid' ? p.pid : 0)
-    list.sort((a, b) => (sort === 'name' ? a.name.localeCompare(b.name) : key(b) - key(a)))
+    list.sort((a, b) => (sort === 'name' ? a.name.localeCompare(b.name) : key(b) - key(a) || b.cpu - a.cpu))
    return rev ? list.reverse() : list
  }, [pt, filter, sort, rev])
@@ -342,10 +342,19 @@ export default function Top() {
  const pct = (a: number, b: number) => (b > 0 ? (a / b) * 100 : 0)
  // the top of a graph's scale: the peak of what it is drawing — the window, not everything kept
  const peak = (id: string) => Math.max(1, ...(curves.get(id) ?? []).slice(-cap))
  // the GPUs, with each one's sensor stem: the same rule the server uses, so two identical
  // cards are gpu.<name> and gpu.<name>_2 here as there
  const gpus = (pt?.gpus ?? []).map((g, i, all) => {
    const seg = segOf(g.name), n = all.slice(0, i).filter((o) => segOf(o.name) === seg).length
    return { g, stem: 'gpu.' + seg + (n ? `_${n + 1}` : '') }
  })
  // a GPU with nothing to read can't light a process's Gpu% either: the column, and its sort,
  // are for a machine where at least one GPU says something
  const gpuCol = gpus.some(({ g }) => !unmeasured(g))
  return (
    <Pal.Provider value={pal}>
-    <div className="top" style={pal.vars as React.CSSProperties}>
+    <div className={'top' + (gpus.length ? ' has-gpu' : '') + (gpus.length && !gpuCol ? ' gpu-slim' : '')} style={pal.vars as React.CSSProperties}>
      <Box n={1} name="cpu" className="top-cpu" right={<>
        <span className="t-clock">{clock.toLocaleTimeString([], { hour12: false })}</span>
        <span className="t-rate">
@@ -383,6 +392,16 @@ export default function Top() {
        </div>
      </Box>
      {gpus.length > 0 && (
        <div className="top-gpus">
          {gpus.map(({ g, stem }, i) => (
            <Box key={stem} name={`gpu${i}`} className={'top-gpu' + (unmeasured(g) ? ' slim' : '')} right={<span className="t-gpu-name" title={[g.name, g.driver, g.bus].filter(Boolean).join(' · ')}>{g.name}</span>}>
              {unmeasured(g) ? <GPULine g={g} /> : <GPUBox g={g} values={curves.get(stem + '.usage') ?? []} tick={tick} advance={advance} cap={cap} />}
            </Box>
          ))}
        </div>
      )}
      <Box n={2} name="mem" className="top-mem">
        <div className="t-row"><b>Total:</b><span className="spacer" />{bytes(mem.total)}</div>
        {([['Used', mem.used, 'used'], ['Available', mem.avail, 'avail'], ['Cached', cached, 'cached'], ['Free', free, 'free']] as const).map(([k, v, g]) => (
@@ -445,17 +464,17 @@ export default function Top() {
      <Box n={5} name="proc" className="top-proc" right={<>
        <span className="t-sorts">
-          {(['pid', 'name', 'mem', 'cpu'] as const).map((k) => <button key={k} className={sort === k ? 'on' : ''} onClick={() => setSort(k)}>{k}</button>)}
+          {(['pid', 'name', 'mem', ...(gpuCol ? ['gpu' as const] : []), 'cpu'] as const).map((k) => <button key={k} className={sort === k ? 'on' : ''} onClick={() => setSort(k)}>{k}</button>)}
          <button className={rev ? 'on' : ''} onClick={() => setRev(!rev)}>reverse</button>
        </span>
      </>}>
-        <div className="t-proc">
+        <div className={'t-proc' + (gpuCol ? ' has-gpu' : '')}>
          <input className="t-filter" value={filter} onChange={(e) => setFilter(e.target.value)} placeholder="filter" spellCheck={false} autoCapitalize="off" autoCorrect="off" />
          <div className="t-thead">
            <span className="c-pid">Pid:</span><span className="c-name">Program:</span><span className="c-cmd">Command:</span>
-            <span className="c-thr">Threads:</span><span className="c-user">User:</span><span className="c-mem">MemB</span><span className="c-g" /><span className="c-cpu">Cpu%</span>
+            <span className="c-thr">Threads:</span><span className="c-user">User:</span><span className="c-mem">MemB</span><span className="c-g" />{gpuCol && <span className="c-gpu">Gpu%</span>}<span className="c-cpu">Cpu%</span>
          </div>
-          <Rows procs={procs} sel={sel} onSel={setSel} curves={pcurve} empty={err || 'no process matches'} />
+          <Rows procs={procs} sel={sel} onSel={setSel} curves={pcurve} gpu={gpuCol} empty={err || 'no process matches'} />
          <div className="t-foot">
            {sel !== null && procs.find((p) => p.pid === sel) ? (() => { const p = procs.find((x) => x.pid === sel)!; return <>
              <span className="t-dim">{p.pid} {p.name}</span><span className="spacer" />
@@ -473,7 +492,7 @@ export default function Top() {
 // The rows are the heaviest thing on the page — hundreds of them — so they are redrawn when the
 // table is read again, not on every tick of the curves above them.
-const Rows = memo(function Rows({ procs, sel, onSel, curves, empty }: { procs: Proc[]; sel: number | null; onSel: (p: number | null) => void; curves: React.RefObject<Map<number, number[]>>; empty: string }) {
+const Rows = memo(function Rows({ procs, sel, onSel, curves, gpu, empty }: { procs: Proc[]; sel: number | null; onSel: (p: number | null) => void; curves: React.RefObject<Map<number, number[]>>; gpu: boolean; empty: string }) {
  return (
    <div className="t-rows">
      {procs.map((p) => {
@@ -487,6 +506,7 @@ const Rows = memo(function Rows({ procs, sel, onSel, curves, empty }: { procs: P
            <span className="c-user">{p.user}</span>
            <span className="c-mem">{shortBytes(p.rss)}</span>
            <span className="c-g">{braille(c, 7, Math.max(5, ...(c.length ? c : [0])))}</span>
            {gpu && <span className={'c-gpu' + (p.gpu ? ' lit' : '')}>{p.gpu ? p.gpu.toFixed(1) : ''}</span>}
            <span className="c-cpu">{p.cpu.toFixed(1)}</span>
          </div>
        )
@@ -496,6 +516,78 @@ const Rows = memo(function Rows({ procs, sel, onSel, curves, empty }: { procs: P
  )
 })
 // One GPU, laid out like the cpu box: the busy curve on the left; on the right the device's
 // meters — busy, each engine, temperature — then clock, power and memory, then the processes
 // that have it open with their share. What cannot be measured on this card is left out rather
 // than shown as zero: an Arc on i915 has no temperature sensor, Asahi has no busy counter.
 const ENGINES: [string, string][] = [['render', 'render'], ['compute', 'compute'], ['video', 'video'], ['video-enhance', 'enhance'], ['copy', 'copy']]
 function GPUBox({ g, values, tick, advance, cap }: { g: GPU; values: number[]; tick: number; advance: number; cap: number }) {
  const engines = ENGINES.filter(([k]) => g.engines && g.engines[k] !== undefined)
  const temp = g.temp !== undefined && g.temp > 0 ? g.temp : undefined
  const how = g.stalled ? 'a sysfs read of this card stopped answering; its readings are paused until the console restarts'
    : g.usage_src === 'idle' ? 'busy, from idle residency' : g.usage_src === 'busy' ? 'busy, as the driver reports it' : g.usage_src === 'clients' ? 'the busiest engine, summed over its clients' : 'this driver exposes no activity counter'
  return (
    <div className="t-gpu">
      <div className="t-gpu-graph">
        <Graph values={values} tick={tick} advance={advance} cap={cap} max={100} grad="cpu" />
        <div className="t-up">{g.usage_src ? `${Math.round(g.usage)}% ` : ''}{how}</div>
      </div>
      <div className="t-gpu-stats">
        <div className="t-cores-head"><b>{g.name}</b><span className="spacer" />{g.driver}{g.integrated ? ' · integrated' : ''}</div>
        <div className="t-core">
          <span className="t-core-n">GPU</span>
          <Meter pct={g.usage_src ? g.usage : 0} grad="cpu" />
          <span className="t-core-p">{g.usage_src ? `${Math.round(g.usage)}%` : '—'}</span>
          <span className="t-core-t">{temp !== undefined && <Meter pct={Math.min(100, temp)} grad="temp" />}</span>
          <span className="t-core-c" title={temp !== undefined ? (g.temp_src === 'package' ? 'the CPU package: an integrated GPU shares the die, and this is its only sensor' : "the card's own sensor") : 'no temperature sensor on this device'}>{temp !== undefined ? `${Math.round(temp)}°C${g.temp_src === 'package' ? '*' : ''}` : ''}</span>
        </div>
        {engines.map(([k, label]) => (
          <div className="t-core" key={k}>
            <span className="t-core-n t-eng">{label}</span>
            <Meter pct={g.engines![k]} grad="cpu" />
            <span className="t-core-p">{Math.round(g.engines![k])}%</span>
            <span className="t-core-t" /><span className="t-core-c" />
          </div>
        ))}
        <div className="t-gpu-facts">
          {(g.mhz !== undefined || g.max_mhz) ? <div className="t-row t-sub"><span>Clock</span><span className="spacer" />
            <span className="t-fact" title={g.req_mhz ? `${Math.round(g.req_mhz)} MHz requested${g.max_mhz ? `, ${Math.round(g.max_mhz)} MHz max` : ''}` : undefined}>{Math.round(g.mhz ?? 0)}{g.max_mhz ? ` / ${Math.round(g.max_mhz)}` : ''} MHz</span></div> : null}
          {(g.power_w !== undefined && g.power_w > 0) || g.power_cap_w ? <div className="t-row t-sub"><span>Power</span><span className="spacer" />
            <span className="t-fact">{(g.power_w ?? 0).toFixed(1)}{g.power_cap_w ? ` / ${Math.round(g.power_cap_w)}` : ''} W</span></div> : null}
          {g.fan_rpm ? <div className="t-row t-sub"><span>Fan</span><span className="spacer" /><span className="t-fact">{g.fan_rpm} rpm</span></div> : null}
          {g.mem_total ? <div className="t-row t-sub"><span>Mem</span><span className="t-pct">{Math.round(((g.mem_used ?? 0) / g.mem_total) * 100)}%</span><Meter pct={((g.mem_used ?? 0) / g.mem_total) * 100} grad="used" /><span className="t-amt">{bytes(g.mem_used ?? 0, 1)}</span></div>
            : g.mem_used ? <div className="t-row t-sub"><span>Mem</span><span className="spacer" /><span className="t-fact" title="resident memory of the processes that have it open">{bytes(g.mem_used, 1)}</span></div> : null}
        </div>
        <div className="t-gpu-clients">
          <div className="t-thead"><span className="c-pid">Pid:</span><span className="c-name">Program:</span><span className="c-mem">MemB</span><span className="c-cpu">Gpu%</span></div>
          {g.clients.map((c) => (
            <div className="t-crow" key={c.pid} title={c.engines ? Object.entries(c.engines).map(([k, v]) => `${k} ${v.toFixed(1)}%`).join('  ') : undefined}>
              <span className="c-pid">{c.pid}</span><span className="c-name">{c.name}</span><span className="c-mem">{c.mem ? shortBytes(c.mem) : ''}</span><span className="c-cpu">{c.usage.toFixed(1)}</span>
            </div>
          ))}
          {g.clients.length === 0 && <div className="t-dim t-crow-none">{g.clients_ok ? 'no process has it open' : 'other users’ processes are hidden: the console has no root here'}</div>}
        </div>
      </div>
    </div>
  )
 }
 // Nothing on this GPU can be read — no busy counter, engine, temperature, clock, power, memory,
 // nor a process with it open: the Mac's Asahi, today. Its box would be a graph that can never
 // draw beside a column of dashes, so it is one line instead. A client opening it brings the box back.
 const unmeasured = (g: GPU) => !g.usage_src && !g.engines && !(g.temp && g.temp > 0) && !g.mhz && !g.max_mhz && !g.power_w && !g.power_cap_w && !g.fan_rpm && !g.mem_total && !g.mem_used && g.clients.length === 0
 function GPULine({ g }: { g: GPU }) {
  return (
    <div className="t-row t-gpu-line">
      <span>{g.driver}{g.integrated ? ' · integrated' : ''}</span>
      {g.stalled
        ? <span className="t-dim" title="a sysfs read of this card stopped answering">readings paused until the console restarts</span>
        : <span className="t-dim" title="the driver exposes no busy counter, clock, temperature, power or memory reading">nothing measurable</span>}
      <span className="t-dim t-gpu-who" title={g.clients_ok ? undefined : 'the console has no root here'}>{g.clients_ok ? 'no process has it open' : 'other users’ processes are hidden'}</span>
    </div>
  )
 }
 // One physical drive. `used` is the bytes of the filesystems on it, left out when none can
 // be attributed to it — a pool member's share of a pool is not a number worth inventing.
 function Disk({ name, note, size, used, io, temp, mounts, pool }: { name: string; note?: string; size: number; used?: number; io?: number; temp?: number; mounts?: string[]; pool?: string }) {
@@ -52,7 +52,11 @@ export default function Updates() {
    setOpen(p.name)
    if (!changelog[p.name]) {
      try { const r = await api.changelog(p.name); setChangelog((c) => ({ ...c, [p.name]: r.text || '(empty)' })) }
-      catch (e) { setChangelog((c) => ({ ...c, [p.name]: e instanceof Error ? e.message : 'unavailable' })) }
+      catch (e) {
        const m = e instanceof Error ? e.message : 'unavailable'
        // apt's own words for a third-party repository that publishes none are a wall of "E: Failed to fetch"
        setChangelog((c) => ({ ...c, [p.name]: /changelog unavailable|failed to fetch/i.test(m) ? 'No changelog is published for this package — its repository does not provide one.' : m }))
      }
    }
  }
  const toggle = (n: string) => setSel((s) => { const x = new Set(s); if (x.has(n)) x.delete(n); else x.add(n); return x })
@@ -99,7 +103,7 @@ export default function Updates() {
                <span className="n">{p.name}{p.security && <span className="badge bad">security</span>}</span>
                <span className="d">{p.description}</span>
              </span>
-              <span className="pkg-ver"><span className="muted">{p.current}</span> → {p.candidate}<span className="muted"> · {p.suite}</span></span>
+              <span className="pkg-ver"><span className="muted">{p.current}</span> → {p.candidate}{p.suite && <span className="muted"> · {p.suite}</span>}</span>
            </div>
            {open === p.name && <pre className="run-out changelog">{changelog[p.name] ?? <span className="muted"><Spinner /> fetching changelog…</span>}</pre>}
          </div>
@@ -3,19 +3,98 @@
 // a token rather than a login, so a node's accounts are managed on that node's console.
 import { FormEvent, useCallback, useEffect, useState } from 'react'
 import { createRoot } from 'react-dom/client'
-import { api, Me, Session, User } from './api'
+import { api, FreshGuest, Guest, Me, Session, User } from './api'
 import { Btn, Loading, askYes, toast, useEscape } from './ui'
 import { fmtLeft } from './guest'
 export default function Users({ me }: { me: Me }) {
  return (
    <div className="dash settings">
      <Account me={me} />
      <Sessions />
      <Guests />
      <Accounts me={me} />
    </div>
  )
 }
 const GUEST_MINUTES = [15, 30, 60, 120, 480] as const
 // A login to hand somebody you are showing the console to: an ordinary account with a clock
 // on it, which expires on its own and can be ended here at any moment. Making one is the only
 // time its password is readable, so it is shown until dismissed rather than in a toast.
 function Guests() {
  const [list, setList] = useState<Guest[] | null>(null)
  const [minutes, setMinutes] = useState<number>(30)
  const [fresh, setFresh] = useState<FreshGuest | null>(null)
  const [busy, setBusy] = useState(false)
  const load = useCallback(async () => { try { setList(await api.guests()) } catch { setList([]) } }, [])
  useEffect(() => { load() }, [load])
  // The clock is the point of the thing, so it keeps up without a reload.
  useEffect(() => { const t = setInterval(load, 30000); return () => clearInterval(t) }, [load])
  const create = async () => {
    setBusy(true)
    try { setFresh(await api.createGuest(minutes)); load() }
    catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
    finally { setBusy(false) }
  }
  const revoke = async (g: Guest) => {
    if (!await askYes(`End the guest login ${g.name}? Whoever has it is signed out at once, everywhere.`, 'Revoke')) return
    try { await api.revokeGuest(g.name); toast(`${g.name} revoked`); if (fresh?.name === g.name) setFresh(null); load() }
    catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  return (
    <>
      <div className="dash-bar"><h2 className="section">Guest logins</h2><span className="spacer" /><span className="muted small">temporary, and can be revoked</span></div>
      <section className="card">
        {list === null && <Loading label="Loading…" />}
        {list?.length === 0 && <div className="empty">None. One lets somebody in for a while without giving them an account.</div>}
        {list?.map((g) => {
          const left = new Date(g.expires).getTime() - Date.now()
          return (
            <div className="row" key={g.name}>
              <span className="k mono">{g.name}</span>
              <span className="muted small">{left > 0 ? `expires in ${fmtLeft(left)}` : 'expiring'} · {g.sessions > 0 ? `${g.sessions} signed in` : g.used ? 'signed in earlier' : 'not used yet'}{g.issued_by ? ` · by ${g.issued_by}` : ''}</span>
              <Btn className="btn danger" onClick={() => revoke(g)}>Revoke</Btn>
            </div>
          )
        })}
      </section>
      {fresh && <FreshCard g={fresh} onClose={() => setFresh(null)} />}
      <div className="card addform" style={{ marginTop: 8 }}>
        <label>Lasts
          <select value={minutes} onChange={(e) => setMinutes(Number(e.target.value))}>
            {GUEST_MINUTES.map((m) => <option key={m} value={m}>{m < 60 ? `${m} minutes` : `${m / 60} hour${m === 60 ? '' : 's'}`}</option>)}
          </select>
        </label>
        <div className="svc-actions"><Btn className="btn on" busy={busy} onClick={create}>Create guest login</Btn><span className="muted small">A guest can do everything you can, on every machine — including the terminal — until the time is up or you revoke it. Give one out accordingly.</span></div>
      </div>
    </>
  )
 }
 // The credentials, once. Copy takes the whole invitation, because that is what gets pasted
 // into a message.
 function FreshCard({ g, onClose }: { g: FreshGuest; onClose: () => void }) {
  const invite = `${location.origin}\nuser: ${g.name}\npassword: ${g.password}\nexpires ${new Date(g.expires).toLocaleTimeString()}`
  const copy = async (text: string, what: string) => {
    try { await navigator.clipboard.writeText(text); toast(`${what} copied`) }
    catch { toast('select and copy it by hand', 'err') }
  }
  return (
    <section className="card fresh" style={{ marginTop: 8 }}>
      <div className="row"><span className="k">User</span><span className="mono">{g.name}</span><Btn className="btn" onClick={() => copy(g.name, 'user')}>Copy</Btn></div>
      <div className="row"><span className="k">Password</span><span className="mono">{g.password}</span><Btn className="btn" onClick={() => copy(g.password, 'password')}>Copy</Btn></div>
      <div className="svc-actions" style={{ marginTop: 8 }}>
        <Btn className="btn on" onClick={() => copy(invite, 'invitation')}>Copy invitation</Btn>
        <span className="muted small">Shown once. Expires {new Date(g.expires).toLocaleTimeString()}.</span>
        <span className="spacer" />
        <button className="btn" onClick={onClose}>Done</button>
      </div>
    </section>
  )
 }
 function Account({ me }: { me: Me }) {
  const [cur, setCur] = useState('')
  const [pw1, setPw1] = useState('')
@@ -54,6 +133,13 @@ function Sessions() {
    catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  const others = list?.filter((s) => !s.current).length ?? 0
  // a day unused is the line between a device that is used and a script that signed in once
  const IDLE_H = 24
  const idle = list?.filter((s) => !s.current && Date.now() - new Date(s.last_seen).getTime() > IDLE_H * 3600e3).length ?? 0
  const revokeIdle = async () => {
    if (!await askYes(`Sign out ${idle} ${idle === 1 ? 'session' : 'sessions'} not used in the last ${IDLE_H} hours? Devices used since then, and this one, stay signed in.`, 'Revoke')) return
    try { const r = await api.revokeIdleSessions(IDLE_H); toast(`${r.revoked} signed out`); load() } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  const revokeOthers = async () => {
    if (!await askYes(`Sign out ${others} other ${others === 1 ? 'device' : 'devices'}? This one stays signed in.`)) return
    try { const r = await api.revokeOtherSessions(); toast(`${r.revoked} signed out`); load() } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
@@ -66,11 +152,15 @@ function Sessions() {
        {list && list.map((s) => (
          <div className="row" key={s.id}>
            <span className="k">{device(s.agent)}{s.current && <span className="badge" style={{ marginLeft: 6 }}>this device</span>}</span>
-            <span className="muted small">{s.ip ?? '?'} · seen {ago(s.last_seen)} · since {new Date(s.created).toLocaleDateString()}</span>
+            <span className="muted small">{s.ip ?? '?'} · seen {ago(s.last_seen)} · since {new Date(s.created).toLocaleDateString()}{!s.current && Date.now() - new Date(s.last_seen).getTime() > IDLE_H * 3600e3 ? <span className="badge" style={{ marginLeft: 6 }}>idle</span> : null}</span>
            <Btn className={'btn' + (s.current ? '' : ' danger')} onClick={() => revoke(s)}>{s.current ? 'Sign out' : 'Revoke'}</Btn>
          </div>
        ))}
-        {others > 0 && <div className="svc-actions" style={{ marginTop: 8 }}><Btn className="btn danger" onClick={revokeOthers}>Sign out everywhere else</Btn></div>}
+        {others > 0 && <div className="svc-actions" style={{ marginTop: 8 }}>
          {idle > 0 && <Btn className="btn" onClick={revokeIdle}>Sign out {idle} idle</Btn>}
          <Btn className="btn danger" onClick={revokeOthers}>Sign out everywhere else</Btn>
          <span className="muted small">A session unused for a week signs itself out.</span>
        </div>}
      </section>
    </>
  )
@@ -0,0 +1,218 @@
 // VPN: this machine's WireGuard tunnels, and the devices on them.
 //
 // Per machine, like every other tab, which here is the whole point: the tunnel people dial
 // into and a tunnel this machine dials out through are different machines, and showing them
 // together would invite the question "which box is this actually on".
 import { FormEvent, useCallback, useEffect, useState } from 'react'
 import qrcode from 'qrcode-generator'
 import { api, CreatedVPNPeer, VPNReport, VPNPeer, VPNTunnel } from './api'
 import { fmtBytes } from './fmt'
 import { Btn, Loading, askText, toast, useEscape } from './ui'
 import { useRefreshMs } from './prefs'
 // WireGuard says nothing when it has nothing to say, so "online" is a handshake in the last
 // few minutes rather than a connection — there is no connection to have.
 const ago = (iso?: string) => {
  if (!iso) return 'never'
  const s = Math.max(0, (Date.now() - new Date(iso).getTime()) / 1000)
  if (s < 60) return `${Math.floor(s)}s ago`
  if (s < 3600) return `${Math.floor(s / 60)} min ago`
  if (s < 86400) return `${Math.floor(s / 3600)} h ago`
  return `${Math.floor(s / 86400)} d ago`
 }
 // A public key is long and nobody reads the middle of it; the ends are what you compare.
 const shortKey = (k: string) => (k.length > 16 ? `${k.slice(0, 8)}…${k.slice(-6)}` : k)
 function Peer({ p, server, iface, editable, onChanged }: { p: VPNPeer; server: boolean; iface: string; editable: boolean; onChanged: () => void }) {
  const rename = async () => {
    const name = await askText(`Rename ${p.name || 'this device'}`, p.name ?? '')
    if (!name || name.trim() === p.name) return
    try { await api.editVPNPeer(iface, p.section!, { name: name.trim() }); toast('renamed'); onChanged() }
    catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  const remove = async () => {
    const typed = await askText(`Remove ${p.name}? Its profile stops working at once and cannot be brought back — the device would need a new one.\n\nType its name to confirm.`, '', 'Remove')
    if (typed === null) return
    try { await api.removeVPNPeer(iface, p.section!, typed.trim()); toast(`${p.name} removed`); onChanged() }
    catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  return (
    <div className="card iface vpn-peer">
      <div className="sub-head">
        <span className={'sdot ' + (p.online ? 'ok' : p.handshake ? '' : 'bad')} />
        <span className="dname">{p.name || <span className="muted">unnamed</span>}</span>
        {p.online && <span className="badge ok">online</span>}
        {!p.handshake && <span className="badge">never connected</span>}
        <span className="spacer" />
        <span className="muted small mono">{(p.allowed_ips ?? []).join(', ')}</span>
        {editable && p.section && <>
          <Btn className="btn" onClick={rename}>Rename</Btn>
          <Btn className="btn danger" onClick={remove}>Remove</Btn>
        </>}
      </div>
      <div className="props">
        <div><span className="k">Last handshake</span><span>{ago(p.handshake)}</span></div>
        {p.endpoint && <div><span className="k">{server ? 'Comes from' : 'Endpoint'}</span><span className="mono">{p.endpoint}</span></div>}
        <div><span className="k">Moved</span><span className="mono">↓ {fmtBytes(p.rx)} · ↑ {fmtBytes(p.tx)}</span></div>
        {!!p.keepalive && <div><span className="k">Keepalive</span><span>every {p.keepalive}s</span></div>}
        <div><span className="k">Public key</span><span className="mono" title={p.public_key}>{shortKey(p.public_key)}</span></div>
        {p.psk && <div><span className="k">Pre-shared key</span><span>in use</span></div>}
      </div>
    </div>
  )
 }
 function Tunnel({ t, onChanged }: { t: VPNTunnel; onChanged: () => void }) {
  const server = t.kind === 'server'
  const [adding, setAdding] = useState(false)
  const [made, setMade] = useState<CreatedVPNPeer | null>(null)
  const peers = t.peers ?? []
  const online = peers.filter((p) => p.online).length
  return (
    <>
      <div className="dash-bar">
        <h2 className="section">{t.name}</h2>
        <span className="badge">{server ? 'devices dial in' : t.kind === 'client' ? 'this machine dials out' : 'tunnel'}</span>
        <span className="spacer" />
        <span className="muted small">{server ? `${online} of ${peers.length} online` : t.exit ? `via ${t.exit}` : ''}</span>
        {t.managed && server && <Btn className="btn on" onClick={() => { setMade(null); setAdding(true) }}>+ Add device</Btn>}
      </div>
      <section className="card">
        <div className="props">
          {!!t.addrs?.length && <div><span className="k">Address</span><span className="mono">{t.addrs.join(', ')}</span></div>}
          {!!t.listen_port && <div><span className="k">Listening on</span><span className="mono">{t.listen_port}/udp</span></div>}
          {!!t.mtu && <div><span className="k">MTU</span><span>{t.mtu}</span></div>}
          <div><span className="k">Public key</span><span className="mono" title={t.public_key}>{shortKey(t.public_key)}</span></div>
          {t.fwmark && <div><span className="k">Firewall mark</span><span className="mono">{t.fwmark}</span></div>}
        </div>
      </section>
      {adding && <AddDevice t={t} onClose={() => setAdding(false)} onMade={(c) => { setAdding(false); setMade(c); onChanged() }} />}
      {made && <Profile made={made} onClose={() => setMade(null)} />}
      {peers.map((p) => <Peer key={p.public_key} p={p} server={server} iface={t.name} editable={!!t.managed} onChanged={onChanged} />)}
      {peers.length === 0 && <section className="card"><div className="empty">No peers are configured on this tunnel.</div></section>}
    </>
  )
 }
 // The form for a new device. The endpoint is what the device will dial from outside, so it is
 // the one thing Ham Lab cannot work out for itself — remembered per machine once given.
 function AddDevice({ t, onClose, onMade }: { t: VPNTunnel; onClose: () => void; onMade: (c: CreatedVPNPeer) => void }) {
  const key = `hamlab.wg.endpoint.${t.name}`
  const [name, setName] = useState('')
  const [endpoint, setEndpoint] = useState(localStorage.getItem(key) ?? '')
  const [address, setAddress] = useState('')
  const [dns, setDns] = useState('1.1.1.1, 1.0.0.1')
  const [split, setSplit] = useState(false)
  const [busy, setBusy] = useState(false)
  const [err, setErr] = useState('')
  useEscape(onClose)
  const submit = async (e: FormEvent) => {
    e.preventDefault()
    setBusy(true); setErr('')
    try {
      const made = await api.addVPNPeer(t.name, { name: name.trim(), endpoint: endpoint.trim(), address: address.trim(), dns: dns.trim(), split })
      localStorage.setItem(key, endpoint.trim())
      onMade(made)
    } catch (e) { setErr(e instanceof Error ? e.message : 'failed') }
    finally { setBusy(false) }
  }
  return (
    <form className="card addform" onSubmit={submit} style={{ marginTop: 8 }}>
      <label>Device name <input autoFocus value={name} onChange={(e) => setName(e.target.value)} placeholder="phone" spellCheck={false} required /></label>
      <label>What it dials <input value={endpoint} onChange={(e) => setEndpoint(e.target.value)} placeholder="home.example.com:51820" spellCheck={false} autoCapitalize="off" required /></label>
      <label>Address in the tunnel <input value={address} onChange={(e) => setAddress(e.target.value)} placeholder="the next free one" spellCheck={false} /></label>
      <label>DNS while connected <input value={dns} onChange={(e) => setDns(e.target.value)} placeholder="none" spellCheck={false} /></label>
      <label className="row-check"><input type="checkbox" checked={split} onChange={(e) => setSplit(e.target.checked)} /> Only route the home networks, not everything</label>
      {err && <div className="error">{err}</div>}
      <div className="svc-actions">
        <button className="btn on" disabled={busy || !name.trim() || !endpoint.trim()}>Create device</button>
        <span className="spacer" />
        <button type="button" className="btn" onClick={onClose}>Cancel</button>
      </div>
    </form>
  )
 }
 // The profile, once. Everything a device needs is in it, including a private key that exists
 // nowhere else — so it is shown until dismissed, with the QR the WireGuard apps expect.
 function Profile({ made, onClose }: { made: CreatedVPNPeer; onClose: () => void }) {
  const qr = qrcode(0, 'M')
  qr.addData(made.config)
  qr.make()
  const copy = async () => {
    try { await navigator.clipboard.writeText(made.config); toast('profile copied') }
    catch { toast('select and copy it by hand', 'err') }
  }
  const download = () => {
    const url = URL.createObjectURL(new Blob([made.config], { type: 'text/plain' }))
    const a = document.createElement('a')
    a.href = url
    a.download = `${made.name.replace(/[^a-zA-Z0-9._-]+/g, '-')}.conf`
    a.click()
    URL.revokeObjectURL(url)
  }
  return (
    <section className="card fresh vpn-profile" style={{ marginTop: 8 }}>
      <div className="sub-head"><span className="dname">{made.name}</span><span className="badge ok">created</span><span className="spacer" /><span className="muted small mono">{made.address}</span></div>
      <div className="vpn-qr">
        <img src={qr.createDataURL(5, 8)} alt={`WireGuard profile for ${made.name}`} width={220} height={220} />
        <div>
          <div className="muted small">Scan it in the WireGuard app, or copy the profile into one.</div>
          <pre className="mono vpn-conf">{made.config}</pre>
        </div>
      </div>
      <div className="svc-actions" style={{ marginTop: 8 }}>
        <Btn className="btn on" onClick={copy}>Copy profile</Btn>
        <Btn className="btn" onClick={download}>Download .conf</Btn>
        <span className="muted small">Shown once. The private key in it is not stored here — if it is lost, make the device again.</span>
        <span className="spacer" />
        <button className="btn" onClick={onClose}>Done</button>
      </div>
    </section>
  )
 }
 export default function VPN() {
  const [d, setD] = useState<VPNReport | null>(null)
  const [err, setErr] = useState('')
  const rate = useRefreshMs()
  const load = useCallback(async () => {
    try { setD(await api.vpn()); setErr('') } catch (e) { setErr(e instanceof Error ? e.message : 'failed') }
  }, [])
  useEffect(() => { load() }, [load])
  // Handshakes and counters move; nothing here is expensive to read.
  useEffect(() => {
    const t = setInterval(() => { if (document.visibilityState === 'visible') load() }, Math.max(5000, rate))
    return () => clearInterval(t)
  }, [load, rate])
  if (!d) return <div className="dash">{err ? <div className="notice">{err === 'not found' ? "This machine's Ham Lab is older than the VPN tab. Update it from Machines." : err}</div> : <Loading label="Reading the tunnels…" />}</div>
  // A machine running an older build answers with null rather than an empty list, and taking
  // the length of that is what a blank page is made of.
  const tunnels = d.tunnels ?? []
  return (
    <div className="dash net">
      {tunnels.length === 0 && (
        <>
          <div className="dash-bar"><h2 className="section">VPN</h2></div>
          <section className="card"><div className="empty">{d.note || 'Nothing here.'}</div>
            <div className="muted small" style={{ marginTop: 6 }}>
              Tunnels are the machine's own, so look on the one that carries them — the switcher at the top of the rail.
            </div>
          </section>
        </>
      )}
      {tunnels.map((t) => <Tunnel key={t.name} t={t} onChanged={load} />)}
      {tunnels.length > 0 && (
        <div className="muted small method" style={{ marginTop: 12 }}>
          Read from the kernel with <span className="mono">wg show</span>; names come from the configuration beside it.
          Online means a handshake in the last three minutes — WireGuard is silent when idle, so there is no connection to be in.
          Private and pre-shared keys are never read.
        </div>
      )}
    </div>
  )
 }
@@ -0,0 +1,509 @@
 // Virtualization: LXC containers and KVM virtual machines through Incus, per machine. The
 // page has four states — ready, repair, installable, unavailable — decided by what the
 // selected machine can do, never by a rule about what should run where. Everything the
 // tab makes lives under Ham Lab's own folder on that machine.
 import { FormEvent, Fragment, useCallback, useEffect, useMemo, useState } from 'react'
 import { api, LibraryFile, TaskInfo, VirtImage, VirtInstance, VirtOverview, VirtStatus } from './api'
 import TaskOutput from './TaskOutput'
 import { useHashSel } from './hashsel'
 import { useRefreshMs } from './prefs'
 import { fmtBytes } from './fmt'
 import { askText, askYes, Btn, CloseBtn, Loading, toast, useEscape } from './ui'
 type Tab = 'instances' | 'images' | 'library'
 const GiB = 1 << 30
 const dot = (i: VirtInstance) => (i.status === 'Running' ? 'ok' : i.status === 'Stopped' ? '' : i.status === 'Error' ? 'bad' : 'warn')
 const kind = (i: VirtInstance) => (i.type === 'virtual-machine' ? 'VM' : 'LXC')
 const when = (iso?: string) => { if (!iso) return '—'; const d = new Date(iso); return isNaN(d.getTime()) || d.getFullYear() < 1980 ? '—' : d.toLocaleString() }
 const errMsg = (e: unknown) => (e instanceof Error ? e.message : 'failed')
 const imgLabel = (im: VirtImage) => im.aliases[0] ?? im.fingerprint.slice(0, 12)
 export default function Virtualization() {
  const [ov, setOv] = useState<VirtOverview | null>(null)
  const [err, setErr] = useState('')
  const [tab, setTab] = useState<Tab>('instances')
  const [sel, setSel] = useHashSel('virt')
  const [making, setMaking] = useState<null | { image?: string; vm?: boolean }>(null)
  const [task, setTask] = useState<{ id: string; title: string } | null>(null)
  const [filter, setFilter] = useState('')
  const rate = useRefreshMs()
  const load = useCallback(async () => { try { setOv(await api.virt()); setErr('') } catch (e) { setErr(errMsg(e)) } }, [])
  useEffect(() => { load(); const t = setInterval(() => { if (document.visibilityState === 'visible') load() }, Math.max(5000, rate)); return () => clearInterval(t) }, [load, rate])
  const startTask = (t: TaskInfo) => setTask({ id: t.id, title: t.title })
  if (!ov) return <div className="dash">{err ? <div className="notice">{err === 'not found' ? "This machine's Ham Lab is older than the Virtualization tab. Update it from Machines." : err}</div> : <Loading label="Asking Incus…" />}</div>
  const st = ov.status
  if (st.state !== 'ready') return <Gate st={st} onChanged={load} />
  const instances = ov.instances ?? []
  const images = ov.images ?? []
  const cur = instances.find((i) => i.name === sel)
  const running = instances.filter((i) => i.status === 'Running').length
  const f = filter.trim().toLowerCase()
  const shown = instances.filter((i) => !f || `${i.name} ${i.image} ${i.type} ${i.status}`.toLowerCase().includes(f))
  const libCount = ov.library.images.length + ov.library.isos.length
  return (
    <div className={'svc virt' + (cur && tab === 'instances' ? ' has-sel' : '') + (tab !== 'instances' ? ' wide' : '')}>
      <div className="svc-list">
        <div className="svc-tools">
          <div className="chips">
            {(['instances', 'images', 'library'] as Tab[]).map((t) => (
              <button key={t} className={'chip' + (tab === t ? ' on' : '')} onClick={() => { setTab(t); if (sel) setSel('') }}>
                {t}{t === 'instances' ? ` · ${running}/${instances.length}` : t === 'images' ? ` · ${images.length}` : ` · ${libCount}`}
              </button>
            ))}
          </div>
          {tab === 'instances' && (
            <div className="svc-actions">
              <button className="btn on" onClick={() => setMaking({})}>+ New instance</button>
              <span className="spacer" />
              <span className="muted small">{st.kvm ? 'containers and VMs' : 'containers only — no KVM on this machine'}</span>
            </div>
          )}
          {tab === 'instances' && instances.length > 3 && <input value={filter} onChange={(e) => setFilter(e.target.value)} placeholder="filter instances" />}
          {ov.error && <div className="error">{ov.error}</div>}
          {task && <TaskOutput id={task.id} onDone={(t) => { load(); if (t.exit_code === 0 && !t.error) toast(`${task.title}: done`); else toast(`${task.title}: failed`, 'err') }} />}
          {task && <div className="svc-actions"><button className="btn" onClick={() => setTask(null)}>Hide output</button></div>}
        </div>
        {tab === 'instances' && (
          <div className="svc-rows">
            {shown.map((i) => (
              <div key={i.name} className={'svc-row' + (i.name === sel ? ' active' : '')} onClick={() => setSel(i.name)}>
                <span className={'sdot ' + dot(i)} />
                <div className="svc-name"><div className="n">{i.name} <span className="badge">{kind(i)}</span></div><div className="d">{i.image || '—'}{i.ipv4.length ? ` · ${i.ipv4.join(' ')}` : ''}</div></div>
                <div className="svc-state">{i.status === 'Running' ? <>{fmtBytes(i.mem_usage)}<br />{i.cpu_pct.toFixed(0)}%</> : i.status.toLowerCase()}</div>
              </div>
            ))}
            {instances.length === 0 && <div className="muted small" style={{ padding: 12 }}>Nothing here yet. “+ New instance” makes a container{st.kvm ? ' or a VM' : ''} from the catalogue.</div>}
          </div>
        )}
        {tab === 'images' && <Images st={st} images={images} instances={instances} onTask={startTask} onNew={(image, vm) => { setTab('instances'); setMaking({ image, vm }) }} onChanged={load} />}
        {tab === 'library' && <Library lib={ov.library} dir={st.dir} />}
      </div>
      {tab === 'instances' && (cur ? <Detail i={cur} st={st} onBack={() => setSel('')} onChanged={load} onTask={startTask} /> : <Summary ov={ov} />)}
      {making && <NewInstance st={st} ov={ov} preset={making} onClose={() => setMaking(null)} onStarted={(t) => { setMaking(null); startTask(t) }} />}
    </div>
  )
 }
 // ---- the three states before "ready" ----
 function Gate({ st, onChanged }: { st: VirtStatus; onChanged: () => void }) {
  const [open, setOpen] = useState(false)
  const [task, setTask] = useState('')
  const [waiting, setWaiting] = useState(false)
  // the install ends by restarting Ham Lab, so poll until it answers again
  useEffect(() => {
    if (!waiting) return
    const t = setInterval(async () => { try { const o = await api.virt(); if (o.status.state === 'ready') { setWaiting(false); onChanged() } } catch { /* restarting */ } }, 3000)
    return () => clearInterval(t)
  }, [waiting, onChanged])
  const headline = st.state === 'installable'
    ? `Incus isn't installed on this machine. Ham Lab can install it with ${st.pm}: ${st.kvm ? 'containers and virtual machines' : 'containers only, since this kernel has no KVM'}.`
    : st.reason
  return (
    <div className="dash">
      <div className="dash-bar"><h2 className="section">Virtualization</h2></div>
      <section className="card">
        <div className="notice">{headline}</div>
        <div className="props" style={{ marginTop: 8 }}>
          <div><span className="k">KVM</span><span>{st.kvm ? 'yes — virtual machines possible' : 'no — containers only'}</span></div>
          <div><span className="k">Architecture</span><span className="mono">{st.arch}</span></div>
          {st.pm && <div><span className="k">Packages</span><span>{st.pm}</span></div>}
          <div><span className="k">Pool drivers</span><span>{st.drivers.join(', ')}</span></div>
          <div><span className="k">Memory</span><span>{fmtBytes(st.mem_avail)} available of {fmtBytes(st.mem_total)}</span></div>
          <div><span className="k">Disk</span><span>{fmtBytes(st.disk_free)} free where Ham Lab's data lives</span></div>
          <div><span className="k">Docker</span><span>{st.docker ? 'present — its FORWARD policy needs two rules, Install adds them' : 'not here'}</span></div>
          <div><span className="k">Data</span><span className="mono small">{st.dir}</span></div>
          {st.layout && <div><span className="k">Last attempt</span><span>{when(st.layout.attempted)} · {st.layout.driver} pool · {st.layout.subnet}</span></div>}
        </div>
        {(st.state === 'installable' || st.state === 'repair') && !task && (
          <div className="svc-actions"><Btn className="btn on" onClick={() => setOpen(true)}>{st.state === 'repair' ? 'Finish setup' : 'Install Incus'}</Btn></div>
        )}
        {task && <TaskOutput id={task} onDone={(t) => { if (t.exit_code === 0 && !t.error) setWaiting(true); else { toast('install failed — the output says where', 'err'); setTask('') } }} />}
        {waiting && <div className="muted small" style={{ marginTop: 8 }}>Ham Lab is restarting to pick up its new group. The page reconnects by itself.</div>}
      </section>
      {open && <InstallDialog st={st} onClose={() => setOpen(false)} onStarted={(id) => { setOpen(false); setTask(id) }} />}
    </div>
  )
 }
 function InstallDialog({ st, onClose, onStarted }: { st: VirtStatus; onClose: () => void; onStarted: (id: string) => void }) {
  const [driver, setDriver] = useState(st.layout?.driver ?? st.drivers[0] ?? 'dir')
  const [size, setSize] = useState(String(st.layout?.pool_size_gib || st.pool_size_gib || 20))
  const [subnet, setSubnet] = useState(st.layout?.subnet || st.subnet || '172.16.100.1/24')
  const [docker, setDocker] = useState(st.docker)
  const [arc, setArc] = useState(st.layout?.arc_max_gib ? String(st.layout.arc_max_gib) : '')
  const [err, setErr] = useState('')
  const [busy, setBusy] = useState(false)
  useEscape(onClose)
  const submit = async (e: FormEvent) => {
    e.preventDefault(); setErr(''); setBusy(true)
    try { const t = await api.virtInstall({ driver, pool_size_gib: Number(size) || 0, subnet: subnet.trim(), docker_rules: docker, arc_max_gib: Number(arc) || 0 }); onStarted(t.id) }
    catch (e) { setErr(errMsg(e)); setBusy(false) }
  }
  return (
    <div className="modal-back" onClick={onClose}>
      <form className="modal" onClick={(e) => e.stopPropagation()} onSubmit={submit}>
        <h3>{st.state === 'repair' ? 'Finish setting up Incus' : 'Install Incus'}</h3>
        <div className="notice">Everything goes under <span className="mono">{st.dir}</span>: the pool's backing file, ISOs, images, exports. Where is not a choice; how big is.</div>
        <label className="muted">Pool driver
          <select value={driver} onChange={(e) => setDriver(e.target.value)}>
            {st.drivers.map((d) => <option key={d} value={d}>{d}{d === 'zfs' ? ' — snapshots and clones are instant' : d === 'btrfs' ? ' — snapshots, on a loop file' : ' — plain files, no snapshots to speak of'}</option>)}
          </select>
        </label>
        {driver !== 'dir' && (
          <label className="muted">Pool size, GiB — a sparse file, grows as used · {fmtBytes(st.disk_free)} free
            <input type="number" min={5} max={Math.floor(st.disk_free / GiB)} value={size} onChange={(e) => setSize(e.target.value)} required />
          </label>
        )}
        <label className="muted">Bridge address — picked clear of this machine's routes and Docker's pools
          <input className="mono" value={subnet} onChange={(e) => setSubnet(e.target.value)} spellCheck={false} required />
        </label>
        {st.docker && (
          <label className="muted" style={{ flexDirection: 'row', alignItems: 'center', gap: 8 }}>
            <input type="checkbox" checked={docker} onChange={(e) => setDocker(e.target.checked)} style={{ width: 'auto' }} />
            Add the two DOCKER-USER rules that let the bridge's traffic through Docker's FORWARD drop
          </label>
        )}
        {st.zfs_host && (
          <label className="muted">Cap the ZFS ARC, GiB — blank leaves it at {st.arc_max_gib} GiB; a cap makes guest memory deterministic
            <input type="number" min={1} max={Math.floor(st.mem_total / GiB)} value={arc} onChange={(e) => setArc(e.target.value)} placeholder="leave as is" />
          </label>
        )}
        <div className="muted small">Ham Lab's account ({st.user}) joins the incus-admin group and Ham Lab restarts at the end.</div>
        {err && <div className="error">{err}</div>}
        <div className="svc-actions">
          <button type="submit" className="btn on" disabled={busy}>{busy ? 'Starting…' : 'Install'}</button>
          <span className="spacer" />
          <button type="button" className="btn" onClick={onClose}>Cancel</button>
        </div>
      </form>
    </div>
  )
 }
 // ---- the right-hand pane when nothing is selected: what this machine's Incus is ----
 function Summary({ ov }: { ov: VirtOverview }) {
  const st = ov.status
  return (
    <div className="svc-detail">
      <div className="svc-head"><span className="dname">Incus {st.version}</span><span className="badge">{st.arch}</span>{!st.kvm && <span className="badge warn">no KVM</span>}</div>
      <div className="props">
        <div><span className="k">Data</span><span className="mono small">{st.dir}</span></div>
        <div><span className="k">Memory</span><span>{fmtBytes(st.mem_avail)} available of {fmtBytes(st.mem_total)}</span></div>
        {st.layout && <div><span className="k">Installed</span><span>{when(st.layout.attempted)} by Ham Lab {st.layout.hamlab_version ?? ''}</span></div>}
        {!st.layout && <div><span className="k">Installed</span><span>outside Ham Lab — pools and networks below are as found</span></div>}
      </div>
      <h3 className="stack-head">Storage pools</h3>
      <table className="virt-table"><thead><tr><th>Pool</th><th>Driver</th><th>Source</th><th className="num">Used</th><th className="num">Of</th><th className="num">Users</th></tr></thead>
        <tbody>{(ov.pools ?? []).map((p) => <tr key={p.name}><td>{p.name}</td><td>{p.driver}</td><td className="mono small">{p.source}</td><td className="num">{p.total ? fmtBytes(p.used) : '—'}</td><td className="num">{p.total ? fmtBytes(p.total) : '—'}</td><td className="num">{p.used_by}</td></tr>)}</tbody></table>
      <h3 className="stack-head">Networks</h3>
      <table className="virt-table"><thead><tr><th>Network</th><th>Type</th><th>Address</th><th>NAT</th><th className="num">Users</th></tr></thead>
        <tbody>{(ov.networks ?? []).filter((n) => n.managed).map((n) => <tr key={n.name}><td>{n.name}</td><td>{n.type}</td><td className="mono small">{n.address ?? '—'}</td><td>{n.nat ? 'yes' : 'no'}</td><td className="num">{n.used_by}</td></tr>)}</tbody></table>
      <div className="muted small">Pick an instance on the left, or make one. Consoles open as tabs on the Terminal page.</div>
    </div>
  )
 }
 // ---- one instance ----
 function Detail({ i, st, onBack, onChanged, onTask }: { i: VirtInstance; st: VirtStatus; onBack: () => void; onChanged: () => void; onTask: (t: TaskInfo) => void }) {
  const [tab, setTab] = useState<'overview' | 'snapshots' | 'config'>('overview')
  const running = i.status === 'Running'
  const stopped = i.status === 'Stopped'
  const act = (a: 'start' | 'stop' | 'restart' | 'kill') => async () => { try { await api.virtAction(i.name, a); toast(`${i.name}: ${a} ok`); onChanged() } catch (e) { toast(`${i.name}: ${errMsg(e)}`, 'err') } }
  const term = (k: 'console' | 'exec') => async () => {
    try { const t = await api.virtTerminal(i.name, k); toast(`${k === 'exec' ? 'shell' : 'console'} open as “${t.name}” on the Terminal tab`); window.dispatchEvent(new CustomEvent('hamlab:terminal', { detail: t.id })) }
    catch (e) { toast(errMsg(e), 'err') }
  }
  const rename = async () => { const to = await askText(`Rename ${i.name} to`, i.name, 'Rename'); if (!to || to === i.name) return; try { await api.virtRename(i.name, to); toast('renamed'); onBack(); onChanged() } catch (e) { toast(errMsg(e), 'err') } }
  const del = async () => {
    const typed = await askText(`Delete ${i.name}${running ? ', which is running' : ''}? Its disk and ${i.snapshots.length} snapshot${i.snapshots.length === 1 ? '' : 's'} go with it. Type its name to confirm.`, '', 'Delete')
    if (typed === null) return
    if (typed !== i.name) { toast('name did not match', 'err'); return }
    try { await api.virtDelete(i.name, running); toast(`${i.name} deleted`); onBack(); onChanged() } catch (e) { toast(errMsg(e), 'err') }
  }
  const publish = async () => {
    const alias = await askText(`Publish ${i.name} as an image — an alias for it`, `${i.name}-image`, 'Publish')
    if (!alias) return
    try { onTask(await api.virtImagePublish(i.name, alias)) } catch (e) { toast(errMsg(e), 'err') }
  }
  const snap = async () => {
    const name = await askText(`Snapshot ${i.name} — a name, or blank for the date`, '', 'Snapshot')
    if (name === null) return
    try { const r = await api.virtSnapshot(i.name, name); toast(`snapshot ${r.name} taken`); onChanged() } catch (e) { toast(errMsg(e), 'err') }
  }
  return (
    <div className="svc-detail">
      <CloseBtn onClick={onBack} />
      <div className="svc-head"><button className="btn back" onClick={onBack}>‹ Back</button><span className={'sdot ' + dot(i)} /><span className="dname">{i.name}</span><span className="badge">{kind(i)}</span><span className={'badge' + (i.status === 'Error' ? ' bad' : '')}>{i.status.toLowerCase()}</span></div>
      <div className="props">
        <div><span className="k">Image</span><span>{i.image || '—'}{i.arch ? <span className="muted"> · {i.arch}</span> : ''}</span></div>
        {i.description && <div><span className="k">About</span><span>{i.description}</span></div>}
        {running && <div><span className="k">Using</span><span>{i.cpu_pct.toFixed(0)}% of a core · {fmtBytes(i.mem_usage)}{i.mem_peak ? ` (peak ${fmtBytes(i.mem_peak)})` : ''}{i.processes ? ` · ${i.processes} processes` : ''}</span></div>}
        <div><span className="k">Limits</span><span>{i.cpu ? `${i.cpu} CPU` : 'all CPUs'} · {i.memory || 'unlimited memory'}</span></div>
        {i.disk_total > 0 && <div><span className="k">Disk</span><span>{fmtBytes(i.disk_usage)} of {fmtBytes(i.disk_total)}</span></div>}
        <div><span className="k">Network</span><span className="mono small">{i.ipv4.length ? i.ipv4.join('  ') : running ? 'no address yet' : '—'}</span></div>
        <div><span className="k">Created</span><span>{when(i.created)}</span></div>
      </div>
      <div className="svc-actions">
        {!running && <Btn className="btn on" onClick={act('start')}>Start</Btn>}
        {running && <Btn className="btn" onClick={term('exec')}>Shell</Btn>}
        <Btn className="btn" onClick={term('console')} title="the serial console: boot output, and a login prompt where the guest has one on it">Console</Btn>
        {running && <Btn className="btn" confirm={`Restart ${i.name}?`} onClick={act('restart')}>Restart</Btn>}
        {running && <Btn className="btn danger" confirm={`Stop ${i.name}? It is asked to shut down cleanly.`} onClick={act('stop')}>Stop</Btn>}
        {running && <Btn className="btn danger" confirm={`Force stop ${i.name}? Like pulling the plug — anything unsaved inside is lost.`} onClick={act('kill')}>Force stop</Btn>}
        <Btn className="btn" onClick={snap}>Snapshot</Btn>
        {stopped && <Btn className="btn" onClick={rename}>Rename</Btn>}
        {stopped && <Btn className="btn" onClick={publish} title="make an image of it, to launch more like it">Publish</Btn>}
        <Btn className="btn danger" onClick={del}>Delete…</Btn>
      </div>
      <div className="tabs">
        <button className={tab === 'overview' ? 'on' : ''} onClick={() => setTab('overview')}>Limits</button>
        <button className={tab === 'snapshots' ? 'on' : ''} onClick={() => setTab('snapshots')}>Snapshots · {i.snapshots.length}</button>
        <button className={tab === 'config' ? 'on' : ''} onClick={() => setTab('config')}>Config</button>
      </div>
      {tab === 'overview' && <Limits i={i} st={st} onChanged={onChanged} />}
      {tab === 'snapshots' && <Snapshots i={i} onChanged={onChanged} />}
      {tab === 'config' && (
        <>
          <div className="virt-kv">{Object.entries(i.config).sort().map(([k, v]) => <Fragment key={k}><span className="muted mono">{k}</span><span className="mono">{v}</span></Fragment>)}</div>
          <h3 className="stack-head">Devices</h3>
          <div className="virt-kv">{Object.entries(i.devices).sort().map(([d, cfg]) => <Fragment key={d}><span className="muted mono">{d}</span><span className="mono">{Object.entries(cfg).map(([k, v]) => `${k}=${v}`).join('  ')}</span></Fragment>)}</div>
          <div className="muted small">Profiles: {i.profiles.join(', ') || '—'}</div>
        </>
      )}
    </div>
  )
 }
 function Limits({ i, st, onChanged }: { i: VirtInstance; st: VirtStatus; onChanged: () => void }) {
  const [cpu, setCpu] = useState(i.cpu ?? '')
  const [mem, setMem] = useState(i.memory ?? '')
  const [auto, setAuto] = useState(i.config['boot.autostart'] ?? '')
  const [desc, setDesc] = useState(i.description ?? '')
  useEffect(() => { setCpu(i.cpu ?? ''); setMem(i.memory ?? ''); setAuto(i.config['boot.autostart'] ?? ''); setDesc(i.description ?? '') }, [i.name, i.cpu, i.memory, i.description, i.config])
  const changed = cpu !== (i.cpu ?? '') || mem !== (i.memory ?? '') || auto !== (i.config['boot.autostart'] ?? '') || desc !== (i.description ?? '')
  const save = async () => { try { await api.virtConfig(i.name, { 'limits.cpu': cpu, 'limits.memory': mem, 'boot.autostart': auto, description: desc }); toast('saved'); onChanged() } catch (e) { toast(errMsg(e), 'err') } }
  return (
    <section className="card">
      <div className="virt-form-row">
        <label className="muted">CPUs<input className="mono" value={cpu} onChange={(e) => setCpu(e.target.value)} placeholder="all" /></label>
        <label className="muted">Memory<input className="mono" value={mem} onChange={(e) => setMem(e.target.value)} placeholder="unlimited · e.g. 2GiB" /></label>
      </div>
      <div className="virt-form-row">
        <label className="muted">Start with the machine
          <select value={auto} onChange={(e) => setAuto(e.target.value)}><option value="">default (as it was when the host went down)</option><option value="true">yes</option><option value="false">no</option></select>
        </label>
        <label className="muted">Description<input value={desc} onChange={(e) => setDesc(e.target.value)} maxLength={200} /></label>
      </div>
      <div className="muted small">{i.type === 'container' ? 'Containers take new limits live.' : 'A VM takes CPU and memory changes live where its agent is running, otherwise on its next start.'} {fmtBytes(st.mem_avail)} is free on this machine right now.</div>
      <div className="svc-actions"><Btn className="btn on" disabled={!changed} onClick={save}>Save</Btn></div>
    </section>
  )
 }
 function Snapshots({ i, onChanged }: { i: VirtInstance; onChanged: () => void }) {
  const [sched, setSched] = useState(i.schedule ?? '')
  const [exp, setExp] = useState(i.expiry ?? '')
  useEffect(() => { setSched(i.schedule ?? ''); setExp(i.expiry ?? '') }, [i.name, i.schedule, i.expiry])
  const saveSched = async () => { try { await api.virtConfig(i.name, { 'snapshots.schedule': sched.trim(), 'snapshots.expiry': exp.trim() }); toast('schedule saved'); onChanged() } catch (e) { toast(errMsg(e), 'err') } }
  const on = async (s: string, a: 'restore' | 'delete') => {
    const ok = await askYes(a === 'restore' ? `Restore ${i.name} to snapshot ${s}? Everything since it is lost${i.status === 'Running' ? '; it is stopped first' : ''}.` : `Delete snapshot ${s} of ${i.name}?`, a === 'restore' ? 'Restore' : 'Delete')
    if (!ok) return
    try { await api.virtSnapshotAction(i.name, s, a); toast(`${a} ${s}: ok`); onChanged() } catch (e) { toast(errMsg(e), 'err') }
  }
  return (
    <>
      {i.snapshots.length === 0 && <div className="muted small">No snapshots. “Snapshot” above takes one now; a schedule below keeps taking them.</div>}
      {i.snapshots.length > 0 && (
        <table className="virt-table"><thead><tr><th>Snapshot</th><th>Taken</th><th>Expires</th><th className="num">Size</th><th /></tr></thead>
          <tbody>{[...i.snapshots].sort((a, b) => b.created.localeCompare(a.created)).map((s) => (
            <tr key={s.name}><td className="mono">{s.name}</td><td>{when(s.created)}</td><td>{s.expires ? when(s.expires) : 'never'}</td><td className="num">{s.size ? fmtBytes(s.size) : '—'}</td>
              <td className="num"><Btn className="btn" onClick={() => on(s.name, 'restore')}>Restore</Btn> <Btn className="btn danger" onClick={() => on(s.name, 'delete')}>Delete</Btn></td></tr>
          ))}</tbody></table>
      )}
      <section className="card">
        <div className="virt-form-row">
          <label className="muted">Schedule — @hourly, @daily, @weekly, or a cron line<input className="mono" value={sched} onChange={(e) => setSched(e.target.value)} placeholder="none" spellCheck={false} /></label>
          <label className="muted">Keep for — 1d, 2w, 3M…<input className="mono" value={exp} onChange={(e) => setExp(e.target.value)} placeholder="forever" spellCheck={false} /></label>
        </div>
        <div className="muted small">Incus takes and expires these itself, so they carry on whether or not this console is up.</div>
        <div className="svc-actions"><Btn className="btn on" disabled={sched === (i.schedule ?? '') && exp === (i.expiry ?? '')} onClick={saveSched}>Save schedule</Btn></div>
      </section>
    </>
  )
 }
 // ---- making one ----
 function NewInstance({ st, ov, preset, onClose, onStarted }: { st: VirtStatus; ov: VirtOverview; preset: { image?: string; vm?: boolean }; onClose: () => void; onStarted: (t: TaskInfo) => void }) {
  const [vm, setVm] = useState(!!preset.vm && st.kvm)
  const [image, setImage] = useState(preset.image ?? '')
  const [name, setName] = useState('')
  const [cpu, setCpu] = useState('')
  const [mem, setMem] = useState('')
  const [disk, setDisk] = useState('')
  const nets = (ov.networks ?? []).filter((n) => n.managed || n.type === 'bridge')
  const [net, setNet] = useState(nets.find((n) => n.name === 'incusbr0')?.name ?? nets[0]?.name ?? '')
  const [cloud, setCloud] = useState('')
  const [showCloud, setShowCloud] = useState(false)
  const [desc, setDesc] = useState('')
  const [start, setStart] = useState(true)
  const [q, setQ] = useState('')
  const [remote, setRemote] = useState<VirtImage[] | null>(null)
  const [remoteErr, setRemoteErr] = useState('')
  const [err, setErr] = useState('')
  const [busy, setBusy] = useState(false)
  const type = vm ? 'virtual-machine' : 'container'
  useEscape(onClose)
  useEffect(() => { setRemote(null); setRemoteErr(''); api.virtRemote(type).then((r) => setRemote(r.images)).catch((e) => setRemoteErr(errMsg(e))) }, [type])
  const local = (ov.images ?? []).filter((im) => im.type === type)
  const picks = useMemo(() => {
    const f = q.trim().toLowerCase()
    const match = (im: VirtImage) => !f || `${im.aliases.join(' ')} ${im.description} ${im.os} ${im.release}`.toLowerCase().includes(f)
    const seen = new Set<string>()
    const out: { ref: string; label: string; sub: string; here: boolean }[] = []
    for (const im of local) { if (!match(im)) continue; const ref = imgLabel(im); if (seen.has(ref)) continue; seen.add(ref); out.push({ ref, label: ref, sub: `${im.description} · on this machine · ${fmtBytes(im.size)}`, here: true }) }
    for (const im of remote ?? []) {
      const a = im.aliases.find((x) => !x.includes('/default')) ?? im.aliases[0]
      if (!a || !match(im)) continue
      const ref = `images:${a}`
      if (seen.has(ref)) continue
      seen.add(ref); out.push({ ref, label: a, sub: `${im.description}${im.variant && im.variant !== 'default' ? ` · ${im.variant}` : ''} · ${fmtBytes(im.size)}`, here: false })
    }
    return out.slice(0, 60)
  }, [q, local, remote])
  const submit = async (e: FormEvent) => {
    e.preventDefault(); setErr(''); setBusy(true)
    try { onStarted(await api.virtCreate({ name: name.trim(), image: image.trim(), vm, cpu: cpu.trim(), memory: mem.trim(), disk: disk.trim(), network: net, cloud_init: showCloud ? cloud : '', profiles: [], description: desc.trim(), start })) }
    catch (e) { setErr(errMsg(e)); setBusy(false) }
  }
  return (
    <div className="modal-back" onClick={onClose}>
      <form className="modal" onClick={(e) => e.stopPropagation()} onSubmit={submit}>
        <h3>New instance</h3>
        <div className="chips">
          <button type="button" className={'chip' + (!vm ? ' on' : '')} onClick={() => setVm(false)}>container</button>
          <button type="button" className={'chip' + (vm ? ' on' : '')} disabled={!st.kvm} title={st.kvm ? '' : 'no KVM on this machine'} onClick={() => setVm(true)}>virtual machine</button>
        </div>
        <label className="muted">Image — from this machine, or the public catalogue for {st.arch}
          <input className="mono" value={image} onChange={(e) => { setImage(e.target.value); setQ(e.target.value.replace(/^images:/, '')) }} placeholder="images:debian/13" spellCheck={false} autoCapitalize="off" autoCorrect="off" required />
        </label>
        <div className="virt-pick">
          {remote === null && !remoteErr && <div className="muted small" style={{ padding: 8 }}>Reading the catalogue…</div>}
          {remoteErr && <div className="error" style={{ margin: 8 }}>catalogue: {remoteErr}</div>}
          {picks.map((p) => (
            <div key={p.ref} className={'svc-row' + (image === p.ref ? ' active' : '')} onClick={() => setImage(p.ref)}>
              <div className="svc-name"><div className="n">{p.label}{p.here && <span className="badge">here</span>}</div><div className="d">{p.sub}</div></div>
            </div>
          ))}
          {remote && picks.length === 0 && <div className="muted small" style={{ padding: 8 }}>Nothing in the catalogue matches.</div>}
        </div>
        <div className="virt-form-row">
          <label className="muted">Name<input value={name} onChange={(e) => setName(e.target.value)} pattern="[a-zA-Z][a-zA-Z0-9\-]{0,62}" placeholder="web-1" spellCheck={false} autoCapitalize="off" autoCorrect="off" required /></label>
          <label className="muted">Network<select value={net} onChange={(e) => setNet(e.target.value)}>{nets.map((n) => <option key={n.name} value={n.name}>{n.name}{n.address ? ` · ${n.address}` : ''}</option>)}</select></label>
        </div>
        <div className="virt-form-row">
          <label className="muted">CPUs<input className="mono" value={cpu} onChange={(e) => setCpu(e.target.value)} placeholder={vm ? '2' : 'all'} /></label>
          <label className="muted">Memory<input className="mono" value={mem} onChange={(e) => setMem(e.target.value)} placeholder={vm ? '2GiB' : 'unlimited'} /></label>
          <label className="muted">Root disk<input className="mono" value={disk} onChange={(e) => setDisk(e.target.value)} placeholder={vm ? '20GiB' : 'pool default'} /></label>
        </div>
        <div className="muted small">{fmtBytes(st.mem_avail)} free on this machine.{vm ? ' A VM’s memory is a reservation.' : ''}</div>
        <label className="muted">Description — optional<input value={desc} onChange={(e) => setDesc(e.target.value)} maxLength={200} /></label>
        <label className="muted" style={{ flexDirection: 'row', alignItems: 'center', gap: 8 }}>
          <input type="checkbox" checked={showCloud} onChange={(e) => setShowCloud(e.target.checked)} style={{ width: 'auto' }} />cloud-init user-data (an SSH key, packages, a first command — needs a /cloud image)
        </label>
        {showCloud && <textarea className="mono" rows={6} value={cloud} onChange={(e) => setCloud(e.target.value)} placeholder={'#cloud-config\nusers:\n  - name: core\n    ssh_authorized_keys:\n      - ssh-ed25519 AAAA…\n    sudo: ALL=(ALL) NOPASSWD:ALL\npackages: [curl]'} spellCheck={false} />}
        <label className="muted" style={{ flexDirection: 'row', alignItems: 'center', gap: 8 }}>
          <input type="checkbox" checked={start} onChange={(e) => setStart(e.target.checked)} style={{ width: 'auto' }} />Start it once made
        </label>
        {err && <div className="error">{err}</div>}
        <div className="svc-actions">
          <button type="submit" className="btn on" disabled={busy || !image.trim() || !name.trim()}>{busy ? 'Starting…' : start ? 'Create and start' : 'Create'}</button>
          <span className="spacer" />
          <button type="button" className="btn" onClick={onClose}>Cancel</button>
        </div>
      </form>
    </div>
  )
 }
 // ---- images: the local store, the catalogue, and files brought in ----
 function Images({ st, images, instances, onTask, onNew, onChanged }: { st: VirtStatus; images: VirtImage[]; instances: VirtInstance[]; onTask: (t: TaskInfo) => void; onNew: (image: string, vm: boolean) => void; onChanged: () => void }) {
  const [q, setQ] = useState('')
  const [type, setType] = useState<'container' | 'virtual-machine'>('container')
  const [remote, setRemote] = useState<VirtImage[] | null>(null)
  const [remoteErr, setRemoteErr] = useState('')
  const [path, setPath] = useState('')
  const [alias, setAlias] = useState('')
  const [itype, setItype] = useState<'container' | 'virtual-machine'>('container')
  const loadRemote = useCallback((refresh = false) => { setRemote(null); setRemoteErr(''); api.virtRemote(type, refresh).then((r) => setRemote(r.images)).catch((e) => setRemoteErr(errMsg(e))) }, [type])
  useEffect(() => { loadRemote() }, [loadRemote])
  const usedBy = (im: VirtImage) => instances.filter((i) => i.config['volatile.base_image'] === im.fingerprint || im.aliases.includes(i.image)).map((i) => i.name)
  const f = q.trim().toLowerCase()
  const found = (remote ?? []).filter((im) => f && `${im.aliases.join(' ')} ${im.description} ${im.os} ${im.release} ${im.variant}`.toLowerCase().includes(f)).slice(0, 40)
  const del = async (im: VirtImage) => {
    if (!(await askYes(`Delete the image ${imgLabel(im)} from this machine's store? Instances made from it are not affected; the file in the library, if any, stays.`, 'Delete'))) return
    try { await api.virtImageDelete(im.fingerprint); toast('image deleted'); onChanged() } catch (e) { toast(errMsg(e), 'err') }
  }
  const imp = async (e: FormEvent) => { e.preventDefault(); try { onTask(await api.virtImageImport(path.trim(), alias.trim(), itype)); setPath(''); setAlias('') } catch (er) { toast(errMsg(er), 'err') } }
  return (
    <div className="svc-rows" style={{ padding: 12, display: 'flex', flexDirection: 'column', gap: 14 }}>
      <section className="card">
        <div className="sub-head"><span className="dname">On this machine</span><span className="spacer" /><span className="muted small">{images.length} image{images.length === 1 ? '' : 's'} · {fmtBytes(images.reduce((a, im) => a + im.size, 0))}</span></div>
        {images.length === 0 && <div className="muted small">No images yet. Make an instance and its image is kept here; or download one below so the next one is instant.</div>}
        {images.length > 0 && (
          <table className="virt-table"><thead><tr><th>Image</th><th>Type</th><th>What</th><th className="num">Size</th><th>Used by</th><th /></tr></thead>
            <tbody>{images.map((im) => (
              <tr key={im.fingerprint}><td className="mono">{im.aliases.join(', ') || im.fingerprint.slice(0, 12)}</td><td>{im.type === 'virtual-machine' ? 'VM' : 'LXC'}</td><td>{im.description}{im.cached ? <span className="muted small"> · cached from {im.source}</span> : ''}</td><td className="num">{fmtBytes(im.size)}</td><td className="small">{usedBy(im).join(' ') || '—'}</td>
                <td className="num"><Btn className="btn" onClick={() => onNew(imgLabel(im), im.type === 'virtual-machine')}>New instance</Btn> <Btn className="btn danger" onClick={() => del(im)}>Delete</Btn></td></tr>
            ))}</tbody></table>
        )}
      </section>
      <section className="card">
        <div className="sub-head"><span className="dname">Download from the catalogue</span><span className="spacer" />
          <div className="chips"><button className={'chip' + (type === 'container' ? ' on' : '')} onClick={() => setType('container')}>containers</button><button className={'chip' + (type === 'virtual-machine' ? ' on' : '')} disabled={!st.kvm} onClick={() => setType('virtual-machine')}>VMs</button></div>
          <Btn className="btn" onClick={() => loadRemote(true)} title="the catalogue is kept for an hour">Refresh</Btn></div>
        <input value={q} onChange={(e) => setQ(e.target.value)} placeholder={`search images.linuxcontainers.org for ${st.arch} — debian, ubuntu 24.04, alpine…`} />
        {remote === null && !remoteErr && <div className="muted small">Reading the catalogue…</div>}
        {remoteErr && <div className="error">{remoteErr}</div>}
        {remote && !f && <div className="muted small">{remote.length} {type === 'virtual-machine' ? 'VM' : 'container'} images for {st.arch}. Type to search. A download keeps the image in the store and a copy of it in the library.</div>}
        {found.length > 0 && (
          <div className="virt-pick">{found.map((im) => { const a = im.aliases.find((x) => !x.includes('/default')) ?? im.aliases[0]; return (
            <div key={im.fingerprint} className="svc-row" style={{ cursor: 'default' }}>
              <div className="svc-name"><div className="n">{a}</div><div className="d">{im.description}{im.variant && im.variant !== 'default' ? ` · ${im.variant}` : ''} · {fmtBytes(im.size)}</div></div>
              <Btn className="btn" onClick={async () => { try { onTask(await api.virtImageDownload(a)) } catch (e) { toast(errMsg(e), 'err') } }}>Download</Btn>
            </div>) })}</div>
        )}
      </section>
      <section className="card">
        <div className="sub-head"><span className="dname">Import a file from this machine</span></div>
        <form onSubmit={imp} className="virt-form-row" style={{ flexWrap: 'wrap' }}>
          <label className="muted" style={{ flex: '2 1 240px' }}>Path — a template tarball, or a qcow2 / raw disk for a VM<input className="mono" value={path} onChange={(e) => setPath(e.target.value)} placeholder="/path/to/image.tar.xz" spellCheck={false} required /></label>
          <label className="muted">Alias<input className="mono" value={alias} onChange={(e) => setAlias(e.target.value)} placeholder="my-template" spellCheck={false} required /></label>
          <label className="muted">Type<select value={itype} onChange={(e) => setItype(e.target.value as 'container' | 'virtual-machine')}><option value="container">container template</option><option value="virtual-machine" disabled={!st.kvm}>VM disk image</option></select></label>
          <div className="svc-actions" style={{ alignSelf: 'flex-end' }}><button className="btn on" disabled={!path.trim() || !alias.trim()}>Import</button></div>
        </form>
        <div className="muted small">The file is copied into <span className="mono">{st.dir}/images/</span> first — the library is Ham Lab's — then loaded into Incus.</div>
      </section>
    </div>
  )
 }
 function Library({ lib, dir }: { lib: { images: LibraryFile[]; isos: LibraryFile[] }; dir: string }) {
  const table = (files: LibraryFile[]) => files.length === 0 ? <div className="muted small">empty</div> : (
    <table className="virt-table"><thead><tr><th>File</th><th className="num">Size</th><th>Changed</th></tr></thead>
      <tbody>{files.map((f) => <tr key={f.name}><td className="mono">{f.name}</td><td className="num">{fmtBytes(f.size)}</td><td>{when(f.modified)}</td></tr>)}</tbody></table>
  )
  return (
    <div className="svc-rows" style={{ padding: 12, display: 'flex', flexDirection: 'column', gap: 14 }}>
      <section className="card"><div className="sub-head"><span className="dname">Images</span><span className="spacer" /><span className="muted small mono">{dir}/images</span></div>{table(lib.images)}
        <div className="muted small">Every image downloaded, imported or published lands here as a file, so the library is whole without Incus and can be reloaded into a fresh install.</div></section>
      <section className="card"><div className="sub-head"><span className="dname">ISOs</span><span className="spacer" /><span className="muted small mono">{dir}/isos</span></div>{table(lib.isos)}
        <div className="muted small">ISO installs for VMs — upload, fetch by URL, attach, and the graphical console they need — are the next phase. Files put here now will be listed when it lands.</div></section>
    </div>
  )
 }
@@ -8,10 +8,12 @@
 // the retention in the name — rather than inventing a second convention for the two tools to
 // disagree over. An existing cron job is detected and shown, so the page tells the truth
 // about the machine whether Ham Lab or something else is doing the work.
-import { FormEvent, useCallback, useEffect, useState } from 'react'
+import { FormEvent, Fragment, useCallback, useEffect, useState } from 'react'
-import { api, ZFSExternal, ZFSInfo, ZFSOverview, ZFSPolicy, ZFSRun, ZFSSnapshot } from './api'
+import { api, ARCView, CacheDev, PoolCaching, ZFSExternal, ZFSInfo, ZFSOverview, ZFSPolicy, ZFSRun, ZFSSnapshot } from './api'
 import { fmtBytes } from './fmt'
 import { useRefreshMs } from './prefs'
 import { fetchHistories } from './sensors'
 import Sparkline from './Sparkline'
 import { askText, askYes, Btn, Loading, Meter, Spinner, toast, useEscape } from './ui'
 const ago = (t?: string | number) => {
@@ -68,18 +70,46 @@ function PoolCard({ p, onChanged }: { p: ZFSInfo['pools'][number]; onChanged: ()
  )
 }
 // A snapshot's name carries when it was taken and how long it lives — 2026-09-24_02.30.01--6w
 // — so the explorer can say when each one goes without asking the server anything.
 function snapMeta(full: string): { taken?: Date; expires?: Date } {
  const snap = full.split('@')[1] ?? ''
  const [body, ttl] = snap.split('--')
  const m = /(\d{4})-(\d{2})-(\d{2})_(\d{2})\.(\d{2})\.(\d{2})$/.exec(body)
  if (!m) return {}
  const taken = new Date(+m[1], +m[2] - 1, +m[3], +m[4], +m[5], +m[6])
  const a = /^(\d+)([mhdw])$/.exec(ttl ?? '')
  if (!a) return { taken }
  const ms = { m: 60e3, h: 3600e3, d: 86400e3, w: 7 * 86400e3 }[a[2]]!
  return { taken, expires: new Date(taken.getTime() + +a[1] * ms) }
 }
 // One dataset in the snapshot explorer: how many it has and what they cost, opening into
 // the snapshots themselves, each of which can be taken now or deleted.
 function DatasetRow({ d, onChanged }: { d: ZFSInfo['datasets'][number]; onChanged: () => void }) {
  const [snaps, setSnaps] = useState<ZFSSnapshot[] | null>(null)
  const [open, setOpen] = useState(false)
  const reload = useCallback(async () => {
    try { setSnaps(await api.zfsSnapshots(d.name)) } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err'); setSnaps([]) }
  }, [d.name])
  async function toggle() {
    setOpen(!open)
-    if (!open && !snaps) { try { setSnaps(await api.zfsSnapshots(d.name)) } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err'); setSnaps([]) } }
+    if (!open && !snaps) await reload()
  }
  async function snapshot() {
-    const def = 'hamlab-' + new Date().toISOString().slice(0, 16).replace(/[-:T]/g, '').replace(/(\d{8})(\d{4})/, '$1-$2')
+    // zfSnap's name is the machine's local time, and snapMeta reads it back as local time — so the
    // default is built from local fields, not toISOString(), which is UTC and read seven hours wrong.
    const n = new Date(), two = (x: number) => String(x).padStart(2, '0')
    const def = `${n.getFullYear()}-${two(n.getMonth() + 1)}-${two(n.getDate())}_${two(n.getHours())}.${two(n.getMinutes())}.${two(n.getSeconds())}`
    const name = await askText(`Snapshot name for ${d.name}@`, def)
    if (!name) return
-    try { await api.zfsSnapshot(d.name, name.trim()); toast(`${d.name}@${name.trim()} created`); setSnaps(null); onChanged() }
+    try { await api.zfsSnapshot(d.name, name.trim()); toast(`${d.name}@${name.trim()} created`); await reload(); onChanged() }
    catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  async function destroy(s: ZFSSnapshot) {
    const short = s.name.split('@')[1]
    if (!await askYes(`Delete ${s.name}? It is removed from this dataset and from every dataset beneath it that has a snapshot of the same name. This cannot be undone.`, 'Delete')) return
    try { await api.deleteZfsSnapshot(s.name, true); toast(`${short} deleted`); await reload(); onChanged() }
    catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') }
  }
  const depth = d.name.split('/').length - 1
@@ -87,18 +117,29 @@ function DatasetRow({ d, onChanged }: { d: ZFSInfo['datasets'][number]; onChange
    <>
      <div className="ds-row" style={{ paddingLeft: 12 + depth * 16 }}>
        <span className="ds-name" onClick={toggle}><span className="chev">{open ? '▾' : '▸'}</span>{d.name.split('/').pop()}{d.type === 'volume' && <span className="badge">zvol</span>}</span>
-        <span className="num">{fmtBytes(d.used)}</span>
+        <span className="num">{d.snapshots}</span>
        <span className="num muted">{fmtBytes(d.refer)}</span>
        <span className="num muted">{d.used_by_snapshots ? fmtBytes(d.used_by_snapshots) : '—'}</span>
-        <span className="num muted">{d.snapshots}</span>
+        <span className="num muted">{fmtBytes(d.used)}</span>
-        <span className="num muted">{d.compressratio}</span>
+        <span className="num muted">{fmtBytes(d.refer)}</span>
        <span className="ds-mount muted">{d.mount}</span>
      </div>
      {open && (
        <div className="ds-snaps">
-          <div className="svc-actions"><Btn className="btn" onClick={snapshot}>Snapshot now</Btn><span className="muted">avail {fmtBytes(d.avail)}</span></div>
+          <div className="svc-actions"><Btn className="btn" onClick={snapshot}>Snapshot now</Btn><span className="muted small">{d.snapshots} snapshot{d.snapshots === 1 ? '' : 's'} · {fmtBytes(d.avail)} free on this dataset</span></div>
          {snaps === null ? <div className="muted"><Spinner /> listing…</div> : snaps.length === 0 ? <div className="muted">No snapshots.</div> : (
-            <div className="snap-list">{snaps.slice().reverse().map((s) => <div key={s.name} className="snap"><span className="mono">{s.name.split('@')[1]}</span><span className="num muted">{new Date(s.created).toLocaleString()} · {fmtBytes(s.used)}</span></div>)}</div>
+            <div className="snap-list">{snaps.slice().reverse().map((s) => {
              const { taken, expires } = snapMeta(s.name)
              const gone = expires && expires.getTime() < Date.now()
              return (
                <div key={s.name} className="snap">
                  <span className="mono snap-name" title={s.name}>{s.name.split('@')[1]}</span>
                  <span className="muted snap-when">{taken ? taken.toLocaleString() : new Date(s.created).toLocaleString()}</span>
                  <span className={'muted snap-exp' + (gone ? ' warn' : '')}>{expires ? (gone ? 'expired' : `expires ${expires.toLocaleDateString()}`) : 'kept'}</span>
                  <span className="num muted">{fmtBytes(s.used)}</span>
                  <Btn className="btn danger snap-del" onClick={() => destroy(s)}>Delete</Btn>
                </div>
              )
            })}</div>
          )}
        </div>
      )}
@@ -126,18 +167,11 @@ function ExternalNotice({ x, mine, onAdopt }: { x: ExternalJob; mine: number; on
 type ExternalJob = ZFSExternal
 function ScheduleCard({ p, cron, next, last, onChanged, onEdit }: { p: ZFSPolicy; cron: string; next?: string; last: ZFSRun; onChanged: () => void; onEdit: () => void }) {
  const [expired, setExpired] = useState<ZFSSnapshot[] | null>(null)
  const [busy, setBusy] = useState('')
  const preview = async () => {
    setBusy('preview')
    try { setExpired((await api.zfsExpired(p.id)).expired) } catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') } finally { setBusy('') }
  }
  const run = async () => {
-    const n = expired?.length
+    if (!await askYes(`Take a snapshot of ${p.dataset} now? Snapshots past their retention are removed at the same time, as they would be on the schedule.`, 'Run now')) return
    const warn = n ? ` It will also destroy ${n} expired snapshot${n === 1 ? '' : 's'}.` : ''
    if (!await askYes(`Take a snapshot of ${p.dataset} now?${warn}`, 'Run now')) return
    setBusy('run')
-    try { const r = await api.runZfsSchedule(p.id); toast(`${r.taken}${r.pruned.length ? ` · ${r.pruned.length} expired removed` : ''}`); setExpired(null); onChanged() }
+    try { const r = await api.runZfsSchedule(p.id); toast(`${r.taken}${r.pruned.length ? ` · ${r.pruned.length} expired removed` : ''}`); onChanged() }
    catch (e) { toast(e instanceof Error ? e.message : 'failed', 'err') } finally { setBusy('') }
  }
  const remove = async () => {
@@ -165,20 +199,7 @@ function ScheduleCard({ p, cron, next, last, onChanged, onEdit }: { p: ZFSPolicy
        {p.keep_min > 0 && <div><span className="k">Floor</span><span>always keep the newest {p.keep_min}</span></div>}
      </div>
      {last.error && <div className="error" style={{ marginTop: 8 }}>{last.error}</div>}
      {expired && (
        <div className="notice" style={{ marginTop: 8 }}>
          {expired.length === 0 ? 'Nothing has expired — a run now would only take a snapshot.' : (
            <>
              <b>{expired.length}</b> snapshot{expired.length === 1 ? '' : 's'} would be destroyed by a run now:
              <div className="snap-list" style={{ marginTop: 6, maxHeight: 180, overflow: 'auto' }}>
                {expired.map((s) => <div key={s.name} className="snap"><span className="mono">{s.name.split('@')[1]}</span><span className="num muted">{new Date(s.created).toLocaleDateString()} · {fmtBytes(s.used)}</span></div>)}
              </div>
            </>
          )}
        </div>
      )}
      <div className="svc-actions">
        <Btn className="btn" busy={busy === 'preview'} onClick={preview}>What would it delete?</Btn>
        <Btn className="btn on" busy={busy === 'run'} onClick={run}>Run now</Btn>
        <button className="btn" onClick={onEdit}>Edit</button>
        <Btn className="btn" onClick={toggle}>{p.enabled ? 'Pause' : 'Resume'}</Btn>
@@ -281,7 +302,7 @@ function adopt(x: ExternalJob, fallbackDataset: string): Partial<ZFSPolicy> {
  const ds = /\b(tank[\w/-]*)\b/.exec(cmd)?.[1] ?? fallbackDataset
  return {
    ...BLANK, name: `${x.tool} (adopted)`, dataset: ds,
-    recursive: /\s-\w*r/.test(cmd) || true,
+    recursive: /\s-[a-zA-Z]*r\b/.test(cmd), // -r, or -r folded into another flag group
    every: simple ? 'daily' : 'custom',
    at: simple ? `${String(hour).padStart(2, '0')}:${String(min).padStart(2, '0')}` : '02:30',
    cron: simple ? '' : x.schedule.trim(),
@@ -289,6 +310,145 @@ function adopt(x: ExternalJob, fallbackDataset: string): Partial<ZFSPolicy> {
  }
 }
 // ---- the cache: the ARC in RAM, and what each pool has in front of its disks ----
 //
 // Two kinds of number. The live ones (size, the hit rate over the last sampling interval)
 // come from the sampler, which also keeps their history for the sparklines; the since-boot
 // counters, the tunables and the pool layout come with the same request from the storage
 // package. A tier a pool does not have is named and said to be absent, with what that means,
 // because "nothing shown" and "none fitted" are different answers.
 const hitPct = (h: number, m: number) => (h + m > 0 ? (h / (h + m)) * 100 : NaN)
 const fmtPct = (v: number) => (Number.isNaN(v) ? '—' : `${v.toFixed(1)}%`)
 const fmtN = (n: number) => n.toLocaleString()
 const cssVar = (n: string) => getComputedStyle(document.documentElement).getPropertyValue(n).trim() || '#3987e5'
 function Tier({ name, what, devs, none, note }: { name: string; what: string; devs: CacheDev[]; none: string; note?: string }) {
  return (
    <div className="arc-tier">
      <span className="k" title={what}>{name}</span>
      <span className="arc-tier-v">
        {devs.length === 0 ? <span className="none">{none}</span> : devs.map((x) => (
          <span key={x.name} className="mono" title={x.name}>
            {x.name}{x.size ? ` · ${fmtBytes(x.alloc ?? 0)} of ${fmtBytes(x.size)} used` : ''}{x.health && x.health !== 'ONLINE' ? ` · ${x.health}` : ''}
          </span>
        ))}
        {devs.length > 0 && note && <span className="muted small">{note}</span>}
      </span>
    </div>
  )
 }
 function PoolTiers({ p, v }: { p: PoolCaching; v: ARCView }) {
  const L = v.detail.lifetime, z = v.detail.zil, a = v.live
  return (
    <section className="card">
      <div className="sub-head">
        <span className="dname">{p.pool}</span>
        <span className="badge">caching tiers</span>
        <span className="spacer" />
        <span className="muted small mono" title="primarycache: what the ARC may hold for this pool · secondarycache: what an L2ARC may hold">primarycache={p.primarycache || '?'} · secondarycache={p.secondarycache || '?'}</span>
      </div>
      <div className="arc-tiers">
        <Tier name="L2ARC" what="cache devices: a second-level read cache on flash, fed with what the ARC evicts" devs={p.cache}
          none={`none — a read that misses RAM goes to the pool's disks. ${fmtBytes(L.EvictL2Eligible)} evicted from RAM since boot would have been eligible for one.`}
          note={a ? `${fmtBytes(a.l2_size)} cached (${fmtBytes(a.l2_asize)} on device) · ${fmtPct(hitPct(L.L2Hits, L.L2Misses))} hit since boot · ${fmtBytes(L.L2ReadBytes)} read, ${fmtBytes(L.L2WriteBytes)} written` : undefined} />
        <Tier name="SLOG" what="log devices: where synchronous writes are committed before they reach the pool" devs={p.log}
          none={`none — the intent log is on the pool's own vdevs: ${fmtN(z.commits)} commits, ${fmtBytes(z.NormalWrite)} written to it since boot`}
          note={`${fmtN(z.commits)} commits · ${fmtBytes(z.SlogWrite)} written to the log device since boot${z.NormalWrite ? `, ${fmtBytes(z.NormalWrite)} to the pool` : ''}`} />
        <Tier name="Special" what="special allocation class: metadata and small blocks kept on their own (faster) vdev" devs={p.special}
          none="none — metadata lives with the data on the same vdevs" />
        <Tier name="Dedup" what="deduplication: a table of every block's checksum, in the ARC and on disk" devs={p.dedup}
          none={!p.dedup_prop || p.dedup_prop === 'off' ? `off · ratio ${p.dedup_ratio || '1.00x'}` : `${p.dedup_prop} · ratio ${p.dedup_ratio} · the table lives in the ARC and on the data vdevs`} />
        {p.spare.length > 0 && <Tier name="Spares" what="hot spares: take over when a vdev member fails" devs={p.spare} none="" />}
      </div>
      <div className="muted small mono" style={{ marginTop: 8 }}>sync={p.sync || '?'} · logbias={p.logbias || '?'} · compression={p.compression || '?'} · recordsize={p.recordsize || '?'}</div>
    </section>
  )
 }
 function ARCSection() {
  const [v, setV] = useState<ARCView | null>(null)
  const [hist, setHist] = useState<{ size: number[]; hit: number[] }>({ size: [], hit: [] })
  const [err, setErr] = useState('')
  const rate = useRefreshMs()
  useEffect(() => {
    let stop = false
    const poll = async () => {
      try { const r = await api.zfsARC(); if (!stop) { setV(r); setErr('') } } catch (e) { if (!stop) setErr(e instanceof Error ? e.message : 'failed') }
      try {
        const hs = await fetchHistories('zfs.arc.size,zfs.arc.hit_pct')
        if (!stop) setHist({ size: hs.find((h) => h.id === 'zfs.arc.size')?.v ?? [], hit: hs.find((h) => h.id === 'zfs.arc.hit_pct')?.v ?? [] })
      } catch { /* the sparklines stay empty; the numbers still show */ }
    }
    poll()
    const t = setInterval(poll, Math.max(2000, rate))
    return () => { stop = true; clearInterval(t) }
  }, [rate])
  if (err && !v) return <div className="error">{err}</div>
  if (!v) return <div className="muted"><Spinner /> reading the cache…</div>
  const d = v.detail, a = v.live
  if (!d.available || !a) return <div className="notice">The ZFS module reports no ARC statistics on this machine.</div>
  const L = d.lifetime
  const ratio = a.compressed > 0 ? a.uncompressed / a.compressed : 0
  const maxSet = d.tunables.zfs_arc_max && d.tunables.zfs_arc_max !== '0'
  const atFloor = a.target <= a.min * 1.02
  const rows: [string, number, number][] = [
    ['Demand data', L.DemandDataHits, L.DemandDataMisses], ['Demand metadata', L.DemandMetaHits, L.DemandMetaMisses],
    ['Prefetch data', L.PrefetchDataHits, L.PrefetchDataMisses], ['Prefetch metadata', L.PrefetchMetaHits, L.PrefetchMetaMisses],
  ]
  return (
    <>
      <section className="card">
        <div className="sub-head">
          <span className={'dot ' + (a.no_grow ? 'warn' : 'ok')} />
          <span className="dname">ARC</span>
          <span className="badge">in RAM</span>
          <span className="spacer" />
          <span className="num">{fmtBytes(a.size)} / {fmtBytes(a.max)} · {Math.round((a.size / a.max) * 100)}%</span>
        </div>
        <Meter frac={a.size / a.max} cls="" />
        {a.no_grow && (
          <div className="notice attn" style={{ marginTop: 8 }}>
            <b>Held by memory pressure.</b> ZFS is not letting the cache grow (<span className="mono">arc_no_grow</span>): the system is short of free memory
            {atFloor ? <>, and the target sits at its floor of <b>{fmtBytes(a.min)}</b> against a ceiling of {fmtBytes(a.max)}</> : ''}.
            {' '}Whatever no longer fits is read from the disks.
          </div>
        )}
        <div className="arc-sparks">
          <div>
            <div className="muted small">Size · ceiling {fmtBytes(a.max)}</div>
            <Sparkline series={[{ name: 'size', values: hist.size, color: cssVar('--s1') }]} max={a.max} format={fmtBytes} height={48} />
          </div>
          <div>
            <div className="muted small">Hit rate · {a.hit_pct_known ? `${a.hit_pct.toFixed(1)}% now` : 'no reads yet'} · {fmtN(Math.round(a.reads_per_sec))} reads/s</div>
            <Sparkline series={[{ name: 'hit rate', values: hist.hit, color: cssVar('--ok') }]} max={100} format={(x) => `${x.toFixed(1)}%`} height={48} />
          </div>
        </div>
        <div className="props" style={{ marginTop: 8 }}>
          <div><span className="k">Target</span><span title="c: what the ARC may hold right now, between its floor (c_min) and ceiling (c_max)">{fmtBytes(a.target)} · floor {fmtBytes(a.min)} · ceiling {fmtBytes(a.max)} {maxSet ? '(zfs_arc_max)' : '(default: half of RAM)'}</span></div>
          <div><span className="k">Hit rate</span><span>{a.hit_pct_known ? `${a.hit_pct.toFixed(1)}% now` : 'idle'} · {a.lifetime_hit_pct.toFixed(1)}% since boot · {fmtN(a.hits)} hits, {fmtN(a.misses)} misses</span></div>
          <div><span className="k">Holds</span><span title="MRU: seen once · MFU: seen again · dirty: written, not yet on disk">MRU {fmtBytes(a.mru)} · MFU {fmtBytes(a.mfu)} · dirty {fmtBytes(a.anon)}</span></div>
          <div><span className="k">Of which</span><span>data {fmtBytes(a.data)} · metadata {fmtBytes(a.metadata)} · headers {fmtBytes(a.headers)}</span></div>
          <div><span className="k">Compressed</span><span>{ratio > 0 ? `${fmtBytes(a.uncompressed)} held in ${fmtBytes(a.compressed)} · ${ratio.toFixed(2)}x` : 'not compressed'}</span></div>
          <div><span className="k">Ghosts</span><span title="what was evicted recently and is still remembered, so a re-read can grow the right list">MRU {fmtBytes(a.mru_ghost)} · MFU {fmtBytes(a.mfu_ghost)}</span></div>
          <div><span className="k">Prefetch</span><span title="zfetch: the read-ahead that guesses the next blocks of a sequential read">{d.tunables.zfs_prefetch_disable === '1' ? 'disabled (zfs_prefetch_disable)' : `${fmtPct(hitPct(d.zfetch.Hits, d.zfetch.Misses))} of ${fmtN(d.zfetch.Hits + d.zfetch.Misses)} predictions since boot`}</span></div>
          <div><span className="k">dbuf cache</span><span title="the small cache of decompressed buffers in front of the ARC">{fmtBytes(d.dbuf.Size)} of {fmtBytes(d.dbuf.Target)} · {fmtPct(hitPct(d.dbuf.Hits, d.dbuf.Misses))} hits</span></div>
          <div><span className="k">Pressure</span><span title="since boot: throttles are allocations ZFS made wait for memory; prunes are metadata evictions asked of the filesystems">{fmtN(a.throttles)} throttles · {fmtN(a.prunes)} prunes · {fmtN(a.evict_skip)} evict skips</span></div>
        </div>
        <div className="arc-break">
          <span className="head">Since boot</span><span className="head num">hit rate</span><span className="head num">hits</span><span className="head num">misses</span>
          {rows.map(([k, h, m]) => <Fragment key={k}><span>{k}</span><span className="num">{fmtPct(hitPct(h, m))}</span><span className="num">{fmtN(h)}</span><span className="num">{fmtN(m)}</span></Fragment>)}
          <span><b>All reads</b></span><span className="num"><b>{fmtPct(hitPct(L.Hits, L.Misses))}</b></span><span className="num">{fmtN(L.Hits)}</span><span className="num">{fmtN(L.Misses)}</span>
        </div>
      </section>
      {d.pools.length > 0 && <div className="cards" style={{ marginTop: 8 }}>{d.pools.map((p) => <PoolTiers key={p.pool} p={p} v={v} />)}</div>}
    </>
  )
 }
 // ---- page ----
 export default function ZFS() {
@@ -317,6 +477,11 @@ export default function ZFS() {
      {zfs.pools.length === 0 && <div className="notice">ZFS is present but no pools are imported.</div>}
      <div className="cards">{zfs.pools.map((p) => <PoolCard key={p.name} p={p} onChanged={load} />)}</div>
      <div className="dash-bar" style={{ marginTop: 12 }}><h2 className="section">Cache</h2><span className="spacer" />
        <span className="muted small">the ARC in RAM, and what each pool has in front of its disks</span>
      </div>
      <ARCSection />
      <div className="dash-bar" style={{ marginTop: 12 }}><h2 className="section">Snapshot schedules</h2><span className="spacer" />
        <Btn className="btn on" onClick={() => setEditing({ ...BLANK, dataset: zfs.pools[0]?.name ?? '' })}>+ New schedule</Btn>
      </div>
@@ -326,8 +491,14 @@ export default function ZFS() {
        <ScheduleCard key={policy.id} p={policy} cron={cron} next={next} last={last ?? {}} onChanged={load} onEdit={() => setEditing(policy)} />
      ))}</div>
-      <div className="dash-bar" style={{ marginTop: 12 }}><h2 className="section">Datasets</h2><span className="spacer" /><span className="muted small">used · referenced · in snapshots · count · compression · mount</span></div>
+      <div className="dash-bar" style={{ marginTop: 12 }}><h2 className="section">Snapshots</h2><span className="spacer" />
        <span className="muted small">{zfs.datasets.reduce((a, d) => a + d.snapshots, 0)} across {zfs.datasets.length} datasets · open one to browse, take or delete</span>
      </div>
      <section className="card ds-table">
        <div className="ds-row head">
          <span>Dataset</span><span className="num">Snapshots</span><span className="num">They use</span>
          <span className="num">Dataset uses</span><span className="num">Referenced</span><span>Mount</span>
        </div>
        {zfs.datasets.map((d) => <DatasetRow key={d.name} d={d} onChanged={load} />)}
      </section>
@@ -55,7 +55,7 @@ async function req<T>(method: string, path: string, body?: unknown, local = fals
 }
 export interface UpdateStatus { current: string; latest: string; arch: string; behind: boolean; checking: boolean; installing: boolean; enabled: boolean; auto: boolean; url: string; notes?: string; published?: string; checked?: string; error?: string }
-export interface Me { addr?: string; host: string; user: string; account?: string; update_available?: string; version: string; load: string; uptime: number; reboot_needed?: boolean; fingerprint?: string; arch?: string }
+export interface Me { addr?: string; host: string; user: string; account?: string; guest?: boolean; guest_until?: string; update_available?: string; version: string; load: string; uptime: number; reboot_needed?: boolean; fingerprint?: string; arch?: string }
 export interface PushResult { name: string; ok: boolean; version?: string; error?: string }
 // What a machine is called on screen: the hostname it reports, which follows a rename within
 // a poll. The name it was added under stays its identity — the URL, its accent colour, and
@@ -72,9 +72,20 @@ export interface NetIf { iface: string; up: boolean; rx_bps: number; tx_bps: num
 export interface Disk { name: string; model?: string; size: number; read_bps: number; write_bps: number; util_pct: number; temp?: number }
 export interface FS { mount: string; type: string; size: number; used: number; remote?: boolean; disk?: string }
 export interface UPS { driver: string; model: string; state: string; source: string; on_battery: boolean; in_v: number; out_v: number; battery_pct: number; runtime_min: number; load_w: number; load_pct: number; last_event?: string }
-export interface Proc { pid: number; ppid: number; name: string; cmd?: string; user: string; cpu: number; cpu_time: number; rss: number; threads: number; state: string; nice: number; started: number }
+export interface Proc { pid: number; ppid: number; name: string; cmd?: string; user: string; cpu: number; gpu?: number; cpu_time: number; rss: number; threads: number; state: string; nice: number; started: number }
 export interface MemInfo { total: number; used: number; avail: number; free?: number; cached?: number; buffers?: number }
-export interface ProcTable { t: number; cpu: number; cores: number[]; cpu_hist: number[]; cpu_model?: string; cpu_mhz?: number; cpu_max_mhz?: number; cpu_temp?: number; core_ids?: number[]; load: [number, number, number]; mem: MemInfo; swap: MemInfo; net: NetIf[]; uptime: number; procs: Proc[]; total: number; running: number }
+// A graphics device: usage_src says what measures it ("idle" residency, the driver's own "busy", or
 // the busiest "clients" engine) and is absent when nothing does; temp_src is "gpu" for the device's own
 // sensor or "package" for the CPU die an integrated GPU shares.
 export interface GPUClient { pid: number; name: string; usage: number; engines?: Record<string, number>; mem?: number }
 export interface GPU {
  name: string; driver: string; bus?: string; integrated?: boolean
  usage: number; usage_src?: 'idle' | 'busy' | 'clients'; engines?: Record<string, number>
  temp?: number; temp_src?: 'gpu' | 'package'; power_w?: number; power_cap_w?: number; fan_rpm?: number
  mhz?: number; req_mhz?: number; max_mhz?: number; mem_used?: number; mem_total?: number
  clients: GPUClient[]; clients_ok: boolean; stalled?: boolean
 }
 export interface ProcTable { t: number; cpu: number; cores: number[]; cpu_hist: number[]; cpu_model?: string; cpu_mhz?: number; cpu_max_mhz?: number; cpu_temp?: number; core_ids?: number[]; load: [number, number, number]; mem: MemInfo; swap: MemInfo; net: NetIf[]; gpus?: GPU[]; uptime: number; procs: Proc[]; total: number; running: number }
 export interface Sample {
  t: number; interval_ms: number; cpu: number; cores: number[]; load: [number, number, number]
  mem: { total: number; used: number; avail: number }; swap: { total: number; used: number; avail: number }
@@ -109,6 +120,39 @@ export interface StorageDevice { name: string; kind: string; model?: string; ven
 export interface SmartAttr { id: number; name: string; value: number; worst: number; thresh: number; raw: string; failed?: string; warn: boolean }
 export interface SmartInfo { device: string; available: boolean; supported: boolean; standby: boolean; passed: boolean | null; temp: number; power_on_hours: number; power_cycles: number; model?: string; serial?: string; firmware?: string; attrs: SmartAttr[]; nvme?: { critical_warning: number; percentage_used: number; available_spare: number; spare_threshold: number; media_errors: number; error_log_entries: number; data_read_bytes: number; data_written_bytes: number; unsafe_shutdowns: number }; mmc?: { life_a: number; life_b: number; pre_eol: number }; self_test?: string; self_test_remaining?: number; self_test_supported: boolean; self_test_note?: string; warnings: string[]; error?: string; fetched_at: number }
 export interface DockerContainer { id: string; name: string; image: string; state: string; status: string; health: string; ports: string; stack: string; service: string; created: string; mem: number }
 export interface UpdateJob { id: string; name: string; stack: string; machine?: string; services?: string[]; cron: string; apply: boolean; enabled: boolean }
 export interface UpdateJobRow { job: UpdateJob; next?: string; last?: { at?: string; taken?: string; error?: string } }
 export interface MachineUpdates { images: ImageCheck[]; checked: number; stacks: Record<string, string[]> }
 export interface DockerHealthProbe { Start: string; End: string; ExitCode: number; Output: string }
 export interface DockerHealth { configured: boolean; state: string; status?: string; failing_streak?: number; log?: DockerHealthProbe[] }
 export interface DockerEvent { t: number; name: string; action: string; image?: string; exit_code?: string; signal?: string; health?: string }
 // ---- hardware (the Hardware tab: what the machine is made of) ----
 export interface HWSystem { model?: string; serial?: string; family?: string; kernel?: string; arch?: string; virt?: string }
 export interface HWSlot { designation: string; type?: string; usage?: string; length?: string; width?: string; bus?: string; device?: string }
 export interface HWBoard {
  maker?: string; product?: string; version?: string; serial?: string; asset_tag?: string; type?: string; location?: string
  chassis?: string; chassis_maker?: string; bios?: string; bios_date?: string; bios_revision?: string; rom_size?: string
  uefi?: boolean; slots?: HWSlot[]
 }
 export interface HWCache { level: string; size: string }
 export interface HWCPU {
  model?: string; vendor?: string; arch?: string; sockets?: number; socket_names?: string[]
  cores?: number; threads?: number; threads_per_core?: number; min_mhz?: number; max_mhz?: number
  family?: string; model_id?: string; stepping?: string; microcode?: string; virtualization?: string
  caches?: HWCache[]; flags?: string[]
 }
 export interface HWModule {
  locator: string; bank?: string; empty?: boolean; size?: number; type?: string; form_factor?: string
  speed_mts?: number; running_mts?: number; maker?: string; part?: string; serial?: string; rank?: number; voltage?: string
 }
 export interface HWMemory { total: number; installed?: number; max_capacity?: number; slots?: number; filled?: number; ecc?: string; modules?: HWModule[] }
 export interface HWDrive { name: string; kind: string; model?: string; vendor?: string; serial?: string; firmware?: string; tran?: string; size: number; rotational?: boolean; removable?: boolean }
 export interface HWGPU { model?: string; vendor?: string; slot?: string; driver?: string; node?: string; memory?: number; class?: string }
 export interface Hardware { at: string; system: HWSystem; board: HWBoard; cpu: HWCPU; memory: HWMemory; drives: HWDrive[] | null; gpus: HWGPU[] | null; notes?: string[] }
 export interface DockerLogLine { t: string; m: string; err?: boolean }
 export interface DockerInfo { available: boolean; message?: string; tool?: string; containers: DockerContainer[]; images: Record<string, unknown>[]; volumes: Record<string, unknown>[]; networks: Record<string, unknown>[] }
 export interface Stack { name: string; containers: string[]; running: number; file: string; host_file: string; files: string[]; host_files: string[] | null; images: string[] | null; via: string; remote: string; held_by: string; editable: boolean; note?: string; has_env: boolean }
 export interface StackHome { dir: string; inner: string; via: string; remote: string; held_by: string; stacks: number; default: boolean }
@@ -125,6 +169,9 @@ export interface UIPrefs { theme?: string; refreshMs?: number; accents?: Record<
 export interface User { name: string; created: string; last_seen?: string; sessions: number; self: boolean }
 export interface Session { id: string; created: string; last_seen: string; expires: string; ip?: string; agent?: string; current: boolean }
 export interface TokenInfo { name: string; created: string; scope?: string; sensors?: string[]; issued_by?: string }
 // A guest login: a full account with a clock on it, gone at `expires` or when revoked.
 export interface Guest { name: string; created: string; expires: string; issued_by?: string; sessions: number; used: boolean }
 export interface FreshGuest { name: string; password: string; expires: string; minutes: number }
 export interface SpaceProgress { files: number; dirs: number; reused: number; bytes: number; denied: number; current: string }
 export interface SpaceScan { disk: string; roots: string[]; started: string; finished?: string; running: boolean; loading?: boolean; error?: string; progress: SpaceProgress; as_root: boolean; method?: string; by_root?: Record<string, string>; partial?: boolean; files: number; dirs: number; size: number; denied: number; read: number; reused: number; partial_dirs: number; bases?: Record<string, string> }
@@ -133,16 +180,93 @@ export interface SpaceNode { name: string; dir: boolean; size: number; apparent:
 export interface SpaceTree extends SpaceNode { path: string; children?: SpaceTree[]; more_n?: number; more_size?: number; total: number }
 export interface SpaceFile { path: string; size: number; apparent: number; mtime: number }
 export interface KVM { id: string; name: string; url: string; machine: string; added: string }
 // ---- the Network tab ----
 export interface IfAddr { ip: string; prefix: number; v6?: boolean }
 export interface Iface {
  name: string; kind: string; mac?: string; mtu: number; up: boolean; carrier: boolean
  speed_mbps?: number; duplex?: string; driver?: string; signal_dbm?: number
  addrs: IfAddr[]; rx_total: number; tx_total: number; default?: boolean
 }
 export interface NetReport { at: string; host?: string; ifaces: Iface[]; gateway?: string; gateway_via?: string; dns?: string[]; search?: string[] }
 export interface Outside { at: string; ip?: string; colo?: string; country?: string; error?: string }
 export interface SpeedResult {
  at: string; down_mbps: number; up_mbps: number; latency_ms: number; jitter_ms: number
  down_bytes: number; up_bytes: number; took_sec: number; colo?: string; ip?: string; source?: string; note?: string; error?: string
 }
 // A test in progress, or the last one to have finished. `done` is 0–1 across the whole test
 // and `mbps` is the live figure for the phase under way.
 export interface SpeedRun { started: string; running: boolean; phase: string; done: number; mbps: number; bytes: number; result?: SpeedResult }
 export interface NetworkInfo { report: NetReport; outside: Outside; run?: SpeedRun; history: SpeedResult[]; method: string }
 // ---- the VPN tab ----
 // A peer is a device that dials in (on a server tunnel) or the way out (on a client one).
 // `online` is a handshake in the last few minutes, not a connection: WireGuard has none.
 export interface VPNPeer {
  name?: string; public_key: string; endpoint?: string; allowed_ips: string[]
  handshake?: string; online: boolean; rx: number; tx: number; keepalive?: number; psk?: boolean
  // The configuration section holding it, and the id anything that edits it uses. Absent when
  // the peer is not in a configuration Ham Lab can edit.
  section?: string
 }
 export interface VPNTunnel {
  name: string; public_key: string; kind: 'server' | 'client' | ''; listen_port?: number
  fwmark?: string; addrs: string[]; mtu?: number; peers: VPNPeer[]; exit?: string
  // Whether devices on this tunnel can be added and removed from here.
  managed?: boolean
 }
 export interface NewVPNPeer { name: string; address?: string; endpoint: string; dns?: string; split?: boolean }
 // A device just made, and the one chance to read its profile: the private key in it is
 // generated for that call and kept nowhere.
 export interface CreatedVPNPeer { section: string; name: string; public_key: string; address: string; config: string }
 export interface VPNReport { at: string; tunnels: VPNTunnel[]; tool: boolean; root: boolean; note?: string }
 export interface ZFSPolicy { id: string; name: string; dataset: string; recursive: boolean; every: 'hourly' | 'daily' | 'weekly' | 'monthly' | 'custom'; at: string; weekday: number; day: number; cron?: string; keep_for: string; keep_min: number; prefix?: string; enabled: boolean }
 export interface ZFSRun { at?: string; taken?: string; pruned?: string[]; error?: string }
 export interface ZFSExternal { tool: string; schedule: string; command: string; user: string; source: string; id: string; next?: number; last?: number }
 export interface ZFSOverview { zfs: ZFSInfo; policies: { policy: ZFSPolicy; cron: string; next?: string; last: ZFSRun }[]; external?: ZFSExternal | null }
 // The ARC as the sampler sees it (sizes in bytes; hit_pct over the last sampling interval).
 export interface ARCLive {
  size: number; target: number; min: number; max: number
  mru: number; mfu: number; anon: number; data: number; metadata: number; meta_used: number; headers: number
  compressed: number; uncompressed: number; mru_ghost: number; mfu_ghost: number; l2_size: number; l2_asize: number
  reads_per_sec: number; misses_per_sec: number; hit_pct: number; hit_pct_known: boolean; l2_hit_pct: number
  hits: number; misses: number; lifetime_hit_pct: number; no_grow: boolean; throttles: number; prunes: number; evict_skip: number
 }
 export interface CacheDev { name: string; size?: number; alloc?: number; health?: string }
 export interface PoolCaching {
  pool: string; cache: CacheDev[]; log: CacheDev[]; special: CacheDev[]; dedup: CacheDev[]; spare: CacheDev[]
  primarycache: string; secondarycache: string; sync: string; logbias: string; dedup_prop: string; compression: string; recordsize: string; dedup_ratio: string
 }
 export interface ARCDetail {
  available: boolean; mem_total: number; tunables: Record<string, string>
  lifetime: { Hits: number; Misses: number; DemandDataHits: number; DemandDataMisses: number; DemandMetaHits: number; DemandMetaMisses: number; PrefetchDataHits: number; PrefetchDataMisses: number; PrefetchMetaHits: number; PrefetchMetaMisses: number; L2Hits: number; L2Misses: number; L2ReadBytes: number; L2WriteBytes: number; L2Feeds: number; L2RebuildSize: number; EvictL2Eligible: number; Deleted: number; EvictSkip: number; MutexMiss: number; Prune: number; Throttle: number }
  zfetch: { Hits: number; Misses: number; IOIssued: number }
  zil: { commits: number; NormalBytes: number; NormalWrite: number; SlogBytes: number; SlogWrite: number }
  dbuf: { Size: number; Target: number; Hits: number; Misses: number }
  pools: PoolCaching[]
 }
 export interface ARCView { detail: ARCDetail; live: ARCLive | null }
 export interface FileEntry { name: string; type: 'dir' | 'file' | 'link' | 'other'; size: number; mode: string; perm: string; mtime: number; user: string; group: string; target?: string; broken?: boolean }
 export interface FileInfo extends FileEntry { path: string; text: boolean }
 const q = (path: string, sudo: boolean, extra = '') => `path=${encodeURIComponent(path)}${sudo ? '&sudo=1' : ''}${extra}`
 // ---- Virtualization: Incus, per machine ----
 export interface VirtLayout { driver: string; pool_source: string; pool_size_gib?: number; bridge: string; subnet: string; docker_rules: boolean; arc_max_gib?: number; attempted: string; hamlab_version?: string }
 export interface VirtStatus { state: 'ready' | 'repair' | 'installable' | 'unavailable'; reason?: string; kvm: boolean; arch: string; version?: string; pm?: string; drivers: string[]; docker: boolean; zfs_host: boolean; arc_max_gib?: number; mem_total: number; mem_avail: number; disk_free: number; dir: string; subnet: string; pool_size_gib?: number; layout?: VirtLayout; user: string; root: boolean; unit?: string; btrfs_tools: boolean }
 export interface VirtSnapshot { name: string; created: string; expires?: string; size?: number }
 export interface VirtInstance { name: string; type: 'container' | 'virtual-machine'; status: string; arch: string; description?: string; image: string; os?: string; release?: string; cpu?: string; memory?: string; created: string; last_used: string; ipv4: string[]; snapshots: VirtSnapshot[]; profiles: string[]; ephemeral: boolean; cpu_pct: number; mem_usage: number; mem_peak: number; disk_usage: number; disk_total: number; pid: number; processes: number; config: Record<string, string>; devices: Record<string, Record<string, string>>; schedule?: string; expiry?: string }
 export interface VirtImage { fingerprint: string; aliases: string[]; arch: string; type: string; size: number; created: string; last_used: string; description: string; os?: string; release?: string; variant?: string; cached: boolean; source?: string; public: boolean }
 export interface VirtPool { name: string; driver: string; source: string; used: number; total: number; used_by: number }
 export interface VirtNetwork { name: string; type: string; managed: boolean; address?: string; nat: boolean; used_by: number }
 export interface LibraryFile { name: string; size: number; modified: string }
 export interface VirtOverview { status: VirtStatus; error?: string; instances?: VirtInstance[]; images?: VirtImage[]; pools?: VirtPool[]; networks?: VirtNetwork[]; library: { images: LibraryFile[]; isos: LibraryFile[] } }
 export interface VirtInstallOpts { driver: string; pool_size_gib: number; subnet: string; docker_rules: boolean; arc_max_gib: number }
 export interface VirtCreate { name: string; image: string; vm: boolean; cpu: string; memory: string; disk: string; network: string; cloud_init: string; profiles: string[]; description: string; start: boolean }
 export const api = {
  procs: () => req<ProcTable>('GET', '/api/procs'),
  space: () => req<{ targets: SpaceTarget[] }>('GET', '/api/space'),
@@ -184,10 +308,14 @@ export const api = {
  storage: () => req<{ devices: StorageDevice[]; zfs: ZFSInfo; smartctl: boolean }>('GET', '/api/storage'),
  smart: (dev: string, force = false) => req<SmartInfo>('GET', `/api/storage/smart/${dev}${force ? '?force=1' : ''}`),
  smartTest: (dev: string, kind: 'short' | 'long') => req<void>('POST', `/api/storage/smart/${dev}/test/${kind}`),
  // Always the hub's: it holds every machine's KVM, so the tab works when the machine does not.
  kvms: () => req<KVM[]>('GET', '/api/kvms', undefined, true),
  addKVM: (name: string, url: string, machine: string) => req<KVM>('POST', '/api/kvms', { name, url, machine }, true),
  removeKVM: (id: string) => req<void>('DELETE', `/api/kvms/${encodeURIComponent(id)}`, undefined, true),
  zfsOverview: () => req<ZFSOverview>('GET', '/api/zfs'),
  zfsARC: () => req<ARCView>('GET', '/api/zfs/arc'),
  saveZfsSchedule: (p: Partial<ZFSPolicy>) => req<ZFSPolicy>('PUT', '/api/zfs/schedules', p),
  deleteZfsSchedule: (id: string) => req<void>('DELETE', `/api/zfs/schedules/${encodeURIComponent(id)}`),
  zfsExpired: (id: string) => req<{ expired: ZFSSnapshot[] }>('GET', `/api/zfs/schedules/${encodeURIComponent(id)}/expired`),
  runZfsSchedule: (id: string) => req<{ taken: string; pruned: string[] }>('POST', `/api/zfs/schedules/${encodeURIComponent(id)}/run`),
  deleteZfsSnapshot: (name: string, recursive: boolean) => req<void>('POST', '/api/zfs/snapshots/delete', { name, confirm: name, recursive }),
  zfsSnapshots: (dataset: string) => req<ZFSSnapshot[]>('GET', `/api/storage/zfs/snapshots?dataset=${encodeURIComponent(dataset)}`),
@@ -223,9 +351,13 @@ export const api = {
  addUser: (name: string, password: string) => req<{ name: string }>('POST', '/api/users', { name, password }, true),
  removeUser: (name: string, current: string) => req<void>('DELETE', `/api/users/${encodeURIComponent(name)}`, { current }, true),
  resetPassword: (name: string, current: string, password: string) => req<void>('POST', `/api/users/${encodeURIComponent(name)}/password`, { current, password }, true),
  guests: () => req<Guest[]>('GET', '/api/guests', undefined, true),
  createGuest: (minutes: number) => req<FreshGuest>('POST', '/api/guests', { minutes }, true),
  revokeGuest: (name: string) => req<void>('DELETE', `/api/guests/${encodeURIComponent(name)}`, undefined, true),
  sessions: () => req<Session[]>('GET', '/api/sessions', undefined, true),
  revokeSession: (id: string) => req<void>('DELETE', `/api/sessions/${encodeURIComponent(id)}`, undefined, true),
  revokeOtherSessions: () => req<{ revoked: number }>('POST', '/api/sessions/revoke-others', undefined, true),
  revokeIdleSessions: (hours = 24) => req<{ revoked: number }>('POST', `/api/sessions/revoke-idle?hours=${hours}`, undefined, true),
  logout: () => req<void>('POST', '/api/logout', undefined, true),
  meHere: () => req<Me>('GET', '/api/me', undefined, true),
  me: () => req<Me>('GET', '/api/me'),
@@ -239,10 +371,30 @@ export const api = {
  installUpdate: (version: string, force = false) => req<{ version: string; previous: string }>('POST', `/api/update/install${force ? '?force=1' : ''}`, { version }),
  tokens: () => req<TokenInfo[]>('GET', '/api/tokens', undefined, true),
  createReadToken: (name: string, sensors: string[] = []) => req<{ token: string; name: string }>('POST', '/api/tokens/read', { name, sensors }, true),
  hardware: (refresh = false) => req<Hardware>('GET', `/api/hardware${refresh ? '?refresh=1' : ''}`),
  network: () => req<NetworkInfo>('GET', '/api/network'),
  vpn: () => req<VPNReport>('GET', '/api/vpn'),
  addVPNPeer: (iface: string, p: NewVPNPeer) => req<CreatedVPNPeer>('POST', `/api/vpn/${encodeURIComponent(iface)}/peers`, p),
  editVPNPeer: (iface: string, section: string, p: { name?: string; address?: string }) =>
    req<void>('PATCH', `/api/vpn/${encodeURIComponent(iface)}/peers/${encodeURIComponent(section)}`, p),
  removeVPNPeer: (iface: string, section: string, confirm: string) =>
    req<void>('DELETE', `/api/vpn/${encodeURIComponent(iface)}/peers/${encodeURIComponent(section)}`, { confirm }),
  speedtest: () => req<SpeedRun>('GET', '/api/network/speedtest'),
  startSpeedtest: () => req<SpeedRun>('POST', '/api/network/speedtest'),
  cancelSpeedtest: () => req<void>('POST', '/api/network/speedtest/cancel'),
  docker: () => req<DockerInfo>('GET', '/api/docker'),
  dockerAction: (id: string, action: 'start' | 'stop' | 'restart' | 'remove') => req<void>('POST', `/api/docker/containers/${encodeURIComponent(id)}/${action}`),
-  dockerLogs: (id: string, since = '') => req<{ lines: { t: string; m: string }[] }>('GET', `/api/docker/containers/${encodeURIComponent(id)}/logs${since ? `?since=${encodeURIComponent(since)}` : '?n=300'}`),
+  dockerLogs: (id: string, since = '') => req<{ lines: DockerLogLine[] }>('GET', `/api/docker/containers/${encodeURIComponent(id)}/logs${since ? `?since=${encodeURIComponent(since)}` : '?n=300'}`),
  dockerInspect: (id: string) => req<Record<string, unknown>>('GET', `/api/docker/containers/${encodeURIComponent(id)}/inspect`),
  machineUpdates: (refresh = false) => req<MachineUpdates>('GET', `/api/docker/updates${refresh ? '?refresh=1' : ''}`),
  updateJobs: () => req<{ jobs: UpdateJobRow[]; file: string }>('GET', '/api/docker/schedules'),
  saveUpdateJob: (j: Partial<UpdateJob>) => req<UpdateJob>('PUT', '/api/docker/schedules', j),
  deleteUpdateJob: (id: string) => req<void>('DELETE', `/api/docker/schedules/${encodeURIComponent(id)}`),
  runUpdateJob: (id: string) => req<{ output: string }>('POST', `/api/docker/schedules/${encodeURIComponent(id)}/run`),
  dockerHealth: (id: string) => req<DockerHealth>('GET', `/api/docker/containers/${encodeURIComponent(id)}/health`),
  dockerEvents: (id: string, since = '7d') => req<{ events: DockerEvent[]; since: string; kept: number; capacity: number }>('GET', `/api/docker/containers/${encodeURIComponent(id)}/events?since=${since}`),
  dockerExec: (id: string) => req<{ id: string; name: string }>('POST', `/api/docker/containers/${encodeURIComponent(id)}/exec`, {}),
  dockerLimits: (id: string, p: { memory?: string; cpus?: string; restart?: string }) => req<void>('POST', `/api/docker/containers/${encodeURIComponent(id)}/limits`, p),
  stacks: () => req<{ stacks: Stack[] }>('GET', '/api/docker/stacks'),
  stackHomes: () => req<{ homes: StackHome[] }>('GET', '/api/docker/stacks/homes'),
  createStack: (p: { name: string; text: string; env: string; home: string; env_from?: string }) => req<{ stack: Stack }>('POST', '/api/docker/stacks', p),
@@ -252,11 +404,13 @@ export const api = {
  pruneImages: (all = false) => req<{ result: string }>('POST', '/api/docker/images/prune', { all }),
  removeVolume: (name: string) => req<void>('POST', '/api/docker/volumes/remove', { name, confirm: name }),
  pruneVolumes: () => req<{ result: string }>('POST', '/api/docker/volumes/prune'),
  removeNetwork: (name: string) => req<void>('POST', '/api/docker/networks/remove', { name }),
  pruneNetworks: () => req<{ result: string }>('POST', '/api/docker/networks/prune'),
  stackDiff: (name: string) => req<{ files: StackDiff[] }>('GET', `/api/docker/stacks/${encodeURIComponent(name)}/diff`),
  stackUpdates: (name: string, refresh = false) => req<{ images: ImageCheck[]; checked: number }>('GET', `/api/docker/stacks/${encodeURIComponent(name)}/updates${refresh ? '?refresh=1' : ''}`),
  stackFile: (name: string, file: number | 'env' = 0) => req<{ stack: Stack; text: string; sum: string }>('GET', `/api/docker/stacks/${encodeURIComponent(name)}?file=${file}`),
  saveStack: (name: string, file: number | 'env', text: string, sum: string) => req<void>('PUT', `/api/docker/stacks/${encodeURIComponent(name)}?file=${file}`, { text, sum }),
-  stackAction: (name: string, action: 'deploy' | 'update' | 'restart' | 'stop' | 'down') => req<TaskInfo>('POST', `/api/docker/stacks/${encodeURIComponent(name)}/${action}`),
+  stackAction: (name: string, action: 'deploy' | 'update' | 'recreate' | 'restart' | 'stop' | 'down') => req<TaskInfo>('POST', `/api/docker/stacks/${encodeURIComponent(name)}/${action}`),
  alerts: () => req<AlertsInfo>('GET', '/api/alerts', undefined, true),
  saveAlerts: (p: { rules?: AlertRule[]; channels?: AlertChannel[] }) => req<void>('PUT', '/api/alerts', p, true),
  testAlerts: () => req<void>('POST', '/api/alerts/test', undefined, true),
@@ -268,6 +422,21 @@ export const api = {
  createTerminal: (name?: string) => req<{ id: string; name: string }>('POST', '/api/terminals', { name }),
  renameTerminal: (id: string, name: string) => req<void>('PATCH', `/api/terminals/${id}`, { name }),
  deleteTerminal: (id: string) => req<void>('DELETE', `/api/terminals/${id}`),
  virt: () => req<VirtOverview>('GET', '/api/virt'),
  virtInstall: (o: VirtInstallOpts) => req<TaskInfo>('POST', '/api/virt/install', o),
  virtRemote: (type: string, refresh = false) => req<{ images: VirtImage[] }>('GET', `/api/virt/images/remote?type=${type}${refresh ? '&refresh=1' : ''}`),
  virtImageDownload: (alias: string) => req<TaskInfo>('POST', '/api/virt/images/download', { alias }),
  virtImageImport: (path: string, alias: string, type: string) => req<TaskInfo>('POST', '/api/virt/images/import', { path, alias, type }),
  virtImagePublish: (instance: string, alias: string) => req<TaskInfo>('POST', '/api/virt/images/publish', { instance, alias }),
  virtImageDelete: (ref: string) => req<void>('POST', '/api/virt/images/delete', { ref }),
  virtCreate: (c: VirtCreate) => req<TaskInfo>('POST', '/api/virt/instances', c),
  virtAction: (name: string, action: 'start' | 'stop' | 'restart' | 'kill') => req<void>('POST', `/api/virt/instances/${encodeURIComponent(name)}/${action}`),
  virtDelete: (name: string, force: boolean) => req<void>('POST', `/api/virt/instances/${encodeURIComponent(name)}/delete`, { confirm: name, force }),
  virtTerminal: (name: string, kind: 'console' | 'exec') => req<{ id: string; name: string }>('POST', `/api/virt/instances/${encodeURIComponent(name)}/${kind}`),
  virtRename: (name: string, to: string) => req<void>('POST', `/api/virt/instances/${encodeURIComponent(name)}/rename`, { name: to }),
  virtConfig: (name: string, cfg: Record<string, string>) => req<void>('PUT', `/api/virt/instances/${encodeURIComponent(name)}/config`, cfg),
  virtSnapshot: (name: string, snap: string) => req<{ name: string }>('POST', `/api/virt/instances/${encodeURIComponent(name)}/snapshots`, { name: snap }),
  virtSnapshotAction: (name: string, snap: string, action: 'restore' | 'delete') => req<void>('POST', `/api/virt/instances/${encodeURIComponent(name)}/snapshots/${encodeURIComponent(snap)}/${action}`),
 }
 export function wsURLPath(path: string): string {
@@ -0,0 +1,134 @@
 // A cron expression in English. Not a parser for running jobs — internal/cron does that,
 // and computes the next run — just a sentence for the person reading the schedule, who
 // should not have to decode "15 3 * * 1" in their head.
 //
 // Anything it can't say plainly it declines to describe, and the caller shows the
 // expression instead. A confident wrong sentence is worse than five fields.
 const DAYS = ['Sunday', 'Monday', 'Tuesday', 'Wednesday', 'Thursday', 'Friday', 'Saturday']
 const MONTHS = ['January', 'February', 'March', 'April', 'May', 'June', 'July', 'August', 'September', 'October', 'November', 'December']
 const DAY_ALIAS: Record<string, number> = { sun: 0, mon: 1, tue: 2, wed: 3, thu: 4, fri: 5, sat: 6 }
 const MONTH_ALIAS: Record<string, number> = { jan: 1, feb: 2, mar: 3, apr: 4, may: 5, jun: 6, jul: 7, aug: 8, sep: 9, oct: 10, nov: 11, dec: 12 }
 // "@daily" and friends, which crontab(5) accepts in place of the five fields.
 const ALIASES: Record<string, string> = {
  '@yearly': 'once a year, on 1 January at midnight',
  '@annually': 'once a year, on 1 January at midnight',
  '@monthly': 'monthly, on the 1st at midnight',
  '@weekly': 'weekly, on Sunday at midnight',
  '@daily': 'daily at midnight',
  '@midnight': 'daily at midnight',
  '@hourly': 'hourly, on the hour',
  '@reboot': 'at every boot',
 }
 const clock = (h: number, m: number) => {
  if (m === 0 && h === 0) return 'midnight'
  if (m === 0 && h === 12) return 'noon'
  const ampm = h < 12 ? 'am' : 'pm'
  const h12 = h % 12 === 0 ? 12 : h % 12
  return m === 0 ? `${h12}${ampm}` : `${h12}:${String(m).padStart(2, '0')}${ampm}`
 }
 // Monday-to-Friday and Saturday-Sunday are worth naming; the rest get listed.
 function dayNames(d: number[]): string {
  const set = d.join(',')
  if (set === '1,2,3,4,5') return 'weekdays'
  if (set === '0,6') return 'weekends'
  return list(d.map((x) => DAYS[x]))
 }
 const list = (parts: string[]) =>
  parts.length <= 1 ? (parts[0] ?? '') : `${parts.slice(0, -1).join(', ')} and ${parts[parts.length - 1]}`
 const ordinal = (n: number) => {
  const s = ['th', 'st', 'nd', 'rd'][(n % 100 - 20) % 10] ?? ['th', 'st', 'nd', 'rd'][n % 100] ?? 'th'
  return `${n}${s}`
 }
 // values of one field, or null when it is a form this doesn't describe
 function values(field: string, min: number, max: number, alias: Record<string, number> = {}): number[] | null {
  const out = new Set<number>()
  for (const part of field.split(',')) {
    const [range, stepText] = part.split('/')
    const step = stepText === undefined ? 1 : Number(stepText)
    if (!Number.isInteger(step) || step < 1) return null
    let lo: number, hi: number
    if (range === '*') {
      lo = min; hi = max
    } else {
      const ends = range.split('-').map((v) => {
        const n = alias[v.toLowerCase()] ?? Number(v)
        return Number.isInteger(n) ? n : NaN
      })
      if (ends.some(Number.isNaN) || ends.length > 2) return null
      lo = ends[0]
      hi = ends.length === 2 ? ends[1] : (stepText === undefined ? ends[0] : max)
    }
    if (lo < min || hi > max || lo > hi) return null
    for (let v = lo; v <= hi; v += step) out.add(v)
  }
  return [...out].sort((a, b) => a - b)
 }
 const everyValue = (v: number[], min: number, max: number) => v.length === max - min + 1
 // stepOf reports the interval when a field is a plain "every n" over its whole range.
 function stepOf(v: number[], min: number, max: number): number | null {
  if (v.length < 2 || v[0] !== min) return null
  const step = v[1] - v[0]
  for (let i = 1; i < v.length; i++) if (v[i] - v[i - 1] !== step) return null
  return v[v.length - 1] + step > max ? step : null
 }
 /** describeCron turns "15 3 * * *" into "daily at 3:15am", or returns "" if it can't. */
 export function describeCron(expr: string): string {
  const raw = (expr ?? '').trim()
  if (!raw) return ''
  if (raw.startsWith('@')) return ALIASES[raw.toLowerCase()] ?? ''
  const f = raw.split(/\s+/)
  if (f.length !== 5) return '' // 6-field (seconds) and Quartz forms aren't ours to guess at
  const [minF, hourF, domF, monF, dowF] = f
  const mins = values(minF, 0, 59)
  const hours = values(hourF, 0, 23)
  const doms = values(domF, 1, 31)
  const mons = values(monF, 1, 12, MONTH_ALIAS)
  const dows = values(dowF.replace(/7/g, '0'), 0, 6, DAY_ALIAS)
  if (!mins || !hours || !doms || !mons || !dows) return ''
  // --- when in the day ---
  let time: string
  const everyMin = everyValue(mins, 0, 59)
  const everyHour = everyValue(hours, 0, 23)
  const minStep = stepOf(mins, 0, 59)
  const hourStep = stepOf(hours, 0, 23)
  if (everyMin && everyHour) time = 'every minute'
  else if (everyMin) time = `every minute of ${list(hours.map((h) => clock(h, 0).replace(/(am|pm)/, '$1')))}`
  else if (minStep && everyHour) time = `every ${minStep} minutes`
  else if (minStep) time = `every ${minStep} minutes, ${hours.length === 1 ? `during the ${clock(hours[0], 0)} hour` : `during ${list(hours.map((h) => clock(h, 0)))}`}`
  else if (everyHour && mins.length === 1) time = mins[0] === 0 ? 'hourly, on the hour' : `hourly, at ${mins[0]} past`
  else if (hourStep && mins.length === 1) time = `every ${hourStep} hours` + (mins[0] === 0 ? ', on the hour' : `, at ${mins[0]} past`)
  else if (mins.length === 1 && hours.length >= 1) time = `at ${list(hours.map((h) => clock(h, mins[0])))}`
  else {
    const times: string[] = []
    for (const h of hours) for (const m of mins) times.push(clock(h, m))
    if (times.length > 6) return '' // a wall of times is not a translation
    time = `at ${list(times)}`
  }
  // --- which days ---
  const allDom = everyValue(doms, 1, 31)
  const allDow = everyValue(dows, 0, 6)
  const allMon = everyValue(mons, 1, 12)
  let days = ''
  if (allDom && allDow) days = ''
  else if (allDom) days = `on ${dayNames(dows)}`
  else if (allDow) days = `on the ${list(doms.map(ordinal))}`
  // crontab(5): with both set, the job runs when *either* matches — worth saying out loud
  else days = `on the ${list(doms.map(ordinal))}, and on ${dayNames(dows)}`
  const months = allMon ? '' : `in ${list(mons.map((m) => MONTHS[m - 1]))}`
  const daily = !days && !months
  const head = daily ? (time.startsWith('every') || time.startsWith('hourly') ? '' : 'daily ') : ''
  return [head + time, days, months].filter(Boolean).join(' ').replace(/\s+/g, ' ').trim()
 }
@@ -0,0 +1,24 @@
 // A guest login's clock. That is the whole difference between a guest and an account: it
 // can do everything, until it can't, so what the console has to show is how long is left.
 import { useEffect, useState } from 'react'
 // fmtLeft says how long is left, in the roundest terms that are still honest.
 export function fmtLeft(ms: number): string {
  const min = Math.floor(ms / 60000)
  if (min >= 120) return `${Math.floor(min / 60)} h ${min % 60} min`
  if (min >= 1) return `${min} min`
  return 'under a minute'
 }
 // useTimeLeft is the milliseconds until `until`, re-rendering the caller every 15 seconds so
 // the number keeps up. 0 when there is no expiry, which is every normal account.
 export function useTimeLeft(until?: string): number {
  const [, tick] = useState(0)
  const active = !!until
  useEffect(() => {
    if (!active) return
    const t = setInterval(() => tick((n) => n + 1), 15000)
    return () => clearInterval(t)
  }, [active])
  return until ? Math.max(0, new Date(until).getTime() - Date.now()) : 0
 }
@@ -8,7 +8,7 @@ export type Values = Record<string, number | boolean | string>
 export type Catalog = Map<string, Sensor>
 // Everything the modules draw, as one stream query.
-export const MONITOR_SHAPE = ['cpu.*', 'load.*', 'uptime', 'hamlab.*', 'mem.*', 'swap.*', 'temp.*', 'fan.*', 'net.*', 'disk.*', 'fs.*', 'ups.*']
+export const MONITOR_SHAPE = ['cpu.*', 'load.*', 'uptime', 'hamlab.*', 'mem.*', 'swap.*', 'temp.*', 'fan.*', 'net.*', 'disk.*', 'fs.*', 'ups.*', 'gpu.*']
 // The stream answers a list with a list; the modules want one flat map.
 export function mergeValues(parts: unknown[]): Values {
  const out: Values = {}
@@ -20,7 +20,14 @@ button { font: inherit; color: inherit; background: none; border: 0; cursor: poi
 .login input { padding: 10px 12px; background: var(--bg2); border: 1px solid var(--border); border-radius: 6px; color: var(--fg); font-size: 16px; }
 .login button, .primary { padding: 10px 12px; background: var(--accent); color: var(--btn-fg); border-radius: 6px; font-weight: 600; }
 .login button:disabled { opacity: .5; }
 /* the password field and its show/hide eye, which sits inside the field's right edge */
 .login .pwfield { position: relative; display: flex; }
 .login .pwfield input { flex: 1; min-width: 0; padding-right: 44px; }
 .login button.pweye { position: absolute; right: 1px; top: 1px; bottom: 1px; width: 42px; padding: 0; background: none; color: var(--muted); font-weight: 400; border-radius: 0 6px 6px 0; }
 .login button.pweye:hover, .login button.pweye[aria-pressed="true"] { color: var(--fg); }
 .login button.pweye:disabled { opacity: .4; }
 .error { color: var(--bad); font-size: 13px; }
 .svc-list > .error, .cron-list > .error { margin: 8px 12px; }
 /* layout */
 .app { height: 100%; display: flex; flex-direction: column; padding-top: env(safe-area-inset-top); }
@@ -94,6 +101,10 @@ button { font: inherit; color: inherit; background: none; border: 0; cursor: poi
 .tile .sub { font-size: 12px; color: var(--muted); margin-top: 2px; font-variant-numeric: tabular-nums; }
 .tile .spark { margin-top: 8px; }
 .card .empty { color: var(--muted); font-size: 13px; }
 /* the out-of-date list: image, what the registry said, which stacks, and the button */
 .upd-list .row { grid-template-columns: minmax(130px, 1.3fr) minmax(110px, 1.5fr) auto auto; }
 .upd-list .row .btn { padding: 2px 10px; }
@media (max-width: 560px) { .upd-list .row { grid-template-columns: 1fr auto; } .upd-list .row > :nth-child(2) { grid-column: 1 / -1; } }
 .row { display: grid; grid-template-columns: minmax(90px, 1fr) 2fr auto; gap: 10px; align-items: center; padding: 4px 0; font-size: 13px; min-width: 0; }
 .row .k { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; } .row .k small { color: var(--muted); margin-left: 6px; }
 .num { font-variant-numeric: tabular-nums; font-family: ui-monospace, monospace; font-size: 12px; white-space: nowrap; }
@@ -163,11 +174,28 @@ button { font: inherit; color: inherit; background: none; border: 0; cursor: poi
  .svc:not(.has-sel) .svc-list, .cron:not(.has-sel) .cron-list { width: auto; flex: 1 1 auto; border-right: 0; }
  .svc:not(.has-sel) .svc-empty, .cron:not(.has-sel) .svc-empty { display: none; }
 }
 /* A tab that is a page rather than a list-and-detail — Docker's Updates — stacks: the tab
   strip on top, the page below at full width. Otherwise the empty list column takes half
   the window and the page is squeezed into the other half, which looks like a phone. */
 .svc.wide { flex-direction: column; }
 /* :not(.has-sel) to outrank the wide-window rule above, which grows the list to fill */
 .svc.wide:not(.has-sel) .svc-list { width: auto; flex: 0 0 auto; border-right: 0; border-bottom: 1px solid var(--border); }
 .svc.wide .svc-rows { display: none; }
 .svc.wide > .svc-detail { flex: 1 1 auto; min-height: 0; }
 .svc-head { display: flex; align-items: center; gap: 10px; } .svc-head h2 { margin: 0; font-size: 18px; font-weight: 600; word-break: break-all; }
 .svc-actions { display: flex; flex-wrap: wrap; gap: 6px; }
 .props { display: grid; grid-template-columns: repeat(auto-fill, minmax(260px, 1fr)); gap: 4px 16px; font-size: 13px; }
 .props > div { display: flex; gap: 8px; min-width: 0; } .props .k { color: var(--muted); width: 80px; flex-shrink: 0; }
 .props span:not(.k) { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; } .mono { font-family: ui-monospace, monospace; font-size: 12px; }
 /* One property per row, label above its value. For panes like Inspect, whose values are
   long paths and lists — side by side they were half truncated and read as columns. */
 .props.stacked { grid-template-columns: 1fr; gap: 10px 0; }
 .props.stacked > div { flex-direction: column; gap: 2px; }
 .props.stacked .k { width: auto; font-size: 11px; letter-spacing: .04em; text-transform: uppercase; }
 .props.stacked span:not(.k) { overflow: visible; white-space: normal; overflow-wrap: anywhere; }
 .props.stacked .k { text-transform: none; } .props.stacked > div > .k { text-transform: uppercase; }
 .props.stacked .k .muted { text-transform: none; letter-spacing: 0; }
 .props.stacked span:not(.k) > div + div { margin-top: 3px; }
 .tabs { display: flex; gap: 2px; border-bottom: 1px solid var(--border); }
 .tabs button { padding: 6px 12px; color: var(--muted); border-bottom: 2px solid transparent; margin-bottom: -1px; }
 .tabs button.on { color: var(--fg); border-bottom-color: var(--accent); }
@@ -177,6 +205,8 @@ button { font: inherit; color: inherit; background: none; border: 0; cursor: poi
 .log-lines { flex: 1; overflow: auto; padding: 6px 8px; font: 12px/1.45 ui-monospace, SFMono-Regular, Menlo, monospace; }
 .log { display: flex; gap: 8px; white-space: pre-wrap; word-break: break-word; }
 .log .lt { color: var(--muted); flex-shrink: 0; } .log .li { color: var(--accent2); flex-shrink: 0; }
 /* a phone has no room for three columns: the stamp and the unit go on one line, the message underneath */
@media (max-width: 560px) { .log { flex-wrap: wrap; gap: 0 8px; padding: 2px 0; border-bottom: 1px solid var(--border); } .log .lm { flex-basis: 100%; } .log .lt { font-size: 11px; } }
 .log.err .lm { color: var(--bad); } .log.warn .lm { color: var(--s2); } .log.debug .lm { color: var(--muted); }
@media (max-width: 800px) {
  .svc-list { width: 100%; border-right: 0; } .has-sel .svc-list { display: none; } .svc-empty { display: none; }
@@ -285,14 +315,22 @@ h2.section { margin: 0; font-size: 13px; font-weight: 600; color: var(--muted);
   the fallback above stays a fallback. */
 .disk-cards { grid-template-columns: repeat(auto-fill, minmax(min(420px, 100%), 1fr)); }
 .datasets { padding: 0; overflow-x: auto; }
-.ds-row { display: grid; grid-template-columns: minmax(160px, 1fr) 80px 80px 80px 50px 50px minmax(120px, 1fr); gap: 8px; align-items: center; padding: 6px 12px; border-bottom: 1px solid var(--border); font-size: 13px; min-width: 700px; } /* the columns and gaps add up to 692: at 640 the row ended before its own cells did */
+.ds-row { display: grid; grid-template-columns: minmax(160px, 1fr) 80px 90px 90px 90px minmax(120px, 1fr); gap: 8px; align-items: center; padding: 6px 12px; border-bottom: 1px solid var(--border); font-size: 13px; min-width: 700px; }
 .ds-row.head { color: var(--muted); font-size: 11px; text-transform: uppercase; letter-spacing: .5px; }
 .ds-row .num { text-align: right; }
@media (max-width: 800px) { .ds-row { grid-template-columns: minmax(90px, 1fr) 84px 100px; min-width: 0; gap: 6px; padding: 6px 8px; } .ds-row > :nth-child(3), .ds-row > :nth-child(5), .ds-row > :nth-child(6) { display: none; } .ds-row .ds-name, .ds-row.head > span { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; } .ds-row.head { letter-spacing: 0; } }
 .ds-name { cursor: pointer; display: flex; align-items: center; gap: 6px; white-space: nowrap; overflow: hidden; text-overflow: ellipsis; }
 .ds-mount { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; font-family: ui-monospace, monospace; font-size: 12px; }
 .ds-snaps { padding: 8px 12px 10px 40px; border-bottom: 1px solid var(--border); background: var(--bg); display: flex; flex-direction: column; gap: 6px; }
 .snap-list { display: flex; flex-direction: column; gap: 2px; max-height: 40vh; overflow-y: auto; }
-.snap { display: flex; justify-content: space-between; gap: 12px; font-size: 12px; }
+.snap { display: grid; grid-template-columns: minmax(150px, 1fr) 150px 120px 70px auto; gap: 10px; align-items: center; font-size: 12px; padding: 2px 0; border-bottom: 1px solid var(--border); }
 .snap .snap-name { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
 .snap .snap-when, .snap .snap-exp { white-space: nowrap; }
 .snap .snap-exp.warn { color: var(--s2); }
 .snap .num { text-align: right; }
 .snap .snap-del { padding: 0 6px; font-size: 11px; opacity: .55; }
 .snap:hover .snap-del { opacity: 1; }
@media (max-width: 800px) { .snap { grid-template-columns: minmax(120px, 1fr) auto auto; } .snap .snap-when, .snap .snap-exp { display: none; } }
 /* files */
 .files { position: absolute; inset: 0; display: flex; flex-direction: column; background: var(--bg); }
@@ -388,10 +426,37 @@ h2.section { margin: 0; font-size: 13px; font-weight: 600; color: var(--muted);
 .mitem { width: 100%; display: flex; align-items: center; gap: 8px; padding: 9px 12px; text-align: left; }
 .mitem:hover { background: var(--bg3); } .mitem.on { box-shadow: inset 3px 0 0 var(--accent); }
 .mitem .mname { flex: 1; } .mitem.manage { color: var(--accent2); }
 /* VPN tab: the profile handed over once, with its QR beside the file itself. */
 .vpn-profile .vpn-qr { display: flex; gap: 14px; align-items: flex-start; flex-wrap: wrap; }
 .vpn-profile .vpn-qr img { background: #fff; border-radius: 6px; padding: 6px; flex: none; }
 .vpn-profile .vpn-conf { margin: 6px 0 0; padding: 8px 10px; background: var(--bg); border: 1px solid var(--border); border-radius: 6px; font-size: 11px; line-height: 1.5; white-space: pre-wrap; word-break: break-all; user-select: all; }
 .vpn-peer .sub-head .btn { padding: 2px 8px; }
 .addform label.row-check { flex-direction: row; align-items: center; gap: 8px; }
 .addform label.row-check input { width: auto; }
 /* Network tab: the speed test's four figures, large enough to read at arm's length, and the
   bar that fills while one is running. */
 .net .speedgrid { display: grid; grid-template-columns: repeat(4, 1fr); gap: 10px; }
@media (max-width: 560px) { .net .speedgrid { grid-template-columns: repeat(2, 1fr); } }
 .net .speedfig { min-width: 0; }
 .net .speedfig .label { font-size: 11px; font-weight: 600; color: var(--muted); text-transform: uppercase; letter-spacing: .5px; }
 .net .speedfig .big { font-size: 28px; font-weight: 600; line-height: 1.15; font-variant-numeric: tabular-nums; }
 .net .speedfig .big i { font-size: 12px; font-weight: 400; font-style: normal; color: var(--muted); margin-left: 4px; }
 .net .speedfig.live .big { color: var(--accent); }
 .net .speedbar { display: flex; flex-direction: column; gap: 4px; margin-top: 10px; }
 .net .card.speedtest .svc-actions { margin-top: 10px; }
 .net .card.speedtest .method { margin-top: 8px; padding-top: 8px; border-top: 1px solid var(--border); }
 .net .card.iface { margin-top: 8px; }
 .net .card + .dash-bar { margin-top: 14px; }
 .addform { display: flex; flex-direction: column; gap: 10px; }
 .addform label { display: flex; flex-direction: column; gap: 4px; font-size: 12px; color: var(--muted); }
-.addform input { padding: 8px 10px; background: var(--bg); border: 1px solid var(--border); border-radius: 6px; color: var(--fg); font-size: 15px; }
+.addform input, .addform select { padding: 8px 10px; background: var(--bg); border: 1px solid var(--border); border-radius: 6px; color: var(--fg); font-size: 15px; }
 .header .badge.bad { color: var(--bad); border-color: var(--bad); }
 /* a guest login says so, and how long it has, where the power button would be */
 .badge.guest { color: var(--accent2); border-color: var(--accent2); text-transform: none; margin-left: 0; white-space: nowrap; }
 .card.fresh { border-color: var(--accent2); }
 .card.fresh .mono { user-select: all; }
 /* settings + themes */
 .themes { display: grid; grid-template-columns: repeat(auto-fill, minmax(220px, 1fr)); gap: 12px; }
@@ -542,7 +607,28 @@ a.btn { text-decoration: none; display: inline-flex; align-items: center; }
 .docker .svc-list { min-width: 0; max-width: 100%; } .docker .svc-rows { overflow-x: hidden; } .docker .svc-name .d { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
 .docker:not(.has-sel) .svc-list { flex: 1 1 0; }
 .stack-editor { flex: 1; min-height: 46vh; font-size: 13px; tab-size: 2; white-space: pre; overflow: auto; }
-@media (max-width: 800px) { .docker .svc-list { width: 100%; flex: 1 1 auto; } .stack-editor { min-height: 52vh; font-size: 12px; } }
+
 /* The code editor: a gutter of line numbers, a painted copy of the text behind a textarea
   whose own text is transparent, both in the same monospace at the same size so they line up
   glyph for glyph. .code takes the sizing class the textarea used to (.stack-editor), so it
   sits where that did. */
 .code { display: flex; align-items: stretch; background: var(--term-bg); border: 1px solid var(--border); border-radius: 6px; overflow: hidden; position: relative; padding: 0; min-height: 260px; }
 .code.ro .code-text { color: var(--muted); }
 .code-gutter { flex: 0 0 auto; min-width: 2.6em; padding: 10px 6px 10px 0; text-align: right; color: var(--muted); opacity: .7; background: var(--bg2); border-right: 1px solid var(--border); overflow: hidden; user-select: none; font: 12px/1.5 ui-monospace, SFMono-Regular, Menlo, Consolas, "Liberation Mono", monospace; }
 .code-body { flex: 1; position: relative; min-width: 0; }
 .code-paint, .code-text { margin: 0; padding: 10px; font: 12px/1.5 ui-monospace, SFMono-Regular, Menlo, Consolas, "Liberation Mono", monospace; tab-size: 2; white-space: pre; letter-spacing: 0; word-spacing: 0; }
 .code-paint { position: absolute; inset: 0; overflow: hidden; color: var(--fg); pointer-events: none; }
 .code-text { position: absolute; inset: 0; width: 100%; height: 100%; background: transparent; color: transparent; caret-color: var(--fg); border: 0; outline: 0; resize: none; overflow: auto; }
 .code-text::placeholder { color: var(--muted); }
 .code-text::selection { background: rgba(127, 127, 127, .35); }
 .stack-editor.code, .stack-editor .code-paint, .stack-editor .code-text, .stack-editor .code-gutter { font-size: 13px; }
 .y-key { color: var(--accent2); } .y-colon, .y-dash { color: var(--muted); } .y-string { color: var(--ok); } .y-number, .y-bool { color: var(--s2); } .y-comment { color: var(--muted); font-style: italic; } .y-anchor, .y-block { color: var(--accent); } .y-value { color: var(--fg); }
 /* a phone has no Tab key: a bar of the strokes a compose file needs, pinned to the editor's foot */
 .code-keys { position: absolute; left: 0; right: 0; bottom: 0; display: flex; gap: 6px; padding: 6px 8px; background: var(--bg2); border-top: 1px solid var(--border); z-index: 1; }
 .code-keys button { min-width: 40px; height: 32px; padding: 0 10px; border: 1px solid var(--border); border-radius: 6px; background: var(--bg3); color: var(--fg); font: 15px ui-monospace, monospace; }
 .code-keys ~ * { display: none; }
@media (pointer: coarse) { .code .code-body { padding-bottom: 45px; } .code .code-gutter { padding-bottom: 55px; } }
@media (max-width: 800px) { .docker .svc-list { width: 100%; flex: 1 1 auto; } .stack-editor { min-height: 52vh; font-size: 12px; } .stack-editor.code, .stack-editor .code-paint, .stack-editor .code-text, .stack-editor .code-gutter { font-size: 12px; } }
 /* stacks: diff, images, .env */
 .stack-extra { display: flex; flex-direction: column; gap: 6px; border-top: 1px solid var(--border); padding-top: 10px; } .stack-extra .sub-head .k { font-size: 12px; font-weight: 600; color: var(--muted); text-transform: uppercase; letter-spacing: .5px; }
@@ -575,6 +661,31 @@ pre.diff .add { color: var(--ok); background: color-mix(in srgb, var(--ok) 12%,
 .top-mem, .top-disks { min-height: 165px; } /* a grid row set to auto will otherwise squeeze them to nothing */
 .top-cpu { grid-column: 1 / -1; } .top-mem { grid-column: 1; grid-row: 2; } .top-disks { grid-column: 2; grid-row: 2; }
 .top-net { grid-column: 1 / 3; grid-row: 3; } .top-proc { grid-column: 3; grid-row: 2 / 4; }
 /* a machine with a GPU gets a row of gpu boxes under the cpu, as btop lays it out; the rest move down one */
 .top.has-gpu { grid-template-rows: minmax(200px, 1fr) minmax(210px, auto) auto minmax(170px, 1fr); }
 .top-gpus { grid-column: 1 / -1; grid-row: 2; display: grid; grid-auto-flow: column; grid-auto-columns: minmax(0, 1fr); gap: 4px; min-width: 0; }
 .top.has-gpu .top-mem, .top.has-gpu .top-disks { grid-row: 3; } .top.has-gpu .top-net { grid-row: 4; } .top.has-gpu .top-proc { grid-row: 3 / 5; }
 .top-gpu { min-height: 210px; }
 /* a GPU with nothing to measure is one line, and on a machine with no other its row is only as
   tall as that line. Same weight as .top.has-gpu and after it, so the narrower layouts' own
   .top.has-gpu, further down, still wins there (their gpu row is already auto). */
 .top.gpu-slim { grid-template-rows: minmax(200px, 1fr) auto auto minmax(170px, 1fr); }
 .top-gpu.slim { align-self: start; } /* beside a full gpu box it keeps its one line, not that box's height */
 .t-row.t-gpu-line { flex-wrap: wrap; gap: 0 14px; } .t-gpu-line > span { white-space: nowrap; } .t-gpu-who { margin-left: auto; } /* on a phone it wraps to its own line, still on the right */
 .t-gpu { display: flex; gap: 8px; min-width: 0; flex: 1; min-height: 0; }
 /* the graph and its caption share this box: tall enough for both, so the caption never lands on the stats */
 .t-gpu-graph { flex: 1; min-width: 0; min-height: 130px; display: flex; flex-direction: column; }
 .t-gpu-stats { width: 310px; flex-shrink: 0; border: 1px solid var(--border); border-radius: 3px; padding: 4px 6px; display: flex; flex-direction: column; gap: 1px; overflow-y: auto; scrollbar-width: thin; }
 .t-gpu-stats .t-core-n { width: 58px; } /* room for "enhance" — the cpu box's 30px fits "C15", not an engine's name */
 .t-gpu-name { color: var(--fg); overflow: hidden; text-overflow: ellipsis; max-width: 42vw; }
 .t-core-n.t-eng { color: var(--muted); font-size: 12px; }
 .t-gpu-facts { margin-top: 2px; } .t-gpu-facts .t-row { gap: 6px; } .t-fact { color: var(--fg); white-space: nowrap; }
 .t-gpu-facts .tm { flex: 1; }
 .t-gpu-clients { margin-top: 4px; border-top: 1px solid var(--border); padding-top: 2px; }
 .t-gpu-clients .t-thead, .t-crow { display: grid; grid-template-columns: 52px minmax(0, 1fr) 50px 46px; gap: 7px; align-items: center; white-space: nowrap; }
 .t-gpu-clients .t-thead { border-bottom: 0; padding-right: 0; }
 .t-crow > span { min-width: 0; overflow: hidden; } .t-crow .c-name { color: var(--ok); text-overflow: ellipsis; } .t-crow .c-cpu { color: var(--fg); }
 .t-crow-none { display: block; font-size: 12px; padding: 1px 0; }
 .tbox { position: relative; border: 1px solid var(--border); border-radius: 3px; min-width: 0; min-height: 0; display: flex; flex-direction: column; }
 .tbox-in { flex: 1; min-height: 0; min-width: 0; padding: 9px 8px 6px; display: flex; flex-direction: column; overflow: hidden; }
 .tbox-name, .tbox-right { position: absolute; top: -9px; background: var(--term-bg); padding: 0 4px; font-size: 12px; white-space: nowrap; z-index: 1; }
@@ -599,7 +710,12 @@ pre.diff .add { color: var(--ok); background: color-mix(in srgb, var(--ok) 12%,
 .t-load { display: flex; margin-top: 2px; }
 /* the dot graph and the dotted meter */
-.tg { width: 100%; height: 100%; flex: 1; min-height: 0; overflow: hidden; }
+/* The canvas is taken out of flow. Its drawing size is set from the box it sits in; if it also
   sat *in* that box and the box's height was ever content-sized, each resize would set a bigger
   canvas, which would make a bigger box, which would set a bigger canvas — the Processes tab
   growing without end, as it did when the gpu box landed in an auto-height row. */
 .tg { position: relative; width: 100%; height: 100%; flex: 1; min-height: 0; overflow: hidden; }
 .tg canvas { position: absolute; inset: 0; }
 .tm { position: relative; display: inline-block; height: 9px; flex: 1; min-width: 24px; vertical-align: middle; }
 .tm-t, .tm-f { position: absolute; inset: 0; -webkit-mask-image: repeating-linear-gradient(90deg, #000 0 3px, transparent 3px 5px); mask-image: repeating-linear-gradient(90deg, #000 0 3px, transparent 3px 5px); }
 .tm-t { background: var(--border); }
@@ -638,16 +754,22 @@ pre.diff .add { color: var(--ok); background: color-mix(in srgb, var(--ok) 12%,
 .t-prow .c-user { text-overflow: ellipsis; }
 .c-pid, .c-thr, .c-mem, .c-cpu { text-align: right; } .t-thead .c-pid, .t-thead .c-name, .t-thead .c-cmd, .t-thead .c-user { text-align: left; }
 .c-g { color: var(--muted); text-align: right; }
 /* the Gpu% column, on a machine that has one: dim at rest, lit on the rows that are using it */
 .c-gpu { text-align: right; color: var(--muted); } .c-gpu.lit { color: var(--t-magenta); }
 .t-proc.has-gpu .t-thead, .t-proc.has-gpu .t-prow { grid-template-columns: 58px 116px minmax(0, 1fr) 62px 58px 50px 56px 42px 42px; }
@container tproc (max-width: 640px) {
  .t-thead, .t-prow { grid-template-columns: 58px 116px minmax(0, 1fr) 62px 58px 50px 42px; }
  .t-proc.has-gpu .t-thead, .t-proc.has-gpu .t-prow { grid-template-columns: 58px 116px minmax(0, 1fr) 62px 58px 50px 42px 42px; }
  .t-thead .c-g, .t-prow .c-g { display: none; }
 }
@container tproc (max-width: 560px) {
  .t-thead, .t-prow { grid-template-columns: 58px 110px minmax(0, 1fr) 58px 50px 42px; }
  .t-proc.has-gpu .t-thead, .t-proc.has-gpu .t-prow { grid-template-columns: 58px 110px minmax(0, 1fr) 58px 50px 42px 42px; }
  .t-thead .c-thr, .t-prow .c-thr { display: none; }
 }
@container tproc (max-width: 470px) {
  .t-thead, .t-prow { grid-template-columns: 52px minmax(0, 1fr) 52px 48px 40px; gap: 6px; }
  .t-proc.has-gpu .t-thead, .t-proc.has-gpu .t-prow { grid-template-columns: 52px minmax(0, 1fr) 52px 48px 40px 40px; }
  .t-thead .c-cmd, .t-prow .c-cmd { display: none; }
 }
 .t-foot { display: flex; align-items: center; gap: 8px; border-top: 1px solid var(--border); padding-top: 3px; margin-top: 2px; font-size: 12px; }
@@ -660,15 +782,30 @@ pre.diff .add { color: var(--ok); background: color-mix(in srgb, var(--ok) 12%,
  .top-mem, .top-disks { min-height: 220px; }
  .top-cpu { grid-column: 1 / -1; grid-row: 1; } .top-mem { grid-column: 1; grid-row: 2; } .top-disks { grid-column: 2; grid-row: 2; }
  .top-net { grid-column: 1 / -1; grid-row: 3; } .top-proc { grid-column: 1 / -1; grid-row: 4; }
  .top.has-gpu { grid-template-rows: minmax(230px, auto) auto minmax(220px, auto) minmax(200px, auto) minmax(420px, 1fr); }
  .top.has-gpu .top-mem, .top.has-gpu .top-disks { grid-row: 3; } .top.has-gpu .top-net { grid-row: 4; } .top.has-gpu .top-proc { grid-row: 5; }
 }
@media (max-width: 800px) {
  .top { display: flex; flex-direction: column; font-size: 12px; }
  .top > .tbox { flex: 0 0 auto; }
  .top-gpus { display: flex; flex-direction: column; flex: 0 0 auto; } .top-gpu.slim { align-self: stretch; }
  .t-gpu { flex-direction: column; } .t-gpu-graph { height: 130px; flex: 0 0 130px; } .t-gpu-stats { width: auto; max-height: none; }
  .t-cpu { flex-direction: column; } .t-cpu-graph { height: 130px; flex: 0 0 130px; } .t-cores { width: auto; max-height: 40vh; }
  .t-net { flex-direction: column; } .t-net-nums { width: auto; } .t-net-graphs { flex: 0 0 130px; }
  .top-proc { height: 70vh; min-height: 420px; }
 }
 /* ---- ZFS: the ARC card and each pool's caching tiers ---- */
 .arc-sparks { display: grid; grid-template-columns: 1fr 1fr; gap: 12px; margin-top: 8px; } .arc-sparks .small { margin-bottom: 2px; }
@media (max-width: 640px) { .arc-sparks { grid-template-columns: 1fr; } }
 .arc-break { display: grid; grid-template-columns: minmax(140px, 1fr) 80px 110px 110px; gap: 2px 10px; font-size: 12px; margin-top: 10px; }
 .arc-break .head { color: var(--muted); } .arc-break .num { text-align: right; }
@media (max-width: 640px) { .arc-break { grid-template-columns: minmax(0, 1fr) auto auto auto; gap: 2px 8px; font-size: 11px; } .arc-break > span { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; } }
 .arc-tiers { display: flex; flex-direction: column; gap: 5px; font-size: 13px; margin-top: 4px; }
 .arc-tier { display: flex; gap: 8px; min-width: 0; } .arc-tier .k { color: var(--muted); width: 80px; flex-shrink: 0; cursor: help; }
 .arc-tier-v { display: flex; flex-direction: column; gap: 1px; min-width: 0; } .arc-tier-v > span { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
 .arc-tier-v .none { color: var(--muted); white-space: normal; }
 /* ---- Space: what is using a disk ---- */
 .space { position: absolute; inset: 0; display: flex; flex-direction: column; background: var(--bg); min-height: 0; }
 .space-bar { display: flex; flex-direction: column; gap: 8px; padding: 10px 12px; border-bottom: 1px solid var(--border); background: var(--bg2); }
@@ -690,6 +827,7 @@ pre.diff .add { color: var(--ok); background: color-mix(in srgb, var(--ok) 12%,
 .space .warn { color: var(--s2); }
 .space-tabs { padding: 8px 12px 0; }
 .space-split { flex: 1; display: flex; flex-direction: column; min-height: 0; padding: 8px 12px 12px; gap: 8px; }
@media (max-width: 800px) { .space-split { overflow-y: auto; padding: 8px; } .space-tree { flex: 0 0 auto; height: 40vh; } .space-map { flex: 0 0 auto; height: 60vh; } }
 .space-tree { flex: 1 1 55%; min-height: 120px; display: flex; flex-direction: column; border: 1px solid var(--border); border-radius: 6px; background: var(--bg2); overflow: hidden; }
 .space-rows { flex: 1; overflow: auto; min-height: 0; }
 .space-row { display: grid; grid-template-columns: minmax(160px, 1fr) 84px 120px 70px 64px 88px; gap: 8px; align-items: center; padding: 3px 8px; font-size: 13px; border-bottom: 1px solid var(--border); cursor: pointer; white-space: nowrap; }
@@ -733,3 +871,34 @@ pre.diff .add { color: var(--ok); background: color-mix(in srgb, var(--ok) 12%,
 /* ZFS: the schedule dialog's controls match the rest of the modal forms */
 .modal select, .modal input[type=time], .modal input[type=number] { padding: 9px 10px; background: var(--bg); border: 1px solid var(--border); border-radius: 6px; color: var(--fg); font-size: 16px; }
 .modal .chips { margin-top: 2px; }
 /* ---- KVM: a device's own web page, framed ---- */
 .kvm { position: absolute; inset: 0; display: flex; flex-direction: column; background: var(--bg); min-height: 0; }
 .kvm-bar { display: flex; align-items: center; gap: 8px; padding: 8px 12px; border-bottom: 1px solid var(--border); background: var(--bg2); flex-wrap: wrap; }
 .kvm-bar .kvm-url { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; max-width: 28ch; }
 .kvm-frame { flex: 1; min-height: 0; background: #000; position: relative; }
 .kvm-frame iframe { width: 100%; height: 100%; border: 0; display: block; background: #000; }
 .kvm-frame.full { background: #000; }
 .kvm-hint { padding: 6px 12px; border-top: 1px solid var(--border); }
@media (max-width: 800px) { .kvm-bar .kvm-url { display: none; } }
 .kvm-none { text-align: center; } .kvm-none p { margin: 4px 0; }
 /* No true fullscreen for a page element (Safari on iOS): fill the viewport in-page instead. */
 .kvm-frame.expanded { position: fixed; inset: 0; z-index: 60; }
 .kvm-frame.expanded iframe { height: 100%; }
 .kvm-shrink { position: absolute; right: max(10px, env(safe-area-inset-right)); bottom: max(10px, env(safe-area-inset-bottom)); z-index: 61; background: var(--bg2); border: 1px solid var(--border); opacity: .75; }
 .kvm-shrink:hover, .kvm-shrink:active { opacity: 1; }
 .updot { margin-left: 6px; font-size: 11px; font-weight: 700; }
 .updot.behind { color: var(--s2); } .updot.unknown { color: var(--muted); }
 /* ---- Virtualization ---- */
 .virt .svc-tools .task { max-height: 45vh; }
 .virt-table { width: 100%; border-collapse: collapse; font-size: 13px; }
 .virt-table th, .virt-table td { text-align: left; padding: 6px 8px; border-bottom: 1px solid var(--border); vertical-align: top; }
 .virt-table th { font-weight: 500; opacity: .7; }
 .virt-table td.num, .virt-table th.num { text-align: right; white-space: nowrap; }
 .virt-pick { max-height: 240px; overflow-y: auto; border: 1px solid var(--border); border-radius: 6px; }
 .virt-pick .svc-row { padding: 6px 10px; }
 .virt-kv { display: grid; grid-template-columns: max-content 1fr; gap: 2px 12px; font-size: 12px; word-break: break-all; }
 .virt-form-row { display: flex; gap: 8px; }
 .virt-form-row > label { flex: 1; min-width: 0; }
 .virt .modal { max-width: 640px; }
@@ -1,4 +1,4 @@
-import { ReactNode, useEffect, useRef, useState } from 'react'
+import { Component, ReactNode, useEffect, useRef, useState } from 'react'
 // ---- spinners ----
 export function Spinner({ size = 14 }: { size?: number }) {
@@ -155,3 +155,34 @@ export async function okToLeave(): Promise<boolean> {
  if (!unsaved || (await askYes(`Discard unsaved changes to ${unsaved}?`, 'Discard'))) { unsaved = ''; return true }
  return false
 }
 // A page that throws takes the whole console down with it: React unmounts the tree, and what
 // is left is a blank window with no clue in it. The VPN tab did exactly that on a machine
 // whose tunnel list came back null. So each page is rendered inside one of these, and a fault
 // costs that page rather than the session — the rail, the switcher and every other tab still
 // work, and the error says which page and what it said.
 export class PageBoundary extends Component<{ page: string; children: ReactNode }, { err: Error | null }> {
  constructor(props: { page: string; children: ReactNode }) {
    super(props)
    this.state = { err: null }
  }
  static getDerivedStateFromError(err: Error) { return { err } }
  componentDidCatch(err: Error) { console.error('page failed:', err) }
  // A different page is a different mount; clear the fault when the page changes.
  componentDidUpdate(prev: { page: string }) { if (prev.page !== this.props.page && this.state.err) this.setState({ err: null }) }
  render() {
    if (!this.state.err) return this.props.children
    return (
      <div className="dash">
        <div className="dash-bar"><h2 className="section">{this.props.page}</h2></div>
        <section className="card">
          <div className="error">This page hit a fault and stopped drawing. The rest of the console is unaffected.</div>
          <div className="muted small mono" style={{ marginTop: 8 }}>{String(this.state.err)}</div>
          <div className="svc-actions" style={{ marginTop: 10 }}>
            <button className="btn" onClick={() => this.setState({ err: null })}>Try again</button>
          </div>
        </section>
      </div>
    )
  }
 }