R-330: the agent sends SMART 187/188/199 (raw; omitted when unknown)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0159rPz1ZhFKsS53msqPYxtS

.claude/rules/health-checks.md

@@ -21,6 +21,7 @@ fast, and wrong.
 
 This rule used to be duplicated verbatim in felhom-agent/CLAUDE.md with a note explaining that
 felhom.eu/CLAUDE.md "does not load in an agent-only session". That reasoning was correct before
-path-scoped rules existed. The single source is now felhom.eu/CLAUDE.md "Code quality rules"; this
+path-scoped rules existed. Deliberate scoped copies now live in felhom.eu/.claude/rules/hub.md and
-file is the scoped copy that loads exactly where health checks are written. (2026-08-06)
+felhom-controller/.claude/rules/gates.md (hub.md's comment names them); none is the single source. This
+file is the copy that loads exactly where agent health checks are written. (2026-08-06; corrected 2026-10-06)
 -->

.claude/rules/unprompted-work.md

@@ -0,0 +1,69 @@
+---
+unconditional: true
+---
+# Unprompted work — rules for any session without a task file
+
+> Goal sessions, nightly sessions, "work the register" sessions. **A session that starts from
+> `/goal` or a standing brief inherits these rules exactly as it inherits the gates.** They are the
+> part of `PROMPT-TEMPLATE.md` that a task file used to carry and a goal does not. Same wording lives
+> in `felhom.eu`, `felhom-controller`, `felhom-agent` and `app-catalog-felhom.eu` `.claude/rules/`, and in the workspace
+> root's unversioned `.claude/rules/`; change all five or none.
+
+## 1. What you may pick up on your own
+
+- A register row **you or another CC session filed**, with owner CC, at P3 or a bounded P2, that
+  needs **no operator decision**, touches **no customer data by design**, and introduces **no
+  mechanism nobody has measured**. Smallest first.
+- A defect you find while exercising the product, filed as a row **before** you fix it — **unless it is small**:
+  a small finding is fixed in the session and never filed (the size rule, `OPEN-ITEMS.md` „How a row is filed").
+- Hygiene: register compression, stale citations, rows with no owner, documents that contradict
+  live source.
+
+**Not yours, ever, without a task file or an operator word:** money; anything that changes risk to
+customer data; anything that changes a promise the product makes to a customer; anything that
+reverses a documented design decision (`documentation/architecture/` — a design decision is not a
+defect, R-370); anything on DooPlex or ep0; baking or vouching a golden; promoting a
+catalog version; a new external dependency.
+
+## 2. When you may decide instead of ask (operator grant, 2026-09-14)
+
+You may take a decision yourself when **all** of these hold: the architecture folder and the register
+give a clear direction; your choice follows that direction; it is reversible without customer-data
+risk; and you can write it in the `09-update-architecture.md` §3 shape — one answerable sentence, the
+options, what each costs, why this one. **Then record it** as a dated decision in `CONTEXT.md` and
+the owning architecture document, tagged *decided by CC unattended — operator may reverse*, and put
+it **first** in the morning note. A decision you cannot write in that shape is one you do not take.
+
+## 3. The discipline a task file used to carry
+
+1. **Baselines first.** Read each repo's `main` hash and version from live source before touching it.
+2. **Read the architecture document for the area, and name it** in the report, before any claim.
+3. **Red-proof every correctness fix.** A test never seen failing has not been shown to test anything.
+4. **Live-validate on a Tier-0 box** through the endpoints the UI invokes. `demo-hp` is `ssh hp`.
+   Throwaway apps only; the standing apps and `bentopdf` stay.
+5. **Evidence off the machine at the end of each phase**, before any revert (R-320).
+6. **One release per repo per session**, with a CHANGELOG entry (controller: with its `MinAgent`
+   line), REPORT overwritten, floor raised to deliver it. **No golden unless a drill or fresh install
+   needs one** (the waiver, R-468). **No `--no-verify`.**
+7. **An enumerated gap becomes a row in the same session — or, if it is small, is fixed in it** (the size rule).
+   Prose is not a record.
+8. **Hungarian text is searched with ASCII fragments**, with a positive and a negative control.
+9. **Never leave a half-state.** If time runs out, revert to clean and say what was reverted.
+10. **Teardown, three layers, stated** — machine, host, hub — or "provisioned nothing".
+
+## 4. The morning note
+
+One screen, plain language, in this order: **decisions you took** (§2) first; what you exercised;
+what broke and whether you fixed it; rows opened and closed with the register size before and after;
+what needs the operator, each with what happens if they do nothing. No file paths, no function
+names, no row numbers as the subject of a sentence.
+
+## 5. Instruction files
+
+**Instruction files (`CLAUDE.md`, `.claude/rules/*`) are kept true by the session that finds them wrong**
+(operator ruling 2026-10-06, `09` §3 decision 150). A session MAY, without asking: correct a stale fact (a command, a
+count, a version, a path, a description of what a gate does), add a fact it proved, and remove a reference to something
+that no longer exists. Each edit is named in the report (file, line, before, after, why). A session MAY NOT, without the
+operator's word: loosen a safety rule, a fence, a „never", a protected machine, a secret rule, or a review step; or
+remove a rule. When in doubt, it is a rule change, and it goes to the operator. If Claude Code's own permission check
+asks before such an edit, wait for the operator's click; if it refuses, record that and file the exact line.

CHANGELOG.md

@@ -1,5 +1,57 @@
+## Unreleased (2026-10-06 night, later) — after a restart the agent remembers the last backup per tier (R-894); three more SMART counters on the wire (R-330)
+
+Ships with the memory-kill check below as v0.150.0, AFTER the 2026-10-07 night read-back. Nothing delivered tonight.
+
+- **The defect (measured 2026-10-05 on demo-hp):** the agent restarted at 04:57; at 06:25 the off-site storage answered *Can't connect*; the per-tier backup record is in memory only, so the due-check fell back to an EMPTY record and the 7-day tier (last copy 4 days old) read DUE; the controller asked and vzdump failed.
+- New `internal/backup/backup_state.go` `BackupSuccessState`: the newest SUCCESSFUL backup per tier and guest, on disk (`<oob state dir>/backup-success-state.json`, atomic tmp+rename, 0600). Only successes are written; a corrupt file reads as nothing known.
+- `internal/localapi` `handleBackupDue`: when the tier's storage CANNOT be read, the saved copy stands in for the in-memory record. A fresh copy → not due („… (storage unreadable — age from the last success saved on disk)"); a copy older than the cadence → DUE; no copy → the old answer (DUE, age unknown). A storage that answers stays the ground truth: an archive absent there is due even when the file remembers one.
+- Wired in `buildLocalAPIServer` (`LastKnownBackups`); the local API's backup job saves each success.
+- Tests: `TestBackupDue_R894_*` (restart = a new server and a new state from the same file; fresh / old / none / storage answers / failed backup not saved), `TestBackupSuccessState_*`, `TestR894_LastKnownBackupsIsWiredIntoTheDaemon` (AST). Four red-proofs observed (`felhom.eu/documentation/audits/night-burndown-2026-10-06/s4/`).
+- **R-330 (disk health Phase 2, the wire only):** the SMART summary carries three more SATA raw counters — `reported_uncorrect` (187), `command_timeout` (188, carried as the vendor reports it; some pack several counters), `udma_crc_errors` (199). Pointer + omitempty: an attribute the drive does not report is OMITTED (unknown), never 0. No verdict reads them yet. Tests `TestParseSMART_R330_*` (two red-proofs, `felhom.eu/documentation/audits/night-burndown-2026-10-06/r330/`).
+
+## Unreleased (2026-10-06 night) — the Docker step proves the engine reports a memory kill (`09` §3 decision 157, R-528)
+
+To be released as v0.150.0 with its config bundle AFTER the 2026-10-07 night read-back (the night of 2026-10-06 runs v0.149.0 on purpose).
+
+- `configs/felhom-os-apply`: after a docker-layer APPLY (after `health_after`) the wrapper runs `oom_check()`: a throwaway container from the image the running controller uses (`--pull never`, `--network none`, no volume, label `felhom.oomcheck=1`, 64 MB cap) asks for one 200 MB block; „pass" only when `OOMKilled=true` AND the `oom` event; it waits 2 s and reads the events window to the guest's epoch + 1 (measured: a window closed in the same second missed the event); the container is always removed. Reported as `oom_check`; it never changes the step's outcome or health. A wrapper-only mode `oom-check` runs the check alone (no apt, no engine change), by hand as root.
+- `internal/osupdate`: `WrapperReport` and `Report` carry `oom_check` verbatim, on the normal pass and on the kept-copy path (R-868).
+- `configs/test_felhom_os_apply.py`: its `unittest.main()` sat in the middle of the file, so 11 tests (UnsentReport, SaveReportOnDisk, AgentDiesMidPass, CrashLeftTheJournal) never ran — moved to the end; all pass.
+- Tests: the OOMCheck class (pass, OOMKilled=false, no event, unreadable image, removal on an inspect error, not on other layers or in health mode, the events window after the settle wait, mode oom-check alone and its refusals); TestDocker_OOMCheckReachesTheHubUnchanged, TestR868_KeptCopyCarriesTheOOMCheck. 12 red-proofs in `felhom.eu/documentation/audits/readback-2026-10-07/F/`.
+
+## Unreleased (2026-10-06 evening) — the shared rule file (`09` §3 decision 152); no code change
+
+- `.claude/rules/unprompted-work.md` added, byte-identical to the copies in felhom.eu, felhom-controller, app-catalog-felhom.eu and the workspace root (checked with `diff` against the controller's copy and one md5 across all five). Its copies line names five copies.
+
+## Unreleased (2026-10-06 afternoon) — instruction files kept true (`09` §3 decision 150); no code change
+
+- `CLAUDE.md` „Gates — ONE entry point": the runner runs every gate in its `GATES` table (five: three shared, `published`, `release-complete`); `--fast` skips `published` (network). It said two gates and „all of them".
+- `CLAUDE.md`: the decoy gate and its audit are named with their `felhom.eu/` prefix (they do not exist in this repo).
+- `.claude/rules/health-checks.md` (comment): the health-check rule's copies live in felhom.eu `hub.md` and the controller's `gates.md`; it named felhom.eu `CLAUDE.md` „Code quality rules", which holds no such rule.
+
+## v0.149.0 — a weekly disk trim of each customer guest, the crash-boot fact for the controller, the phantom WARN names its runbook (R-444, R-856, R-99; operator rulings `09` §3 139, 143, 140) (2026-10-06)
+
+Released by `scripts/release-agent.sh`: binary sha256 `6bcae9c2eb5d97e8285316583870059835793893299e291891a53a4ce505585f`
+config bundle sha256 `e182c82dcf4a67faa3bcb74dbe4ffa7b06e0b27dc8451cb7574d6339ce91ad66` (tag `v0.149.0` = `f277e61`).
+**The bundle carries the new sudoers rule for the trim (`FELHOM_FSTRIM`) — deliver it with the binary:** signed
+`agent_update`, then signed `agent_config_update`.
+
+- R-856 (`09` §3 decision 143): new local-API route `GET /host/crash-guard` — passes the host crash guard's last-boot record (present, last_boot_at, last_boot_unclean, tripped) from /var/lib/felhom-crash-guard/state.json to the controller, which waits ~15 min with app mails after a crash boot. Read-only, no Proxmox call, guest-token authed; a missing/unreadable/garbled file answers 200 present:false (never an error page). An older agent answers 404, which the controller reads as unknown (normal 90 s grace) — no controller MinAgent raise needed.
+- R-444 (`09` §3 decision 139): weekly guest disk trim. New sudoers alias FELHOM_FSTRIM with ONE exact rule `/usr/sbin/pct ^fstrim [0-9]+$` (rides the signed config bundle; decoys pinned by TestSudoersFstrimRuleIsExact) and capability guest-fstrim (non-critical). New internal/fstrim job: each owned RUNNING guest gets `pct fstrim <vmid>` once a week - due Wednesday from 10:00 host-local, starts only 10:00-20:59 (never the 01:00-06:59 night), holds the one-heavy-op gate so it never runs beside a backup or restore-test (busy -> deferred to the next hourly tick; a box that was off catches up at its next daytime hour); a failed trim WARNs and is retried at most 3 times that week; bytes parsed from `pct fstrim`'s "(N bytes) trimmed" lines; positive log `fstrim: guest N trimmed X GiB in Ys`; last result per guest persisted in <state_dir>/guest-disk-trim.json and reported as the new omitempty host-report stanza `guest_disk_trim`. Opt-out: agent.json "disk_trim": {"disable": true}.
+- R-99 (`09` §3 decision 140): the agent's WARN for a PBS archive below the 1 MiB plausibility floor now ends with the pointer to the sanctioned cleanup (`documentation/runbooks/pbs-phantom-cleanup.md`); detection only — nothing is deleted automatically. Dir-storage archives keep the old text.
+
 ## unreleased
 
+- R-426: scripts/test_gate_decoys.py (new) — the published gate judged against a fake Gitea (127.0.0.1, via GITEA_BASE; never the real registry): 11 cases; COVERS published — `felhom-agent/published` leaves the decoy-coverage EXEMPT list.
+- R-426: release-complete gate — 10 decoy cases (scratch clone + scratch bare origin + fake Gitea); COVERS release-complete — `felhom-agent/release-complete` leaves the decoy-coverage EXEMPT list.
+- R-426: the shared reuse-refs/instructions/observations gates get agent-side decoys (9 cases on a scratch clone of this repo); COVERS reuse-refs, instructions, observations — three `felhom-agent/*` entries leave the decoy-coverage EXEMPT list.
+- **Fixed without a row:** `configs/test_felhom_config_bundle.py` read the two ISO first-boot files that installer 1.32.0 now NAMES under KEPT (R-275) as files the installer writes — `go test ./internal/osupdate` was red on DooPlex from 21:25 to 01:55 (felhom.eu `85de3f9b`); they are listed with why. Test only; agent v0.148.0's code is unaffected.
+
+## v0.148.0 — the host report names the running binary's sha; the format answer carries the new filesystem's UUID (burn-down night: R-349, R-25 agent halves) (2026-10-06)
+
+Released by `scripts/release-agent.sh`: binary sha256 `3e68a0870e0e2ce262cb4819294edddeb0a73e8c558a31611a20329a6d9ee283`
+config bundle sha256 `a6fa4f589d184b58c9911303bd087e300be1e75b3647e4302c9594df6989c4de` (tag `v0.148.0` = `861d32a`).
+Delivery order as for v0.147.0: signed `agent_update`, then signed `agent_config_update`.
+
 - **R-349:** the host report carries `agent_sha256`, the sha256 of the running agent binary (read once from
   `/proc/self/exe`; empty = unknown), so a hand-built binary under the vouched version name becomes visible. The hub
   comparison is a separate hub change. Test `TestCollect_AgentSHA256IsTheRunningBinary`; red-proved.

CLAUDE.md

@@ -52,11 +52,11 @@ This is in the core because breaching it is how this component stops being audit
 
 ## Gates — ONE entry point
 
-**Run `python3 scripts/agent_gates.py` from the repo root after ANY change here.** It runs this
+**Run `python3 scripts/agent_gates.py` from the repo root after ANY change here.** It runs every
-repo's gates — `reuse_refs_check` and `instructions_gate`, both the **shared** copies in
+gate in its `GATES` table (that table is the list); the shared ones — `reuse_refs_check`,
-`felhom.eu/scripts/`, never copied into this repo (a copy recreates the drift they detect; an absent
+`instructions_gate`, `observations_gate` — are the copies in `felhom.eu/scripts/`, never copied into
-sibling clone FAILS). `--fast` selects the gates touching no network and no container runtime; today
+this repo (a copy recreates the drift they detect; an absent sibling clone FAILS). `--fast` selects the
-that is all of them. **A missing gate is a FAILURE, never a skip.**
+gates touching no network and no container runtime, and skips `published` (network), naming it. **A missing gate is a FAILURE, never a skip.**
 
 **The pre-push hook** (`.githooks/pre-push`) runs it with `--fast` and refuses a failing push. It is
 **per-clone** — switch it on once with `git config core.hooksPath .githooks`, and a manual run WARNS
@@ -104,8 +104,8 @@ the mechanism are exempt.
 
 **A gate ships with a decoy test that has been seen to fail (R-421).** A decoy is the LABEL without
 the FACT — a directory with the right name and no bake log, a note whose prose mentions the marker it
-lacks. `scripts/decoy_coverage_gate.py` refuses a new gate that has neither a decoy nor a named
+lacks. `felhom.eu/scripts/decoy_coverage_gate.py` (run by felhom.eu's `repo_gates.py`, for all four repos) refuses a new gate that has neither a decoy nor a named
 exemption carrying its row. The four shapes, the 2026-09-01 sweep that fooled 16 of 29 gates, and the
-decoys withdrawn as illegitimate: `documentation/audits/AUDIT-gate-decoys-2026-09-01.md` and
+decoys withdrawn as illegitimate: `felhom.eu/documentation/audits/AUDIT-gate-decoys-2026-09-01.md` and
 `felhom-controller/.claude/rules/gates.md`. **Scope is a fact too** — prefer `os.walk` over
 `os.listdir`, and a glob over a hand-maintained list.

REPORT.md

@@ -1,10 +1,7 @@
-# REPORT — agent v0.147.0 (2026-10-05, burn-down round 2)
+# REPORT — the shared rule file (2026-10-06 evening)
 
-Full session report: `felhom.eu/REPORT-burndown2-2026-10-05.md`. Baseline `d833163` (v0.146.1). Code commit `f1b9b41`
+Operator ruling 2026-10-06 14:24 (`09` §3 decision 152): the agent repo gets its copy of the shared rule file. Added
-(CI job 1365 success), tag `v0.147.0`, binary sha256 `642c4d19…`, bundle sha256 `326527d0…` (verified by download).
+`.claude/rules/unprompted-work.md`, byte-identical to the other copies (`diff` against felhom-controller's copy: no
-
+output; md5 `c1e6c881…` across all five before the copies line changed, one md5 after). No code changed; no release.
-Rows: R-124 (recipe root namespace = ""), R-118 (no root size for an absent drive), R-269 (rotated-out token rejected at
+`agent_gates.py --fast`: reuse-refs, instructions, release-complete, observations OK. The session report is
-once), R-317 (dnsmasq install probed by its unit). Tests + red-proofs: `felhom.eu/documentation/audits/burndown2-2026-10-05/`.
+`felhom.eu/REPORT.md`.
-`go build/vet/test ./...` green; `agent_gates.py --fast` green after the release (release-complete needs the tag).
-
-Delivery: see the session report (vouch, signed jobs per box, the hub System page afterwards).

REUSE.md

@@ -121,7 +121,7 @@
 | Anti-retarget durable-id binding | internal/localapi/wipe_reresolve.go | resolve id → re-derive + exact match → re-inspect expected state → act on RE-RESOLVED device only |
 | Atomic single-file JSON store | internal/storage/intent.go | `Open*` loads (missing=empty, corrupt=fail-loud), mutex, tmp+rename 0600, idempotent set |
 | Durable append-only log + index | internal/authz/noncestore.go (`FileNonceStore`) | fsync before returning "new"; replay into index on open; expiry-only compaction |
-| Injectable seam funcs on Server | internal/localapi/server.go (`reresolveWipe`, `deviceDurableID`, `boundCheck`, `deviceCheck`, `livenessCheck`, net-verify: `netTrigger`/`netMounted`/`netJournal`/`netReachable`) | prod default wired in `NewServer`; tests override — no real /dev, /proc/mounts, journalctl or TCP in tests. **For mount-table predicates prefer the DATA seams `procSelfMountinfo` / `procGuestMountinfo` (internal/localapi/intermediary.go) over `boundCheck`/`livenessCheck`**: pointing them at a captured fixture runs the real parser, the real predicate and the real handler, so the test cannot go hollow the way R-116's did |
+| Injectable seam funcs on Server | internal/localapi/server.go (`reresolveWipe`, `deviceDurableID`, `boundCheck`, `deviceCheck`, `livenessCheck`, net-verify: `netTrigger`/`netMounted`/`netJournal`/`netReachable`; R-856 `crashGuardStatePath` — GET /host/crash-guard's state file, internal/localapi/crashguard.go) | prod default wired in `NewServer`; tests override — no real /dev, /proc/mounts, journalctl or TCP in tests. **For mount-table predicates prefer the DATA seams `procSelfMountinfo` / `procGuestMountinfo` (internal/localapi/intermediary.go) over `boundCheck`/`livenessCheck`**: pointing them at a captured fixture runs the real parser, the real predicate and the real handler, so the test cannot go hollow the way R-116's did |
 | `Server.devicePresent` (R-113, v0.114.0) | internal/localapi/disks.go | `devicePresent(rawMountPath) bool`; seam `deviceCheck`, default `isHostMountpoint` | the agent's DEVICE-presence signal — asks whether the drive's RAW mount is still mounted | **Use this, never the bind, to answer "is the drive there".** The raw mount is a device-bound systemd unit and dies with its device; the agent's own bind under the shared parent is NOT device-bound and outlives it as a stale shell. `BoundUnderParent` is now `boundUnderParent(...) && devicePresent(...)` at BOTH /disks construction sites — dropping either half is a regression with its own red-proof. Empty path ⇒ **true** (unknown is never absent: absent stops a customer's apps) |
 | `bindLiveness` + `BindLiveness` (R-117, v0.117.0) | internal/localapi/intermediary.go | `bindLiveness(stable, raw) BindLiveness`; seam `livenessCheck`; read verdicts ONLY via `.Usable()` | the agent's bind-LIVENESS signal — the third term of `BoundUnderParent` | **`devicePresent` and `boundUnderParent` are both PATH-PRESENCE tests and neither is liveness.** They compare only mountinfo field 5, so both stay true over a bind that names the drive that went away while the raw mount healed onto the returning one (measured: raw 8:32 /dev/sdc, bind 8:16 /dev/sdb `shutdown`, EIO both ways, payload healthy). Two dead states, and a fix needs BOTH checks: devno mismatch (the detach/return case) AND the ext4 abort tokens `shutdown`/`emergency_ro` (the steady-state case, where the devnos AGREE because the device never left). **THREE states, never a bool** — `BindUnknown` must exist and `Usable()` treats it as PRESENT (absent stops a customer's apps). **Order matters:** compare devices first and read the abort flag off the RAW mount in the stale case — abort-first classifies the real return state as aborted and refuses the re-bind that repairs it. **NO BLOCK I/O, ever** (CLAUDE.md rule; a probe on a wedged device survives SIGKILL). 6 red-proofs |
 | `AttachDrive` repair ruling (R-117, v0.117.0) | internal/localapi/intermediary.go | the `switch bindLiveness(...)` inside the `n == 1 && GuestSeesMount` arm | decides whether the existing self-heal runs | `BindStaleDevice` ⇒ **re-bind** (the raw mount is a healthy new superblock; repairs live, no guest restart). `BindAborted` ⇒ **quiet no-op** — a re-bind lands on the SAME dead superblock and this runs every 20 s, so re-binding is an infinite silent retry that also masks the state; it must surface via `BoundUnderParent=false`. `BindLive`/`BindUnknown` ⇒ no-op, unchanged. **Do not return an error for the aborted case** — the reconcile loop would log a failure every 20 s |
@@ -160,8 +160,10 @@
 | `lanresolver.hostRoot` + `dnsmasqUnitPaths` (data seam, R-317) | internal/lanresolver/lanresolver.go | prod `hostRoot = "/"`; probe = the `dnsmasq` package's systemd UNIT, never `/usr/sbin/dnsmasq` (owned by `dnsmasq-base`) | internal/lanresolver/ensure_dnsmasq_test.go — fixture root tree + recording `proxmox.Runner`; the REAL `os.Stat` probe and `EnsureDnsmasq` run. `TestEnsureDnsmasq_ProductionProbeIsTheUnit` pins the production wiring |
 | `localapi.GuestAPI` / `BackupService` / `BackupStore` / `TokenAuthority` | internal/localapi/server.go | `*proxmox.Client`, `*backup.BackupRunner`, `*backup.Store`, `*TokenStore` | `fakeGuests`/`fakeBackups`/`fakeStore` internal/localapi/server_test.go |
 | `backup.InFlight` | internal/backup/inflight.go | `TryAcquire(what) (release, busy, ok)` / `Busy()` | THE host-wide "one heavy guest operation at a time" gate — shared by the local-API backup path and the restore-test scheduler (R-85) | A **LINK** guard, not a lock one: the scratch VMID never touches the live guest's vzdump lock, but an offsite restore PULLS multi-GB over the tunnel a backup PUSHES one. Callers **DEFER, never cancel** — a deferred restore-test costs coverage, a cancelled backup costs the backup. A nil gate is ungated (pre-R-85 callers). |
+| `fstrim.Trimmer` (R-444) | internal/fstrim/fstrim.go | `New(runner, guests, gate, statePath, logger)` / `Pass(ctx)` / `GuestDiskTrimStatus(ctx)` / `ParseTrimmed(out)` | the weekly `pct fstrim <vmid>` of owned running guests (Wednesday from 10:00 local, starts 10:00-20:59 only), under `backup.InFlight`; last result per guest persisted and reported as `guest_disk_trim` | A busy gate DEFERS to the next hourly tick, never waits; a failed trim retries at most `MaxAttemptsPerWeek`; the report reads the persisted record, it never runs pct |
 | `capability` store-grant probe (`storeGrantStatuses` / `storeGrantVerdict` / `Client.Permissions`) | cmd/felhom-agent/main.go, internal/proxmox/query.go | *"may the agent READ this backup tier?"*, one `capability.Status` per configured tier | R-185. **Never infer permission from an empty content listing** — `{"data":[]}` is what a FORBIDDEN tier and a NEWBORN tier both return, and that ambiguity hid an unreadable host tier on both demo boxes. Ask `/access/permissions` **as the agent's own token** (root always says yes). **The ungranted answer is not empty and not a 403** — it carries the privileges inherited from the box-wide `/` grant, so test for **`Datastore.AllocateSpace`** specifically; path-presence or `Datastore.Audit` reports a blinded storage healthy. Probed set comes from `BackupTiers()`, never a fixed list. Critical except the `local` fallback. Composes AROUND the sudo prober (the `poolReadStatus` precedent); `Status`'s wire shape is untouched so the hub alert is free. Unreachable PVE ⇒ degraded, never ok. |
 | `backup.RestoreTestState` | internal/backup/restoretest_state.go | `RecordSuccess(target,archive,tier,verified,t)` / `ProvenArchive(target)` / `ProvenRestoreTests(ctx)` / `LastSuccess(target)` / `OldestFirst(targets)` | Per-tier restore-test PROOF state, persisted (atomic tmp+rename) — **which archive** was proven, and when (R-86) | **Credit ONLY on success** — a permanently failing tier must keep sorting first, or it looks freshly proven and stops being retried. Ties break on target id: without it, two tiers proven in the same second rotate by Go's randomised map order. **This one NEEDS persistence unlike R-84** — R-84 had ground truth to consult (the archive is still on the storage); a restore-test destroys its scratch and leaves no artifact. **R-86: the ARCHIVE is the state, the time is metadata** — a time alone cannot answer "have we proven THIS archive", which is the due-check's whole question. A pre-R-86 file (bare RFC3339 per target) keeps its time and yields NO proven archive, so each tier is due once after the upgrade; reading a legacy time as proof of the current archive would invent a guarantee. **R-189: it is also the REPORTABLE half of the restore-test signal.** The in-memory `backup.Store` holds only this process's latest run, and under per-archive due-ness the agent will not re-test a proven archive — so a proof lost to a restart is not repeated for a whole archive generation (observed live: a passing 14.5 GB offsite restore reached no host-report). `ProvenRestoreTests` renders the stored proofs as `hub.RestoreTest` entries and the collector merges them; a record missing the archive or the tier is NOT emitted, because an unproven tier reading as proven is worse than the defect. **Only successes are stored, deliberately:** a success suppresses future work, a failure causes it. |
+| `backup.BackupSuccessState` | internal/backup/backup_state.go | `RecordBackupSuccess(target, b)` / `LastKnownSuccess(target, vmid)` | Newest SUCCESSFUL backup per tier+guest, persisted (atomic tmp+rename) — the due-check's fallback when the tier's storage cannot be read after a restart (R-894) | **Read ONLY when the storage cannot be read** — a storage that answers is the ground truth (R-84), and an archive absent there must make the tier due even when this file remembers one. A saved copy older than the cadence still reads due. Only successes are written. |
 | `hub.ProvenRestoreTestReporter` + `Collector.SetProvenRestoreTests` | internal/hub/collect.go | the DURABLE restore-test source, merged with the in-memory one | R-189. Merge rule: **one entry per tier, newest by `TestedAt` wins** — a fresh failure beats a stored success (the failure is the news, and it lives nowhere else), a stored success beats a stale in-memory entry after a restart, and a tier never appears twice (the hub would read two tests). An unparseable timestamp counts as OLDER, so a malformed entry cannot displace a good one. **The wiring is pinned by an AST test** — the method this replaced (`RestoreTestState.Snapshot`) carried a doc comment naming a host-report gauge and had no caller for weeks. |
 | `backup.SpecBuilder` / `backup.TierPicker` / `(*BackupRunner).PickSettledRestoreCandidateOn` | internal/backup/schedule.go, runner.go | `func(ctx,archive) RestoreTestSpec`; `func(ctx,target,notAfter) (archive,landed,error)` | The per-run restore-test spec + per-tier **settled** candidate lookup (R-85, widened by R-86) | The spec is built **PER RUN**, never frozen at construction — the pre-R-85 immediately-invoked value made the offsite tier unschedulable AND went stale on any config change. `SourceTier` comes from **the archive**, never the configured target (the v0.100.0 rule). A tier with no archive returns `("", zero, nil)` — **`""` is NOT an error**, or every fresh box looks broken for its first week. **R-86: `notAfter` is the settle cutoff** (zero = no cutoff, which is what keeps `PickRestoreCandidateOn` a one-line call into it), and the picker now skips entries failing `archivePlausiblyComplete` — under per-archive due-ness an incomplete phantom would be picked forever, fail forever, never earn proof, and make the tier due at EVERY evaluation. |
 | `localapi.BackupTier` + `normalizeBackupTiers` / `config.BackupConfig.BackupTiers` | internal/localapi/backup_tiers.go, internal/config/config.go | `normalizeBackupTiers(tiers, legacy, cadence) []BackupTier`; `BackupTiers() ([]BackupTier, []string)` | THE R-82 multi-tier resolution — one runner per tier, primary first | **The untargeted local-API contract is FROZEN**: no `?target=` ⇒ primary tier ⇒ pre-R-82 response BYTES (Target is `omitempty` and stays empty). Never default a missing cadence — reject it and log the warning at ERROR. Never share one retention knob between tiers. Jobs are keyed by (vmid,target). |

cmd/felhom-agent/fstrim_wiring_test.go

@@ -0,0 +1,53 @@
+package main
+
+import (
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"testing"
+)
+
+// R-444: the weekly trim has the guestnet shape (component + reporter seam + goroutine), so its wiring is asserted
+// from the AST like TestMainWiresGuestNetWatchdog — a unit-green trim job that main.go never starts is the inert-seam
+// defect. It must also share the ONE heavy-op gate (heavyOps), or it could run beside a backup.
+func TestMainWiresGuestDiskTrim(t *testing.T) {
+	fset := token.NewFileSet()
+	f, err := parser.ParseFile(fset, "main.go", nil, 0)
+	if err != nil {
+		t.Fatalf("parse main.go: %v", err)
+	}
+	var constructedWithGate, reporterWired, started bool
+	ast.Inspect(f, func(n ast.Node) bool {
+		switch node := n.(type) {
+		case *ast.CallExpr:
+			if fn, ok := node.Fun.(*ast.SelectorExpr); ok {
+				switch fn.Sel.Name {
+				case "New":
+					if pkg, ok := fn.X.(*ast.Ident); ok && pkg.Name == "fstrim" && len(node.Args) >= 3 {
+						if id, ok := node.Args[2].(*ast.Ident); ok && id.Name == "heavyOps" {
+							constructedWithGate = true
+						}
+					}
+				case "SetGuestDiskTrimReporter":
+					reporterWired = true
+				}
+			}
+		case *ast.GoStmt:
+			if sel, ok := node.Call.Fun.(*ast.SelectorExpr); ok && sel.Sel.Name == "Run" {
+				if id, ok := sel.X.(*ast.Ident); ok && id.Name == "diskTrim" {
+					started = true
+				}
+			}
+		}
+		return true
+	})
+	if !constructedWithGate {
+		t.Error("main.go never calls fstrim.New(..., heavyOps, ...) — no trim job, or one outside the heavy-op gate")
+	}
+	if !reporterWired {
+		t.Error("main.go never calls collector.SetGuestDiskTrimReporter — the guest_disk_trim stanza never reaches the hub")
+	}
+	if !started {
+		t.Error("main.go never starts the trim job with `go diskTrim.Run(ctx)`")
+	}
+}

cmd/felhom-agent/main.go

@@ -38,6 +38,7 @@ import (
 	"gitea.dooplex.hu/admin/felhom-agent/internal/escrow"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/fasttick"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/felhomsshd"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/fstrim"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/guesthook"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/guestnet"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
@@ -1479,6 +1480,25 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 		}
 		go runJanitor(ctx, jd)
 	}
+	// R-444 (`09` §3 decision 139): the weekly guest disk trim — `pct fstrim <vmid>` of every owned, running guest,
+	// Wednesday from 10:00 local, daytime only, under the one-heavy-op gate. Not part of the errc fan-out: a trim job
+	// must never be able to bring the agent down.
+	if cfg.DiskTrim.Enabled() {
+		dtMode := proxmox.RunnerMode(cfg.Privileged.Mode)
+		if dtMode == "" {
+			dtMode = proxmox.RunnerSudo
+		}
+		dtRunner := &proxmox.ExecRunner{Mode: dtMode, SudoPath: cfg.Privileged.SudoPath}
+		dtGuests := localapi.NewStaleLockController(px, dtRunner, reconcile.DefaultPool, logger)
+		if dtGuests != nil {
+			diskTrim := fstrim.New(dtRunner, dtGuests, heavyOps,
+				filepath.Join(cfg.OOB.WithDefaults().StateDir, "guest-disk-trim.json"), logger)
+			collector.SetGuestDiskTrimReporter(diskTrim)
+			go diskTrim.Run(ctx)
+		}
+	} else {
+		logger.Info("fstrim: weekly guest disk trim disabled by config (disk_trim.disable)")
+	}
 	if lanLoop != nil {
 		lanServers = 1
 		go func() { errc <- lanLoop.Run(ctx) }()
@@ -1881,12 +1901,15 @@ func buildLocalAPIServer(cfg config.Config, px *proxmox.Client, store *backup.St
 		BackupTiers:    apiTiers, // R-82: primary first; untargeted endpoints act on the primary
 		InFlight:       inFlight, // R-85: shared with the restore-test scheduler (Scenario F)
 		Store:          store,
-		Storage:        observer,
+		// R-894: the newest success per tier on disk — the due-check's fallback when the storage cannot be
-		DriveTargets:   driveTargets,                    // Impl-2a: registry+units drives for the /disks view (union w/ Observe storages)
+		// read right after a restart. Same state dir as restore-test-state.json.
-		Smart:          storage.NewSmartReader(hostOps), // v0.95.0 Fix B: SMART for the union-path drives
+		LastKnownBackups: backup.NewBackupSuccessState(filepath.Join(cfg.OOB.WithDefaults().StateDir, "backup-success-state.json")),
-		HostReader:     storage.NewProcHostReader(),     // Impl-2b: durableIDForMount raw-mount fallback + role gate
+		Storage:          observer,
-		Tokens:         tokens,
+		DriveTargets:     driveTargets,                    // Impl-2a: registry+units drives for the /disks view (union w/ Observe storages)
-		BackupCadence:  cfg.Backup.BackupCadence(),
+		Smart:            storage.NewSmartReader(hostOps), // v0.95.0 Fix B: SMART for the union-path drives
+		HostReader:       storage.NewProcHostReader(),     // Impl-2b: durableIDForMount raw-mount fallback + role gate
+		Tokens:           tokens,
+		BackupCadence:    cfg.Backup.BackupCadence(),
 		// Disk management (slice 8C): the privileged host surface + the data-bearing wipe gate.
 		Disks:       hostOps,
 		DiskGate:    storageGateAdapter{gate: gate, hostID: cfg.Hub.HostID},

cmd/felhom-agent/r894_wiring_test.go

@@ -0,0 +1,74 @@
+package main
+
+import (
+	"go/ast"
+	"testing"
+)
+
+// R-894 — the on-disk backup record is WIRED on the daemon path (the built-but-never-wired class).
+// main → runDaemon → buildLocalAPIServer, and inside it the localapi.Options literal carries
+// LastKnownBackups built by backup.NewBackupSuccessState. An AST walk, not a string match, for the
+// reasons in escrow_recover_wiring_test.go.
+//
+// COMPANION RED-PROOF (observed): delete the `LastKnownBackups:` line from buildLocalAPIServer → this
+// fails with "localapi.Options in buildLocalAPIServer has no LastKnownBackups field". Restored.
+func TestR894_LastKnownBackupsIsWiredIntoTheDaemon(t *testing.T) {
+	_, f := parseMain(t)
+	if !callsWithin(f, "main")["runDaemon"] || !callsWithin(f, "runDaemon")["buildLocalAPIServer"] {
+		t.Fatal("main → runDaemon → buildLocalAPIServer is broken — the path this test asserts is not the live one")
+	}
+	var field, built bool
+	for _, d := range f.Decls {
+		fd, ok := d.(*ast.FuncDecl)
+		if !ok || fd.Name == nil || fd.Name.Name != "buildLocalAPIServer" || fd.Body == nil {
+			continue
+		}
+		ast.Inspect(fd.Body, func(n ast.Node) bool {
+			cl, ok := n.(*ast.CompositeLit)
+			if !ok {
+				return true
+			}
+			sel, ok := cl.Type.(*ast.SelectorExpr)
+			if !ok {
+				return true
+			}
+			if pkg, _ := sel.X.(*ast.Ident); pkg == nil || pkg.Name+"."+sel.Sel.Name != "localapi.Options" {
+				return true
+			}
+			for _, el := range cl.Elts {
+				kv, ok := el.(*ast.KeyValueExpr)
+				if !ok {
+					continue
+				}
+				if k, ok := kv.Key.(*ast.Ident); ok && k.Name == "LastKnownBackups" {
+					field = true
+					if callsIn(kv.Value)["backup.NewBackupSuccessState"] {
+						built = true
+					}
+				}
+			}
+			return true
+		})
+	}
+	if !field {
+		t.Fatal("localapi.Options in buildLocalAPIServer has no LastKnownBackups field")
+	}
+	if !built {
+		t.Fatal("LastKnownBackups is not built by backup.NewBackupSuccessState")
+	}
+}
+
+func callsIn(n ast.Node) map[string]bool {
+	out := map[string]bool{}
+	ast.Inspect(n, func(n ast.Node) bool {
+		if ce, ok := n.(*ast.CallExpr); ok {
+			if fn, ok := ce.Fun.(*ast.SelectorExpr); ok {
+				if x, ok := fn.X.(*ast.Ident); ok {
+					out[x.Name+"."+fn.Sel.Name] = true
+				}
+			}
+		}
+		return true
+	})
+	return out
+}

configs/felhom-agent.sudoers

@@ -129,6 +129,14 @@ Cmnd_Alias FELHOM_CONTROLLERSWAP = \
 Cmnd_Alias FELHOM_STALELOCK = \
   /usr/sbin/pct ^unlock [0-9]+$
 
+# Weekly guest disk trim (R-444, operator ruling `09` §3 decision 139). A thin pool only ever grows from blocks the
+# guest has FREED: `fstrim` inside the unprivileged container is refused (FITRIM: Operation not permitted), so the host
+# trims the guest's mounts. Measured on demo-hp 2026-10-06: `pct fstrim 9201` rc 0 in 24.4 s, pool 65.53 % -> 33.40 %,
+# apps kept answering. ONE exact pattern: a vmid and nothing else — no `--ignore-mountpoints`, no second argument
+# (pinned: TestSudoersFstrimRuleIsExact). The agent runs it on a weekly daytime timer under the heavy-op gate.
+Cmnd_Alias FELHOM_FSTRIM = \
+  /usr/sbin/pct ^fstrim [0-9]+$
+
 # Restore-test scratch teardown (F-LEAK, Campaign 8, v0.110.0). A restore-test whose restore FAILS
 # leaves a scratch guest the API token CANNOT destroy: `FelhomAgentGuest` is granted at /pool/felhom and
 # a guest joins that pool only when its restore COMPLETES, so a failed restore leaves a pool-less guest
@@ -304,4 +312,4 @@ Cmnd_Alias FELHOM_GUESTNET = \
   /usr/sbin/pct ^exec [0-9]+ -- pgrep -x dhclient$, \
   /usr/sbin/pct ^exec [0-9]+ -- dhclient -pf /run/dhclient\.eth0\.pid -lf /var/lib/dhcp/dhclient\.eth0\.leases eth0$
 
-felhom-agent ALL=(root) NOPASSWD: FELHOM_MOUNT, FELHOM_DISK, FELHOM_PROVISION, FELHOM_FORMAT, FELHOM_DNSMASQ, FELHOM_GUESTHOOK, FELHOM_INTERMEDIARY, FELHOM_CONTROLLERSWAP, FELHOM_STALELOCK, FELHOM_NETMOUNT, FELHOM_WG, FELHOM_SELFUPDATE, FELHOM_SSHD, FELHOM_OOB, FELHOM_PBSDR, FELHOM_BACKUPTARGET, FELHOM_SELFHEAL, FELHOM_ESCROW, FELHOM_GUESTNET, FELHOM_SCRATCH_TEARDOWN, FELHOM_OSAPPLY
+felhom-agent ALL=(root) NOPASSWD: FELHOM_MOUNT, FELHOM_DISK, FELHOM_PROVISION, FELHOM_FORMAT, FELHOM_DNSMASQ, FELHOM_GUESTHOOK, FELHOM_INTERMEDIARY, FELHOM_CONTROLLERSWAP, FELHOM_STALELOCK, FELHOM_FSTRIM, FELHOM_NETMOUNT, FELHOM_WG, FELHOM_SELFUPDATE, FELHOM_SSHD, FELHOM_OOB, FELHOM_PBSDR, FELHOM_BACKUPTARGET, FELHOM_SELFHEAL, FELHOM_ESCROW, FELHOM_GUESTNET, FELHOM_SCRATCH_TEARDOWN, FELHOM_OSAPPLY

configs/felhom-os-apply

@@ -32,6 +32,14 @@
 #               held packages, kernel taint, the crash guard; guest Debian, Docker engine, containerd, live-restore.
 #   live-restore-on  (v0.142.0, layer guest) the ONE-TIME step of `09` decision 87: merge `"live-restore": true`
 #               into the guest's /etc/docker/daemon.json and `systemctl reload docker`. NEVER a restart (R-835).
+#   oom-check   (R-528, `09` decision 157; layer docker, lane slow) ONLY the memory-kill check below: no apt, no engine
+#               change, no authority needed. Wrapper-only — the agent never writes this plan; run it by hand as root.
+# R-528 (`09` decision 157): a docker-layer APPLY also runs `oom_check()` after health_after and reports it as
+#   "oom_check": {"result": "pass"|"fail"|"error", "oom_killed": bool, "oom_event": bool, "exit_code": int|null,
+#                 "image": str|null, "detail": str}
+#   — one throwaway container (the controller's own image, no network / volume / port, 64 MB cap) is made to exceed its
+#   memory; "pass" only when the engine says OOMKilled=true AND emits the `oom` event. It never changes the step's
+#   outcome or health: the hub decides whether the engine set can be approved. Pinned by the OOMCheck tests.
 # Output: log lines on stderr and the journal (tag felhom-os-apply); the LAST stdout line is
 #   OSAPPLY-REPORT <one JSON object>
 # which is what the agent parses. Exit 0 = done; 2 = refused (nothing changed); 3 = failed during install.
@@ -68,6 +76,16 @@ JOURNAL_MARK = "@@FELHOM-DPKG-JOURNAL@@"
 DPKG_STATE_SCRIPT = "dpkg --audit; echo " + JOURNAL_MARK + "; ls -A /var/lib/dpkg/updates 2>/dev/null; true"
 # The installer's ROOT-OWNED record (felhom-host-install.sh `state_set mode`); the agent cannot write it.
 INSTALL_STATE = "/var/lib/felhom-install/state.json"
+# R-528: the memory-kill check (oom_check). One 200 MB block under a 64 MB cap: measured on Docker 29.8.2 to be
+# OOM-killed with OOMKilled=true and an `oom` event. Every call is bounded: timeouts (the clock read twice) + the
+# settle wait stay within 90 s.
+OOMCHECK_PREFIX = "felhom-oomcheck-"
+OOMCHECK_SCRIPT = "dd if=/dev/zero of=/dev/null bs=200M count=1"
+OOMCHECK_TIMEOUTS = {"image": 10, "clock": 5, "run": 30, "inspect": 10, "events": 10, "rm": 15}
+# Measured on demo-hp 9201 (Docker 29.8.2, 2026-10-06, audits/readback-2026-10-07/F/F1, F2): an `--until` taken right
+# after the run MISSED the oom event although OOMKilled=true; after a 2 s wait and `--until` = guest epoch + 1 it is
+# seen. Pinned by test_events_window_ends_after_the_settle_wait.
+OOMCHECK_SETTLE = 2
 # Kernel, boot and firmware packages are the SLOW lane on the host whatever their origin (`11` C3, §5.2): a host
 # reboot is needed for them to take effect, and a bad one can stop the box from booting.
 HOST_SLOW_RE = re.compile(r"^(linux-(image|headers|kbuild|modules|base)|proxmox-kernel|proxmox-default-kernel|pve-kernel|"
@@ -405,7 +423,7 @@ class Apply:
 
     def check_plan(self, plan):
         mode = plan.get("mode", "apply")
-        if mode not in ("apply", "inventory", "health", "facts", "live-restore-on", "bundle", "agent_update"):
+        if mode not in ("apply", "inventory", "health", "facts", "live-restore-on", "bundle", "agent_update", "oom-check"):
             raise Refused("R11", f"unknown mode {mode!r}")
         if mode == "agent_update":
             if plan.get("layer") != "host":
@@ -427,6 +445,8 @@ class Apply:
             raise Refused("R11", "facts is a host-layer mode (it reads the host and the guest)")
         if mode == "live-restore-on" and layer != "guest":
             raise Refused("R11", "live-restore-on is a guest-layer mode")
+        if mode == "oom-check" and layer != "docker":
+            raise Refused("R11", "oom-check is a docker-layer mode (it checks the guest's Docker engine)")
         if plan.get("undo") and layer != "docker":
             raise Refused("R5", "an undo (downgrade) exists only for the Docker layer, inside a signed job")
         vmid = plan.get("vmid")
@@ -951,6 +971,10 @@ class Apply:
         log = self.r.log
         if self.mode == "live-restore-on":
             return self.live_restore_on()
+        if self.mode == "oom-check":
+            # R-528: the check alone — no apt, no engine change; check_guest above still applies.
+            self.report["oom_check"] = self.oom_check()
+            return 0
         self.who, self.allow_downgrade = ("fast", False)
         if self.layer == "docker" and self.mode == "apply":
             self.who, self.allow_downgrade = self.docker_authority(plan)
@@ -990,8 +1014,87 @@ class Apply:
             self.report["docker_engine"] = out_v.strip() if rc_v == 0 and out_v.strip() else "unknown"
         self.report["reboot_scanned"] = "reboot_needed" in self.report
         self.report["health_after"] = self.health()
+        if self.layer == "docker" and self.mode == "apply":
+            # R-528 (`09` decision 157): does the engine report a memory kill? Reported only — never the outcome.
+            self.report["oom_check"] = self.oom_check()
         return 0
 
+    # ---------- R-528: the memory-kill check ----------
+    def oom_check(self):
+        """Run one throwaway container over its memory cap in the guest and read what the engine says about it.
+        Never raises: any failure becomes result "error". The container is ALWAYS removed (finally), and a failed
+        removal is named in the detail."""
+        t = OOMCHECK_TIMEOUTS
+        res = {"result": "error", "oom_killed": False, "oom_event": False, "exit_code": None, "image": None, "detail": ""}
+        log = self.r.log
+        try:
+            rc, out, err = self.g(["docker", "inspect", "-f", "{{.Config.Image}}", "felhom-controller"], timeout=t["image"])
+        except Exception as e:
+            rc, out, err = -1, "", str(e)
+        img = out.strip() if rc == 0 else ""
+        if not img or any(c.isspace() for c in img):
+            res["detail"] = f"the controller's image could not be read (rc={rc}): {(err or out).strip()[:200]}"
+            log(f"os-apply: OOM-CHECK error — {res['detail']}")
+            return res
+        res["image"] = img
+        name = f"{OOMCHECK_PREFIX}{os.getpid()}-{os.urandom(4).hex()}"
+        notes = []
+        try:
+            t0 = self.guest_epoch()
+            if t0 is None:
+                raise RuntimeError("the guest clock could not be read")
+            rrc, rout, rerr = self.g(["docker", "run", "--name", name, "--pull", "never", "--network", "none",
+                                      "--memory", "64m", "--memory-swap", "64m", "--label", "felhom.oomcheck=1",
+                                      "--entrypoint", "sh", img, "-c", OOMCHECK_SCRIPT], timeout=t["run"])
+            self.r.sleep(OOMCHECK_SETTLE)  # the engine publishes the oom event a moment after the run returns (F1/F2)
+            t1 = self.guest_epoch()
+            if t1 is None:
+                t1 = t0 + t["run"] + OOMCHECK_SETTLE + 1
+            irc, iout, ierr = self.g(["docker", "inspect", "-f", "{{.State.OOMKilled}} {{.State.ExitCode}}", name], timeout=t["inspect"])
+            if irc != 0:
+                raise RuntimeError(f"the check container could not be inspected (run rc={rrc}: {(rerr or rout).strip()[:120]}; "
+                                   f"inspect rc={irc}: {(ierr or iout).strip()[:120]})")
+            f = iout.split()
+            res["oom_killed"] = bool(f) and f[0] == "true"
+            try:
+                res["exit_code"] = int(f[1]) if len(f) > 1 else None
+            except ValueError:
+                res["exit_code"] = None
+            erc, eout, eerr = self.g(["docker", "events", "--since", str(t0 - 1), "--until", str(t1 + 1),
+                                      "--filter", f"container={name}", "--filter", "event=oom",
+                                      "--format", "{{.Action}}"], timeout=t["events"])
+            if erc != 0:
+                notes.append(f"the event read failed (rc={erc}): {(eerr or eout).strip()[:120]}")
+            res["oom_event"] = erc == 0 and any(l.strip() == "oom" for l in eout.splitlines())
+            if res["oom_killed"] and res["oom_event"]:
+                res["result"] = "pass"
+                notes.insert(0, "the engine reported the memory kill: OOMKilled=true and the oom event")
+            else:
+                res["result"] = "fail"
+                miss = [w for w, ok in (("OOMKilled=true", res["oom_killed"]), ("the oom event", res["oom_event"])) if not ok]
+                notes.insert(0, f"the engine did not report the memory kill: missing {' and '.join(miss)} (exit code {res['exit_code']})")
+        except Exception as e:
+            res["result"] = "error"
+            notes.insert(0, f"the check could not finish: {type(e).__name__}: {str(e)[:200]}")
+        finally:
+            try:
+                mrc, mout, merr = self.g(["docker", "rm", "-f", name], timeout=t["rm"])
+                if mrc != 0 and "no such container" not in (merr + mout).lower():
+                    notes.append(f"the check container {name} could not be removed (rc={mrc}): {(merr or mout).strip()[:120]}")
+            except Exception as e:
+                notes.append(f"the check container {name} could not be removed: {type(e).__name__}: {str(e)[:120]}")
+        res["detail"] = "; ".join(notes)
+        log(f"os-apply: OOM-CHECK result={res['result']} oom_killed={res['oom_killed']} oom_event={res['oom_event']} "
+            f"exit={res['exit_code']} image={img} — {res['detail']}")
+        return res
+
+    def guest_epoch(self):
+        rc, out, _ = self.g(["date", "+%s"], timeout=OOMCHECK_TIMEOUTS["clock"])
+        try:
+            return int(out.strip()) if rc == 0 else None
+        except ValueError:
+            return None
+
     def dpkg_state(self):
         """`dpkg --audit` AND dpkg's update journal, in ONE call (R-876, agent v0.145.0). A crash in the middle of an
         install can leave `/var/lib/dpkg/updates/` non-empty while `--audit` reads clean — measured on demo-hp

configs/test_felhom_config_bundle.py

@@ -546,7 +546,10 @@ class Builder(unittest.TestCase):
                                r"/etc/felhom/[a-z.-]+)", text))
         agent_writes = {"/usr/local/sbin/felhom-shared-parent", "/etc/systemd/system/felhom-shared-parent.service"}
         trust = {osapply.TRUST_FILE, osapply.TRUST_SIGNERS, osapply.TRUST_SIGNERS + ".tmp", osapply.BUNDLE_RECORD}
-        missing = sorted(p for p in found if p not in osapply.BUNDLE_DESTS and p not in agent_writes | trust)
+        # Written by the appliance ISO's first boot (felhom.eu scripts/iso/felhom-bootstrap.sh), never by the installer:
+        # since installer 1.32.0 (R-275) the uninstall only NAMES them under KEPT.
+        iso_writes = {"/etc/felhom/.bootstrap-done", "/etc/felhom/appliance-pairing-code"}
+        missing = sorted(p for p in found if p not in osapply.BUNDLE_DESTS and p not in agent_writes | trust | iso_writes)
         self.assertEqual(missing, [], "the installer writes these root files, but the bundle does not carry them")
         # the limits drop-in is named through $AGENT_UNIT in the installer
         self.assertIn("/etc/systemd/system/felhom-agent.service.d/felhom-agent-limits.conf", osapply.BUNDLE_DESTS)

configs/test_felhom_os_apply.py

@@ -83,7 +83,7 @@ class Fake:
         return self.clock
 
     def sleep(self, s):
-        pass
+        self.sleeps = getattr(self, "sleeps", []) + [(len(self.calls), s)]  # (calls made before it, seconds)
 
     def verify_sig(self, signers, key_id, ns, blob, sig):
         self.verified = (signers, key_id, ns, blob, sig)
@@ -196,6 +196,27 @@ class Fake:
             return 0, self.engine + "\n", ""
         if cmd == "docker" and a[1:3] == ["ps", "-q"]:
             return 0, "".join(i + "\n" for i in self.ids), ""
+        # R-528: the memory-kill check's engine (oom_image None = unreadable; oom_state / oom_event the engine's answer)
+        if cmd == "date" and a[1:] == ["+%s"]:
+            # each read is 3 s later than the last, so the order of the reads is visible in the values
+            self.oom_epochs = getattr(self, "oom_epochs", []) + [int(self.clock) + 3 * len(getattr(self, "oom_epochs", []))]
+            return 0, f"{self.oom_epochs[-1]}\n", ""
+        if cmd == "docker" and a[1:4] == ["inspect", "-f", "{{.Config.Image}}"]:
+            img = getattr(self, "oom_image", "gitea.dooplex.hu/admin/felhom-controller:0.300.0")
+            return (0, img + "\n", "") if img is not None else (1, "", "Error: No such object: felhom-controller")
+        if cmd == "docker" and a[1] == "run":
+            self.oom_runs = getattr(self, "oom_runs", []) + [a]
+            return 137, "", ""
+        if cmd == "docker" and a[1:4] == ["inspect", "-f", "{{.State.OOMKilled}} {{.State.ExitCode}}"]:
+            if getattr(self, "oom_inspect_raises", False):
+                raise subprocess.TimeoutExpired(a, 10)
+            return 0, getattr(self, "oom_state", "true 137") + "\n", ""
+        if cmd == "docker" and a[1] == "events":
+            self.oom_events_argv = a
+            return 0, ("oom\n" if getattr(self, "oom_event", True) else ""), ""
+        if cmd == "docker" and a[1:3] == ["rm", "-f"]:
+            self.oom_removed = getattr(self, "oom_removed", []) + a[3:]
+            return 0, a[3] + "\n", ""
         if cmd == "docker" and a[1] == "inspect":
             mounts = {"aaa111": "/felhom-controller|/var/run/docker.sock;/app/data;", "bbb222": "/app|/data;"}
             return 0, mounts.get(a[-1], "/other|;") + "\n", ""
@@ -997,8 +1018,6 @@ class RealSignatureCheck(unittest.TestCase):
         self.assertNotEqual(r.verify_sig(self.signers, "someone-else", "felhom-op-v1", blob, sig), 0)
         self.assertNotEqual(r.verify_sig(self.signers, "felhom-op-1", "felhom-op-v1", blob, self.sign(blob, ns="other-ns")), 0)
 
-if __name__ == "__main__":
-    unittest.main()
 
 
 class UnsentReport(unittest.TestCase):
@@ -1179,3 +1198,131 @@ class CrashLeftTheJournal(unittest.TestCase):
         self.assertEqual(rc, 0, rep)
         self.assertTrue(any("INTERRUPTED" in l for l in f.logs), f.logs)
         self.assertTrue(any(l.startswith("os-apply: REPAIR ") and l.endswith("forced") for l in f.logs), f.logs)
+
+
+class OOMCheck(unittest.TestCase):
+    """R-528 (`09` decision 157): after a Docker engine step the wrapper proves the engine reports a memory kill
+    (OOMKilled=true AND the `oom` event). Reported only; the hub decides. Red-proofs: audits/readback-2026-10-07/F/."""
+
+    def apply(self, **kw):
+        f = docker_fake(signed=signed_job())
+        for k, v in kw.items():
+            setattr(f, k, v)
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        return f, rep
+
+    def test_pass_when_oomkilled_and_the_event(self):
+        f, rep = self.apply()
+        oc = rep["oom_check"]
+        self.assertEqual(oc["result"], "pass", oc)
+        self.assertEqual((oc["oom_killed"], oc["oom_event"], oc["exit_code"]), (True, True, 137))
+        self.assertEqual(oc["image"], "gitea.dooplex.hu/admin/felhom-controller:0.300.0")
+        self.assertEqual(sorted(oc), ["detail", "exit_code", "image", "oom_event", "oom_killed", "result"])
+        run_argv = f.oom_runs[0]
+        name = run_argv[run_argv.index("--name") + 1]
+        self.assertTrue(re.match(r"^felhom-oomcheck-[0-9]+-[0-9a-f]{8}$", name), name)
+        for flag, val in (("--pull", "never"), ("--network", "none"), ("--memory", "64m"), ("--memory-swap", "64m"),
+                          ("--label", "felhom.oomcheck=1"), ("--entrypoint", "sh")):
+            self.assertEqual(run_argv[run_argv.index(flag) + 1], val, flag)
+        self.assertNotIn("-v", run_argv)
+        self.assertNotIn("-p", run_argv)
+        self.assertEqual(run_argv[-3:], ["gitea.dooplex.hu/admin/felhom-controller:0.300.0", "-c", osapply.OOMCHECK_SCRIPT])
+        self.assertIn(f"container={name}", f.oom_events_argv)
+        self.assertIn("event=oom", f.oom_events_argv)
+        self.assertEqual(f.oom_removed, [name], "the check container must be removed")
+        self.assertTrue(rep["health_after"], "health is read before the check")
+        # bounded: every call has a timeout and the clock is read twice — the worst case stays within 90 s
+        self.assertLessEqual(sum(osapply.OOMCHECK_TIMEOUTS.values()) + osapply.OOMCHECK_TIMEOUTS["clock"]
+                             + osapply.OOMCHECK_SETTLE, 90)
+
+    def test_events_window_ends_after_the_settle_wait(self):
+        # Measured (F1/F2): an --until taken right after the run missed the oom event. The window must end after a
+        # wait of >= 2 s that comes AFTER the run, at the guest epoch read after that wait, + 1.
+        f, rep = self.apply()
+        run_i = next(i for i, c in enumerate(f.calls) if c[-1][:2] == ["docker", "run"])
+        date_i = [i for i, c in enumerate(f.calls) if c[-1] == ["date", "+%s"]]
+        waits = [(i, s) for i, s in getattr(f, "sleeps", []) if i > run_i]
+        self.assertTrue(waits and waits[0][1] >= 2, f"no settle wait after the run: {getattr(f, 'sleeps', None)}")
+        self.assertTrue(date_i[-1] >= waits[0][0], "the end epoch must be read after the wait")
+        ev = f.oom_events_argv
+        since, until = int(ev[ev.index("--since") + 1]), int(ev[ev.index("--until") + 1])
+        self.assertEqual(until, f.oom_epochs[-1] + 1, "until = the guest epoch read after the wait, + 1")
+        self.assertGreater(until, f.oom_epochs[0] + 1)
+        self.assertLess(since, f.oom_epochs[0] + 1)
+
+    def test_oomkilled_false_is_fail(self):
+        f, rep = self.apply(oom_state="false 0")
+        oc = rep["oom_check"]
+        self.assertEqual(oc["result"], "fail", oc)
+        self.assertIn("OOMKilled=true", oc["detail"])
+        self.assertTrue(oc["oom_event"])
+        self.assertEqual(len(f.oom_removed), 1)
+
+    def test_no_event_is_fail(self):
+        f, rep = self.apply(oom_event=False)
+        oc = rep["oom_check"]
+        self.assertEqual(oc["result"], "fail", oc)
+        self.assertIn("the oom event", oc["detail"])
+        self.assertTrue(oc["oom_killed"])
+
+    def test_image_unreadable_is_error(self):
+        f, rep = self.apply(oom_image=None)
+        oc = rep["oom_check"]
+        self.assertEqual(oc["result"], "error", oc)
+        self.assertIsNone(oc["image"])
+        self.assertIn("image could not be read", oc["detail"])
+        self.assertFalse(hasattr(f, "oom_runs"), "no container is started without an image")
+
+    def test_container_removed_even_when_inspect_raises(self):
+        f, rep = self.apply(oom_inspect_raises=True)
+        oc = rep["oom_check"]
+        self.assertEqual(oc["result"], "error", oc)
+        self.assertEqual(len(f.oom_removed), 1, "the container must be removed even when inspect raised")
+        self.assertTrue(f.oom_removed[0].startswith(osapply.OOMCHECK_PREFIX))
+        self.assertNotIn("failed", rep, "the check never turns the step into a failure")
+
+    def test_never_on_guest_or_host_or_in_health_mode(self):
+        g = Fake()
+        rc, rep = run(g)
+        self.assertEqual(rc, 0, rep)
+        h = Fake()
+        h.plan["layer"] = "host"
+        h.plan["packages"] = [{"name": "bash", "version": "5.2.37-2+b10", "origin": "Debian"}]
+        h.installed["bash"] = "5.2.37-2+b9"
+        h.live["bash"] = {"5.2.37-2+b10"}
+        rc_h, rep_h = run(h)
+        self.assertEqual(rc_h, 0, rep_h)
+        d = docker_fake(signed=signed_job())
+        d.plan["mode"] = "health"
+        rc_d, rep_d = run(d)
+        self.assertEqual(rc_d, 0, rep_d)
+        for f, r in ((g, rep), (h, rep_h), (d, rep_d)):
+            self.assertNotIn("oom_check", r)
+            self.assertFalse(hasattr(f, "oom_runs"), r.get("layer"))
+            self.assertFalse(any(c[-1][:2] == ["docker", "run"] for c in f.calls))
+
+    def test_mode_oom_check_runs_only_the_check(self):
+        f = docker_fake()  # no authority: the check changes nothing, so it needs none
+        f.plan["mode"], f.plan["packages"] = "oom-check", []
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        self.assertEqual(rep["oom_check"]["result"], "pass", rep)
+        self.assertFalse(any("apt-get" in c[-1] or "dpkg-query" in c[-1] for c in f.calls), f.calls)
+        self.assertEqual(len(f.oom_removed), 1)
+
+    def test_mode_oom_check_keeps_the_refusals(self):
+        f = docker_fake()
+        f.plan["mode"], f.plan["packages"] = "oom-check", []
+        f.files["/etc/pve/lxc/9201.conf"] = "arch: amd64\n"
+        rc, rep = run(f)
+        self.assertEqual((rc, rep["refused"]["code"]), (2, "R10"), rep)
+        self.assertFalse(hasattr(f, "oom_runs"))
+        g = Fake()
+        g.plan["mode"] = "oom-check"
+        rc, rep = run(g)
+        self.assertEqual((rc, rep["refused"]["code"]), (2, "R11"), rep)
+
+
+if __name__ == "__main__":
+    unittest.main()

internal/backup/archive_completeness_test.go

@@ -241,3 +241,23 @@ func TestNewestArchiveTime_DistinctPhantomsEachAnnounced(t *testing.T) {
 		t.Errorf("got %d rejection lines for 2 distinct phantoms across 3 polls, want 2:\n%s", n, buf.String())
 	}
 }
+
+// R-99 (`09` §3 decision 140): the WARN for a PBS phantom ends with the cleanup runbook, so whoever sees it knows the
+// one sanctioned way to remove it; a tiny archive on a dir storage is not a PBS phantom and gets no pointer.
+// RED-PROOF: drop the `msg += phantomCleanupPointer` line → "the PBS phantom WARN does not end with the runbook pointer".
+func TestRejectedArchiveWarnNamesTheCleanupRunbook(t *testing.T) {
+	var buf bytes.Buffer
+	r := runnerWithContent(t, &buf, []proxmox.StorageContent{phantomEntry(), goodPBSEntry()})
+	if _, _, err := r.NewestArchiveTime(context.Background(), 9201); err != nil {
+		t.Fatal(err)
+	}
+	const want = "INCOMPLETE archive when computing tier freshness — it is not a successful backup — a phantom leftover; delete it by felhom.eu documentation/runbooks/pbs-phantom-cleanup.md (09 §3 decision 140)"
+	if !strings.Contains(buf.String(), want) {
+		t.Errorf("the PBS phantom WARN does not end with the runbook pointer:\n%s", buf.String())
+	}
+	local := phantomEntry()
+	local.Format, local.VolID = "tar.zst", "local:backup/vzdump-lxc-9201-2026_07_28-05_31_14.tar.zst"
+	if got := rejectedArchiveMessage(local); strings.Contains(got, "pbs-phantom-cleanup") {
+		t.Errorf("a dir-storage archive got the PBS runbook pointer: %s", got)
+	}
+}

internal/backup/backup_state.go

@@ -0,0 +1,131 @@
+package backup
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"sort"
+	"strconv"
+	"sync"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+)
+
+// BackupSuccessState persists the newest SUCCESSFUL whole-guest backup per tier and guest (R-894).
+//
+// Why it exists. The due-check (`localapi` handleBackupDue) asks the tier's storage when a backup last
+// landed (R-84) and falls back to the in-memory record when the storage cannot be read. The in-memory
+// record is empty after an agent restart (Store, R-348), so "storage unreadable" right after a restart
+// read as "no record — DUE". Measured 2026-10-05 on demo-hp: the agent restarted at 04:57, the off-site
+// storage answered "Can't connect" at 06:25, the 7-day tier — last copy 2026-10-01 — read DUE, the
+// controller asked, and vzdump failed. This file is the last known copy the fallback reads instead.
+//
+// It is read ONLY when the storage cannot be read. A storage that answers is the ground truth and wins,
+// in both directions: an archive found there counts, and an archive absent there is absent even when
+// this file remembers a success (a pruned or deleted archive must make the tier due — the same reason
+// R-84 chose the storage over a persisted record). Pinned by
+// TestBackupDue_R894_SavedCopyIgnoredWhenStorageAnswers.
+//
+// Only SUCCESSES are written (the RestoreTestState rule): a failure must stay due and be retried, so a
+// record of a failure has no reader.
+type BackupSuccessState struct {
+	path string
+	mu   sync.Mutex
+	last map[string]savedSuccess // key(target, vmid) → the newest success
+}
+
+type savedSuccess struct {
+	target string
+	vmid   int
+	at     time.Time
+}
+
+// backupSuccessJSON is one entry on disk.
+type backupSuccessJSON struct {
+	Target    string `json:"target"`
+	VMID      int    `json:"vmid"`
+	StartedAt string `json:"started_at"`
+}
+
+func backupStateKey(target string, vmid int) string { return target + "/" + strconv.Itoa(vmid) }
+
+// NewBackupSuccessState opens (or creates) the state at path. A missing or unreadable file degrades to
+// "nothing known" — the pre-R-894 behaviour, which is DUE — and never wedges the daemon.
+func NewBackupSuccessState(path string) *BackupSuccessState {
+	s := &BackupSuccessState{path: path, last: map[string]savedSuccess{}}
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return s
+	}
+	var entries []backupSuccessJSON
+	if json.Unmarshal(data, &entries) != nil {
+		return s
+	}
+	for _, e := range entries {
+		t, perr := time.Parse(time.RFC3339, e.StartedAt)
+		if perr != nil {
+			continue // one unreadable entry must not lose the others
+		}
+		s.last[backupStateKey(e.Target, e.VMID)] = savedSuccess{target: e.Target, vmid: e.VMID, at: t.UTC()}
+	}
+	return s
+}
+
+// RecordBackupSuccess saves b when it is a success newer than the one on file. target is the tier the
+// job ran on (the due-check's key); a failure or an unparseable time is ignored.
+func (s *BackupSuccessState) RecordBackupSuccess(target string, b hub.Backup) error {
+	if s == nil || !b.Success {
+		return nil
+	}
+	t, err := time.Parse(time.RFC3339, b.StartedAt)
+	if err != nil {
+		return nil
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	k := backupStateKey(target, b.VMID)
+	if old, ok := s.last[k]; ok && !t.After(old.at) {
+		return nil
+	}
+	s.last[k] = savedSuccess{target: target, vmid: b.VMID, at: t.UTC()}
+	return s.saveLocked()
+}
+
+// LastKnownSuccess returns the newest saved success for this tier and guest (ok=false = none on file).
+func (s *BackupSuccessState) LastKnownSuccess(target string, vmid int) (time.Time, bool) {
+	if s == nil {
+		return time.Time{}, false
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	e, ok := s.last[backupStateKey(target, vmid)]
+	return e.at, ok
+}
+
+func (s *BackupSuccessState) saveLocked() error {
+	entries := make([]backupSuccessJSON, 0, len(s.last))
+	for _, e := range s.last {
+		entries = append(entries, backupSuccessJSON{Target: e.target, VMID: e.vmid, StartedAt: e.at.Format(time.RFC3339)})
+	}
+	// Deterministic file content (Go's map order is random).
+	sort.Slice(entries, func(i, j int) bool {
+		if entries[i].Target != entries[j].Target {
+			return entries[i].Target < entries[j].Target
+		}
+		return entries[i].VMID < entries[j].VMID
+	})
+	data, err := json.MarshalIndent(entries, "", "  ")
+	if err != nil {
+		return err
+	}
+	if err := os.MkdirAll(filepath.Dir(s.path), 0o755); err != nil {
+		return err
+	}
+	tmp := s.path + ".tmp"
+	if err := os.WriteFile(tmp, data, 0o600); err != nil {
+		os.Remove(tmp)
+		return err
+	}
+	return os.Rename(tmp, s.path)
+}

internal/backup/backup_state_test.go

@@ -0,0 +1,59 @@
+package backup
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+)
+
+// R-894: the on-disk newest success per tier survives a restart (a new state from the same file).
+func TestBackupSuccessState_SurvivesRestart(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "backup-success-state.json")
+	s := NewBackupSuccessState(path)
+	at := time.Date(2026, 10, 1, 20, 15, 0, 0, time.UTC)
+	if err := s.RecordBackupSuccess("felhom-pbs", hub.Backup{VMID: 9201, Success: true, StartedAt: at.Format(time.RFC3339)}); err != nil {
+		t.Fatal(err)
+	}
+	got, ok := NewBackupSuccessState(path).LastKnownSuccess("felhom-pbs", 9201)
+	if !ok || !got.Equal(at) {
+		t.Fatalf("after a restart the saved copy must read back; got %v ok=%v", got, ok)
+	}
+	if _, ok := NewBackupSuccessState(path).LastKnownSuccess("local", 9201); ok {
+		t.Fatal("another tier must not borrow this tier's copy")
+	}
+}
+
+// Only a NEWER success replaces the saved one; failures and unparseable times are ignored.
+func TestBackupSuccessState_KeepsNewestSuccessOnly(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "s.json")
+	s := NewBackupSuccessState(path)
+	newer := time.Date(2026, 10, 5, 0, 0, 0, 0, time.UTC)
+	older := newer.Add(-48 * time.Hour)
+	for _, b := range []hub.Backup{
+		{VMID: 1, Success: true, StartedAt: newer.Format(time.RFC3339)},
+		{VMID: 1, Success: true, StartedAt: older.Format(time.RFC3339)},                 // older: ignored
+		{VMID: 1, Success: false, StartedAt: newer.Add(time.Hour).Format(time.RFC3339)}, // failure: ignored
+		{VMID: 1, Success: true, StartedAt: "not-a-time"},                               // unparseable: ignored
+	} {
+		if err := s.RecordBackupSuccess("t", b); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if got, _ := NewBackupSuccessState(path).LastKnownSuccess("t", 1); !got.Equal(newer) {
+		t.Fatalf("want the newest success %v, got %v", newer, got)
+	}
+}
+
+// A corrupt file degrades to "nothing known" (the pre-R-894 DUE answer), never a crash.
+func TestBackupSuccessState_CorruptFileIsEmpty(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "s.json")
+	if err := os.WriteFile(path, []byte("{not json"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if _, ok := NewBackupSuccessState(path).LastKnownSuccess("t", 1); ok {
+		t.Fatal("a corrupt file must read as nothing known")
+	}
+}

internal/backup/runner.go

@@ -518,10 +518,25 @@ func (r *BackupRunner) warnRejectedArchiveOnce(e proxmox.StorageContent, why str
 	if seen {
 		return
 	}
-	r.logger.Warn("backup: ignoring an INCOMPLETE archive when computing tier freshness — it is not a successful backup",
+	r.logger.Warn(rejectedArchiveMessage(e),
 		"target", r.target, "vmid", e.VMID, "volid", e.VolID, "size_bytes", e.Size, "reason", why)
 }
 
+// phantomCleanupPointer names the runbook that removes a PBS phantom (R-99, `09` §3 decision 140: a leftover of an
+// aborted upload is deleted on the backup server, by a runbook, when one is seen — never automatically).
+const phantomCleanupPointer = " — a phantom leftover; delete it by felhom.eu documentation/runbooks/pbs-phantom-cleanup.md (09 §3 decision 140)"
+
+// rejectedArchiveMessage is the WARN text for a rejected archive. Only a PBS entry (format pbs-ct / pbs-vm) gets the
+// cleanup pointer: the runbook deletes on a PBS datastore, and a tiny archive on a dir storage is not a PBS phantom.
+// Pinned by TestRejectedArchiveWarnNamesTheCleanupRunbook.
+func rejectedArchiveMessage(e proxmox.StorageContent) string {
+	msg := "backup: ignoring an INCOMPLETE archive when computing tier freshness — it is not a successful backup"
+	if strings.HasPrefix(e.Format, "pbs-") {
+		msg += phantomCleanupPointer
+	}
+	return msg
+}
+
 // demo-felhom in a single afternoon of deploys (2026-07-26).
 //
 // Asking the STORAGE rather than persisting the store is deliberate:

internal/backup/store.go

@@ -30,6 +30,9 @@ import (
 // 2026-08-20, two consecutive host-reports with `0 backups` while `pvesm list` showed archives on both tiers. What
 // is unaffected is the hub's VERDICT: it looks back 7 days over stored reports (felhom.eu hub/internal/monitor/
 // deadline.go backupEvidenceLookback) and the storage stays the ground truth (R-84).
+// The due-check's fallback for an UNREADABLE storage no longer reads this store alone (R-894): the newest
+// success per tier is also on disk (BackupSuccessState), so a restart followed by an unreachable storage
+// reads the last known copy, not "never".
 type Store struct {
 	mu       sync.Mutex
 	byTarget map[string]hub.Backup // latest backup per target id

internal/capability/manifest.go

@@ -151,6 +151,10 @@ var manifest = []Capability{
 	// reboot-during-backup lock can't start → the customer box stays DOWN until this clears it) ----
 	{"stalelock-unlock", "reboot-during-backup stale-lock recovery", "/usr/sbin/pct", []string{"unlock", "9201"}, true, ""},
 
+	// ---- Weekly guest disk trim (FELHOM_FSTRIM, R-444). NON-critical: a missing grant means the thin pool is not
+	// reclaimed this week (the trim job WARNs per guest and the report shows the failure), not a serving outage. ----
+	{"guest-fstrim", "weekly guest disk trim (thin-pool reclaim, R-444)", "/usr/sbin/pct", []string{"fstrim", "9201"}, false, ""},
+
 	// ---- Offsite WG tunnel (FELHOM_WG, S3/v0.64.0; Critical FLIPPED in S4/v0.66.0 — offsite
 	// backups now RIDE the tunnel, so a degraded tunnel capability is operator-alert-worthy: the
 	// conf install, unit enable/restart and the handshake read gate the backup path. apt-install

internal/capability/r861_injection_test.go

@@ -63,3 +63,33 @@ func TestSudoersRefusesTheR861Injections(t *testing.T) {
 		}
 	}
 }
+
+// R-444: the weekly trim's grant is ONE exact shape — `pct fstrim <vmid>` — and nothing smuggled after it.
+// The manifest entry (guest-fstrim) proves the real call is still allowed (TestManifestCoveredBySudoers); this
+// pins the other direction. RED-PROOF: write the rule as the glob `/usr/sbin/pct fstrim [0-9]*` → every decoy
+// below with a trailing argument matches (the glob's `*` eats spaces).
+func TestSudoersFstrimRuleIsExact(t *testing.T) {
+	data, err := os.ReadFile(sudoersPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	entries := parseSudoersEntries(t, string(data))
+	if !matchesAny("/usr/sbin/pct fstrim 9201", entries) {
+		t.Fatal("the sudoers does not allow `pct fstrim 9201` — the weekly trim cannot run")
+	}
+	for _, c := range []string{
+		"/usr/sbin/pct fstrim 9201 --ignore-mountpoints",
+		"/usr/sbin/pct fstrim 9201 --ignore-mountpoints 1",
+		"/usr/sbin/pct fstrim 9201; x",
+		"/usr/sbin/pct fstrim 9201 9202",
+		"/usr/sbin/pct fstrim 92a1",
+		"/usr/sbin/pct fstrim ",
+		"/usr/sbin/pct fstrim -- 9201",
+		"/usr/sbin/pct destroy 9201",
+		"/usr/sbin/pct destroy 9201 --purge",
+	} {
+		if matchesAny(c, entries) {
+			t.Errorf("the sudoers allows a command the trim rule must not: %q", c)
+		}
+	}
+}

internal/config/config.go

@@ -33,6 +33,7 @@ type Config struct {
 	LANResolver LANResolverConfig `json:"lan_resolver"`
 	WGTunnel    WGTunnelConfig    `json:"wg_tunnel"`
 	GuestNet    GuestNetConfig    `json:"guest_net"`
+	DiskTrim    DiskTrimConfig    `json:"disk_trim"`
 	OOB         OOBConfig         `json:"oob"`
 	SelfUpdate  SelfUpdateConfig  `json:"selfupdate"`
 	LogLevel    string            `json:"log_level"` // debug|info|warn|error (default info)
@@ -139,6 +140,16 @@ func (w WGTunnelConfig) WithDefaults() WGTunnelConfig {
 	return w
 }
 
+// DiskTrimConfig configures the R-444 weekly guest disk trim (internal/fstrim). DEFAULT-ON, like GuestNetConfig and for
+// the same reason: it only acts on guests the agent already owns, and the operator ruled every box trims (`09` §3
+// decision 139). Opting out is the explicit act: `"disk_trim": {"disable": true}`.
+type DiskTrimConfig struct {
+	Disable bool `json:"disable"`
+}
+
+// Enabled reports whether the weekly trim should run.
+func (d DiskTrimConfig) Enabled() bool { return !d.Disable }
+
 // GuestNetConfig configures the R-54 guest-network watchdog (internal/guestnet).
 //
 // **This is the repo's first DEFAULT-ON feature gate, and the inversion is deliberate.** Every other

internal/fstrim/fstrim.go

@@ -0,0 +1,346 @@
+// Package fstrim is the weekly guest disk trim (R-444, operator ruling `09` §3 decision 139).
+//
+// Why: a thin pool only ever grows from blocks a guest has already FREED — `fstrim` inside the unprivileged container
+// is refused (FITRIM: Operation not permitted), and nothing else on the box gives the blocks back. A full thin pool
+// takes every guest on the host read-only, so the pool can reach 100 % from deleted data alone. Measured on demo-hp
+// 2026-10-06 09:14Z: `pct fstrim 9201` rc 0 in 24.4 s, pool 65.53 % -> 33.40 %, 18/18 app probes 200, max 1.1 s
+// (audits/ten-answers-2026-10-06/r444-measure.txt).
+//
+// The rule, each part pinned by a test in fstrim_test.go:
+//   - Weekly: a guest is DUE from Wednesday 10:00 local until it has been trimmed once since then (a box that was off
+//     on Wednesday catches up at its next eligible hour).
+//   - Daytime only: a trim starts only between 10:00 and 20:59 local — never in the night window (01:00–06:59) where
+//     the backups and the restore-tests run (TestEligibleHourNeverInTheNight).
+//   - Never beside a backup, a restore-test or another heavy operation: the pass holds the host-wide one-heavy-op gate
+//     (backup.InFlight) for its whole run; a busy gate DEFERS the pass to the next hourly tick.
+//   - A failed trim is retried at the next eligible hour, at most MaxAttemptsPerWeek times in one week.
+//   - The last result per guest (time, bytes, ok/fail) is persisted, so a restart neither loses it nor re-trims.
+//
+// The command is the ONE exact sudoers shape `pct fstrim <vmid>` (FELHOM_FSTRIM). Only guests from the pool-verified
+// source (ListLXC ∩ the felhom pool, audit A1) and only RUNNING ones are trimmed.
+package fstrim
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"regexp"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/proxmox"
+)
+
+// Schedule. The weekday/hours are fixed on purpose (one sentence the operator can read on the System page).
+const (
+	Weekday            = time.Wednesday
+	StartHour          = 10 // first eligible local hour (inclusive)
+	EndHour            = 21 // first NOT-eligible local hour (exclusive): last start is 20:59
+	MaxAttemptsPerWeek = 3
+	// TickInterval is how often the job looks; a deferred or failed pass is therefore retried the next hour.
+	TickInterval = time.Hour
+	// FirstTickDelay lets the agent settle after a start before the first look.
+	FirstTickDelay = 5 * time.Minute
+	// PerGuestTimeout bounds one `pct fstrim` (measured 24.4 s for 84 GiB).
+	PerGuestTimeout = 30 * time.Minute
+)
+
+// ScheduleText is the human description carried on the host report.
+const ScheduleText = "weekly, due Wednesday from 10:00 host-local time; starts only 10:00-20:59; never beside a backup or restore-test"
+
+// Runner runs a host command (proxmox.ExecRunner in production, through `sudo -n`).
+type Runner interface {
+	Run(ctx context.Context, name string, args ...string) (stdout, stderr []byte, err error)
+}
+
+// GuestSource yields the guests this agent OWNS (the pool-verified source, never a bare ListLXC).
+type GuestSource interface {
+	Guests(ctx context.Context) ([]proxmox.Guest, error)
+}
+
+// Gate is the host-wide one-heavy-operation gate (*backup.InFlight).
+type Gate interface {
+	TryAcquire(what string) (release func(), busy string, ok bool)
+}
+
+// GateName is what the gate reports as busy while a trim runs.
+const GateName = "guest-fstrim"
+
+// Record is one guest's last trim attempt, as persisted.
+type Record struct {
+	LastAttemptAt   time.Time `json:"last_attempt_at"`
+	OK              bool      `json:"ok"`
+	BytesTrimmed    int64     `json:"bytes_trimmed"`
+	Mounts          int       `json:"mounts"`
+	DurationSeconds float64   `json:"duration_seconds"`
+	LastOKAt        time.Time `json:"last_ok_at,omitempty"`
+	Error           string    `json:"error,omitempty"`
+	// Attempts counts the attempts since the current week's due time (reset by the first attempt of a new week).
+	Attempts int `json:"attempts"`
+}
+
+// Trimmer is the weekly job.
+type Trimmer struct {
+	runner    Runner
+	guests    GuestSource
+	gate      Gate
+	statePath string
+	logger    *slog.Logger
+	loc       *time.Location
+	now       func() time.Time
+
+	mu      sync.Mutex
+	records map[int]Record
+}
+
+// New builds the job and loads the persisted state. A missing state file is an empty state; a corrupt one is logged
+// and treated as empty (the cost is one extra trim, never a missed one).
+func New(runner Runner, guests GuestSource, gate Gate, statePath string, logger *slog.Logger) *Trimmer {
+	if logger == nil {
+		logger = slog.Default()
+	}
+	t := &Trimmer{runner: runner, guests: guests, gate: gate, statePath: statePath, logger: logger,
+		loc: time.Local, now: time.Now, records: map[int]Record{}}
+	t.load()
+	return t
+}
+
+func (t *Trimmer) load() {
+	data, err := os.ReadFile(t.statePath)
+	if err != nil {
+		if !os.IsNotExist(err) {
+			t.logger.Warn("fstrim: state read failed — starting empty", "path", t.statePath, "err", err)
+		}
+		return
+	}
+	var raw map[string]Record
+	if err := json.Unmarshal(data, &raw); err != nil {
+		t.logger.Warn("fstrim: state file corrupt — starting empty", "path", t.statePath, "err", err)
+		return
+	}
+	for k, r := range raw {
+		if id, err := strconv.Atoi(k); err == nil && id > 0 {
+			t.records[id] = r
+		}
+	}
+}
+
+func (t *Trimmer) saveLocked() error {
+	raw := make(map[string]Record, len(t.records))
+	for id, r := range t.records {
+		raw[strconv.Itoa(id)] = r
+	}
+	data, err := json.MarshalIndent(raw, "", "  ")
+	if err != nil {
+		return err
+	}
+	if err := os.MkdirAll(filepath.Dir(t.statePath), 0o755); err != nil {
+		return err
+	}
+	tmp := t.statePath + ".tmp"
+	if err := os.WriteFile(tmp, data, 0o600); err != nil {
+		os.Remove(tmp)
+		return err
+	}
+	return os.Rename(tmp, t.statePath)
+}
+
+// EligibleHour reports whether a trim may START at local time lt.
+func EligibleHour(lt time.Time) bool {
+	h := lt.Hour()
+	return h >= StartHour && h < EndHour
+}
+
+// weekAnchor is the most recent Wednesday StartHour:00 at or before lt (same location as lt).
+func weekAnchor(lt time.Time) time.Time {
+	daysBack := (int(lt.Weekday()) - int(Weekday) + 7) % 7
+	d := lt.AddDate(0, 0, -daysBack)
+	a := time.Date(d.Year(), d.Month(), d.Day(), StartHour, 0, 0, 0, lt.Location())
+	if a.After(lt) {
+		d = d.AddDate(0, 0, -7)
+		a = time.Date(d.Year(), d.Month(), d.Day(), StartHour, 0, 0, 0, lt.Location())
+	}
+	return a
+}
+
+// due reports whether a guest with record r (ok=false: none) is due at local time lt.
+func due(r Record, has bool, lt time.Time) bool {
+	if !has {
+		return true
+	}
+	anchor := weekAnchor(lt)
+	if r.LastAttemptAt.Before(anchor) {
+		return true // not tried this week
+	}
+	return !r.OK && r.Attempts < MaxAttemptsPerWeek
+}
+
+// Run looks every TickInterval until ctx ends. It never returns an error: a failed trim is a reported fact.
+func (t *Trimmer) Run(ctx context.Context) {
+	t.logger.Info("fstrim: weekly guest disk trim starting", "schedule", ScheduleText)
+	timer := time.NewTimer(FirstTickDelay)
+	defer timer.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-timer.C:
+			t.Pass(ctx)
+			timer.Reset(TickInterval)
+		}
+	}
+}
+
+// Pass is one look: outside the daytime window it does nothing; otherwise it trims every due, running, owned guest
+// while holding the heavy-op gate.
+func (t *Trimmer) Pass(ctx context.Context) {
+	lt := t.now().In(t.loc)
+	if !EligibleHour(lt) {
+		t.logger.Debug("fstrim: outside the daytime window — not looking", "local", lt.Format("Mon 15:04"))
+		return
+	}
+	guests, err := t.guests.Guests(ctx)
+	if err != nil {
+		t.logger.Warn("fstrim: owned-guest list unavailable — skipping this pass", "err", err)
+		return
+	}
+	owned := make(map[int]bool, len(guests))
+	var todo []int
+	t.mu.Lock()
+	for _, g := range guests {
+		owned[g.VMID] = true
+		r, has := t.records[g.VMID]
+		if !due(r, has, lt) {
+			continue
+		}
+		if g.Status != "running" {
+			t.logger.Info("fstrim: guest not running — trimmed when it runs", "vmid", g.VMID, "status", g.Status)
+			continue
+		}
+		todo = append(todo, g.VMID)
+	}
+	// A guest the agent no longer owns has no result to report.
+	pruned := false
+	for id := range t.records {
+		if !owned[id] {
+			delete(t.records, id)
+			pruned = true
+		}
+	}
+	if pruned {
+		if err := t.saveLocked(); err != nil {
+			t.logger.Warn("fstrim: state save failed", "err", err)
+		}
+	}
+	t.mu.Unlock()
+	if len(todo) == 0 {
+		return
+	}
+	sort.Ints(todo)
+	release, busy, ok := t.gate.TryAcquire(GateName)
+	if !ok {
+		t.logger.Info("fstrim: deferred — a heavy operation is in flight; retrying next hour", "busy", busy, "due_guests", len(todo))
+		return
+	}
+	defer release()
+	for _, vmid := range todo {
+		if ctx.Err() != nil {
+			return
+		}
+		t.trimOne(ctx, vmid, lt)
+	}
+}
+
+var trimmedLine = regexp.MustCompile(`\((\d+) bytes\) trimmed`)
+
+// ParseTrimmed sums the "(N bytes) trimmed" lines of `pct fstrim` output and counts them (one per mount point), e.g.
+// `/var/lib/lxc/9201/rootfs/: 30.1 GiB (32277680128 bytes) trimmed`.
+func ParseTrimmed(out string) (bytes int64, mounts int) {
+	for _, m := range trimmedLine.FindAllStringSubmatch(out, -1) {
+		n, err := strconv.ParseInt(m[1], 10, 64)
+		if err != nil {
+			continue
+		}
+		bytes += n
+		mounts++
+	}
+	return bytes, mounts
+}
+
+// GiB renders bytes as "30.1 GiB".
+func GiB(b int64) string { return fmt.Sprintf("%.1f GiB", float64(b)/(1<<30)) }
+
+func (t *Trimmer) trimOne(ctx context.Context, vmid int, lt time.Time) {
+	start := t.now()
+	cctx, cancel := context.WithTimeout(ctx, PerGuestTimeout)
+	stdout, stderr, err := t.runner.Run(cctx, "pct", "fstrim", strconv.Itoa(vmid))
+	cancel()
+	dur := t.now().Sub(start)
+	bytes, mounts := ParseTrimmed(string(stdout) + "\n" + string(stderr))
+
+	t.mu.Lock()
+	prev, has := t.records[vmid]
+	r := Record{LastAttemptAt: start.UTC(), OK: err == nil, BytesTrimmed: bytes, Mounts: mounts,
+		DurationSeconds: float64(dur.Round(100*time.Millisecond)) / float64(time.Second), LastOKAt: prev.LastOKAt}
+	if has && !prev.LastAttemptAt.Before(weekAnchor(lt)) {
+		r.Attempts = prev.Attempts + 1
+	} else {
+		r.Attempts = 1
+	}
+	if err == nil {
+		r.LastOKAt = start.UTC()
+	} else {
+		msg := strings.TrimSpace(err.Error() + ": " + strings.TrimSpace(string(stderr)))
+		if len(msg) > 300 {
+			msg = msg[:300]
+		}
+		r.Error = msg
+	}
+	t.records[vmid] = r
+	saveErr := t.saveLocked()
+	t.mu.Unlock()
+
+	if err == nil {
+		t.logger.Info(fmt.Sprintf("fstrim: guest %d trimmed %s in %.1fs", vmid, GiB(bytes), r.DurationSeconds),
+			"vmid", vmid, "bytes_trimmed", bytes, "mounts", mounts, "duration_s", r.DurationSeconds)
+		if mounts == 0 {
+			t.logger.Warn("fstrim: pct fstrim succeeded but reported no trimmed mount — output not understood",
+				"vmid", vmid, "stdout", strings.TrimSpace(string(stdout)))
+		}
+	} else {
+		t.logger.Warn(fmt.Sprintf("fstrim: guest %d trim FAILED after %.1fs", vmid, r.DurationSeconds),
+			"vmid", vmid, "attempt", r.Attempts, "max_attempts_per_week", MaxAttemptsPerWeek, "err", r.Error)
+	}
+	if saveErr != nil {
+		t.logger.Warn("fstrim: state save failed — the result will not survive a restart", "path", t.statePath, "err", saveErr)
+	}
+}
+
+// GuestDiskTrimStatus implements hub.GuestDiskTrimReporter: a pure read of the persisted results (never runs pct).
+func (t *Trimmer) GuestDiskTrimStatus(context.Context) *hub.GuestDiskTrimStatus {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	out := &hub.GuestDiskTrimStatus{Schedule: ScheduleText}
+	ids := make([]int, 0, len(t.records))
+	for id := range t.records {
+		ids = append(ids, id)
+	}
+	sort.Ints(ids)
+	for _, id := range ids {
+		r := t.records[id]
+		g := hub.GuestDiskTrim{VMID: id, LastAttemptAt: r.LastAttemptAt.UTC().Format(time.RFC3339), OK: r.OK,
+			BytesTrimmed: r.BytesTrimmed, Mounts: r.Mounts, DurationSeconds: r.DurationSeconds, Error: r.Error}
+		if !r.LastOKAt.IsZero() {
+			g.LastOKAt = r.LastOKAt.UTC().Format(time.RFC3339)
+		}
+		out.Guests = append(out.Guests, g)
+	}
+	return out
+}

internal/fstrim/fstrim_test.go

@@ -0,0 +1,267 @@
+package fstrim
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"log/slog"
+	"path/filepath"
+	"reflect"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/backup"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/proxmox"
+)
+
+// The real `pct fstrim 9201` output measured on demo-hp 2026-10-06 (audits/ten-answers-2026-10-06/r444-measure.txt).
+const measuredOut = "/var/lib/lxc/9201/rootfs/: 30.1 GiB (32277680128 bytes) trimmed\n" +
+	"/var/lib/lxc/9201/rootfs/var/lib/felhom: 53.9 GiB (57865633792 bytes) trimmed\n"
+
+const measuredBytes = int64(32277680128 + 57865633792)
+
+type fakeRunner struct {
+	mu    sync.Mutex
+	calls [][]string
+	out   string
+	err   error
+	onRun func()
+}
+
+func (f *fakeRunner) Run(_ context.Context, name string, args ...string) ([]byte, []byte, error) {
+	f.mu.Lock()
+	f.calls = append(f.calls, append([]string{name}, args...))
+	f.mu.Unlock()
+	if f.onRun != nil {
+		f.onRun()
+	}
+	if f.err != nil {
+		return nil, []byte("mount busy"), f.err
+	}
+	return []byte(f.out), nil, nil
+}
+
+type fakeGuests struct {
+	g   []proxmox.Guest
+	err error
+}
+
+func (f fakeGuests) Guests(context.Context) ([]proxmox.Guest, error) { return f.g, f.err }
+
+// A Wednesday 10:30 in a fixed zone (CEST-like), so the tests do not depend on the machine's zone.
+var zone = time.FixedZone("CEST", 2*3600)
+
+func at(day, hour, min int) time.Time { return time.Date(2026, 10, day, hour, min, 0, 0, zone) } // 2026-10-07 = Wednesday
+
+func newT(t *testing.T, r Runner, g GuestSource, gate Gate, now *time.Time) (*Trimmer, *bytes.Buffer, string) {
+	t.Helper()
+	var logs bytes.Buffer
+	path := filepath.Join(t.TempDir(), "guest-disk-trim.json")
+	tr := New(r, g, gate, path, slog.New(slog.NewTextHandler(&logs, &slog.HandlerOptions{Level: slog.LevelDebug})))
+	tr.loc = zone
+	tr.now = func() time.Time { return *now }
+	return tr, &logs, path
+}
+
+func running(ids ...int) fakeGuests {
+	var g []proxmox.Guest
+	for _, id := range ids {
+		g = append(g, proxmox.Guest{VMID: id, Status: "running", Type: "lxc"})
+	}
+	return fakeGuests{g: g}
+}
+
+func TestParseTrimmedTheMeasuredOutput(t *testing.T) {
+	b, m := ParseTrimmed(measuredOut)
+	if b != measuredBytes || m != 2 {
+		t.Fatalf("ParseTrimmed = %d bytes over %d mounts, want %d over 2", b, m, measuredBytes)
+	}
+	if b, m := ParseTrimmed("something else\n"); b != 0 || m != 0 {
+		t.Fatalf("unrelated output parsed as %d/%d", b, m)
+	}
+	if got := GiB(measuredBytes); got != "84.0 GiB" {
+		t.Fatalf("GiB = %q", got)
+	}
+}
+
+// The night window (01:00–06:59) must never be eligible, and the daytime window is exactly 10:00–20:59.
+func TestEligibleHourNeverInTheNight(t *testing.T) {
+	for h := 0; h < 24; h++ {
+		lt := time.Date(2026, 10, 7, h, 30, 0, 0, zone)
+		got := EligibleHour(lt)
+		if h >= 1 && h <= 6 && got {
+			t.Errorf("hour %02d is in the night window and must not be eligible", h)
+		}
+		if want := h >= 10 && h <= 20; got != want {
+			t.Errorf("EligibleHour(%02d:30) = %v, want %v", h, got, want)
+		}
+	}
+}
+
+func TestWeekAnchorIsTheLastWednesdayTen(t *testing.T) {
+	cases := map[time.Time]time.Time{
+		at(7, 10, 0):  at(7, 10, 0),                              // Wednesday 10:00 itself
+		at(7, 9, 59):  time.Date(2026, 9, 30, 10, 0, 0, 0, zone), // before 10:00 Wednesday → the previous week
+		at(8, 15, 0):  at(7, 10, 0),                              // Thursday
+		at(13, 20, 0): at(7, 10, 0),                              // next Tuesday
+		at(14, 11, 0): at(14, 10, 0),                             // next Wednesday
+	}
+	for in, want := range cases {
+		if got := weekAnchor(in); !got.Equal(want) {
+			t.Errorf("weekAnchor(%s) = %s, want %s", in.Format("Mon 01-02 15:04"), got.Format("Mon 01-02 15:04"), want.Format("Mon 01-02 15:04"))
+		}
+	}
+}
+
+// The consequence: on Wednesday 10:30 a running owned guest is trimmed with the ONE exact argv, the bytes are parsed,
+// the positive log line is written, the result is persisted, and the host report carries it.
+func TestPassTrimsADueGuestAndReportsIt(t *testing.T) {
+	now := at(7, 10, 30)
+	r := &fakeRunner{out: measuredOut}
+	tr, logs, path := newT(t, r, running(9201), &backup.InFlight{}, &now)
+	tr.Pass(context.Background())
+
+	if want := [][]string{{"pct", "fstrim", "9201"}}; !reflect.DeepEqual(r.calls, want) {
+		t.Fatalf("runner calls = %q, want %q", r.calls, want)
+	}
+	if !strings.Contains(logs.String(), "fstrim: guest 9201 trimmed 84.0 GiB in ") {
+		t.Fatalf("no positive per-guest log line:\n%s", logs.String())
+	}
+	st := tr.GuestDiskTrimStatus(context.Background())
+	if st == nil || st.Schedule != ScheduleText || len(st.Guests) != 1 {
+		t.Fatalf("report stanza = %+v", st)
+	}
+	g := st.Guests[0]
+	if g.VMID != 9201 || !g.OK || g.BytesTrimmed != measuredBytes || g.Mounts != 2 || g.LastOKAt == "" || g.LastAttemptAt == "" {
+		t.Fatalf("report guest = %+v", g)
+	}
+	// Persisted: a NEW Trimmer over the same file (an agent restart) still has it and does not trim again this week.
+	now = at(8, 11, 0)
+	r2 := &fakeRunner{out: measuredOut}
+	tr2 := New(r2, running(9201), &backup.InFlight{}, path, slog.New(slog.NewTextHandler(&bytes.Buffer{}, nil)))
+	tr2.loc, tr2.now = zone, func() time.Time { return now }
+	if st2 := tr2.GuestDiskTrimStatus(context.Background()); len(st2.Guests) != 1 || st2.Guests[0].BytesTrimmed != measuredBytes {
+		t.Fatalf("result lost over a restart: %+v", st2)
+	}
+	tr2.Pass(context.Background())
+	if len(r2.calls) != 0 {
+		t.Fatalf("trimmed again in the same week after a restart: %q", r2.calls)
+	}
+	// Next week it is due again.
+	now = at(14, 10, 5)
+	tr2.Pass(context.Background())
+	if len(r2.calls) != 1 {
+		t.Fatalf("not trimmed in the next week: %q", r2.calls)
+	}
+}
+
+func TestPassNeverRunsInTheNight(t *testing.T) {
+	for _, h := range []int{1, 3, 6, 9, 21, 23} {
+		now := at(7, h, 15)
+		r := &fakeRunner{out: measuredOut}
+		tr, _, _ := newT(t, r, running(9201), &backup.InFlight{}, &now)
+		tr.Pass(context.Background())
+		if len(r.calls) != 0 {
+			t.Errorf("trimmed at %02d:15: %q", h, r.calls)
+		}
+	}
+}
+
+// A backup (or restore-test) holding the heavy-op gate DEFERS the trim; the next hour, gate free, it runs. And while
+// a trim runs, the gate is held, so a backup cannot start beside it.
+func TestPassDefersToAHeavyOperationAndRetriesNextHour(t *testing.T) {
+	now := at(7, 10, 30)
+	gate := &backup.InFlight{}
+	release, _, _ := gate.TryAcquire("backup:9201")
+	var busyDuringTrim string
+	r := &fakeRunner{out: measuredOut}
+	r.onRun = func() { busyDuringTrim = gate.Busy() }
+	tr, logs, _ := newT(t, r, running(9201), gate, &now)
+
+	tr.Pass(context.Background())
+	if len(r.calls) != 0 {
+		t.Fatalf("trimmed beside a running backup: %q", r.calls)
+	}
+	if !strings.Contains(logs.String(), "fstrim: deferred") || !strings.Contains(logs.String(), "backup:9201") {
+		t.Fatalf("the deferral is not logged with what holds the gate:\n%s", logs.String())
+	}
+	release()
+	now = now.Add(time.Hour)
+	tr.Pass(context.Background())
+	if len(r.calls) != 1 {
+		t.Fatalf("not retried the next hour: %q", r.calls)
+	}
+	if busyDuringTrim != GateName {
+		t.Fatalf("the heavy-op gate was %q during the trim, want %q", busyDuringTrim, GateName)
+	}
+	if gate.Busy() != "" {
+		t.Fatalf("the gate was not released after the pass: %q", gate.Busy())
+	}
+}
+
+func TestFailedTrimWarnsIsRecordedAndRetriedAtMostThreeTimes(t *testing.T) {
+	now := at(7, 10, 30)
+	r := &fakeRunner{err: errors.New("exit status 255")}
+	tr, logs, _ := newT(t, r, running(9201), &backup.InFlight{}, &now)
+	for i := 0; i < 6; i++ {
+		tr.Pass(context.Background())
+		now = now.Add(time.Hour)
+	}
+	if len(r.calls) != MaxAttemptsPerWeek {
+		t.Fatalf("attempts in one week = %d, want %d", len(r.calls), MaxAttemptsPerWeek)
+	}
+	if !strings.Contains(logs.String(), "level=WARN") || !strings.Contains(logs.String(), "fstrim: guest 9201 trim FAILED") {
+		t.Fatalf("no WARN for the failure:\n%s", logs.String())
+	}
+	g := tr.GuestDiskTrimStatus(context.Background()).Guests[0]
+	if g.OK || g.LastOKAt != "" || !strings.Contains(g.Error, "exit status 255") || !strings.Contains(g.Error, "mount busy") {
+		t.Fatalf("failed result not recorded as a failure: %+v", g)
+	}
+	// A success later keeps a clean record.
+	r.err = nil
+	r.out = measuredOut
+	now = at(14, 10, 10)
+	tr.Pass(context.Background())
+	if g := tr.GuestDiskTrimStatus(context.Background()).Guests[0]; !g.OK || g.Error != "" || g.BytesTrimmed != measuredBytes {
+		t.Fatalf("success after failure: %+v", g)
+	}
+}
+
+func TestOnlyRunningOwnedGuestsAndAFailedListActsOnNothing(t *testing.T) {
+	now := at(7, 10, 30)
+	r := &fakeRunner{out: measuredOut}
+	g := fakeGuests{g: []proxmox.Guest{{VMID: 9201, Status: "stopped"}, {VMID: 9202, Status: "running"}}}
+	tr, _, _ := newT(t, r, g, &backup.InFlight{}, &now)
+	tr.Pass(context.Background())
+	if want := [][]string{{"pct", "fstrim", "9202"}}; !reflect.DeepEqual(r.calls, want) {
+		t.Fatalf("calls = %q, want only the running guest", r.calls)
+	}
+	r2 := &fakeRunner{out: measuredOut}
+	tr2, logs, _ := newT(t, r2, fakeGuests{err: errors.New("pool read 403")}, &backup.InFlight{}, &now)
+	tr2.Pass(context.Background())
+	if len(r2.calls) != 0 || !strings.Contains(logs.String(), "owned-guest list unavailable") {
+		t.Fatalf("a failed ownership read must act on nothing: calls %q", r2.calls)
+	}
+}
+
+func TestReportJSONShape(t *testing.T) {
+	now := at(7, 10, 30)
+	tr, _, _ := newT(t, &fakeRunner{out: measuredOut}, running(9201), &backup.InFlight{}, &now)
+	tr.Pass(context.Background())
+	b, err := json.Marshal(tr.GuestDiskTrimStatus(context.Background()))
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, k := range []string{`"schedule":`, `"guests":[{"vmid":9201`, `"last_attempt_at":"2026-10-07T08:30:00Z"`, `"ok":true`,
+		`"bytes_trimmed":90143313920`, `"mounts":2`, `"duration_seconds":`, `"last_ok_at":"2026-10-07T08:30:00Z"`} {
+		if !strings.Contains(string(b), k) {
+			t.Errorf("report JSON lacks %s: %s", k, b)
+		}
+	}
+	if strings.Contains(string(b), `"error"`) {
+		t.Errorf("an ok result must omit error: %s", b)
+	}
+}

internal/hub/collect.go

@@ -90,6 +90,12 @@ type GuestNetReporter interface {
 	GuestNetStatus(ctx context.Context) *GuestNetStatus
 }
 
+// GuestDiskTrimReporter is the R-444 seam the weekly trim job plugs into (same consumer-side pattern — hub does not
+// import fstrim). nil (feature not wired) → no guest_disk_trim stanza.
+type GuestDiskTrimReporter interface {
+	GuestDiskTrimStatus(ctx context.Context) *GuestDiskTrimStatus
+}
+
 // Collector builds a HostReport from read-only sources. All deps are behind narrow
 // interfaces for unit testing.
 type Collector struct {
@@ -108,6 +114,7 @@ type Collector struct {
 	pbsdr            PBSDRReporter                                 // slice 2: PBS DR tier bridge state (nil → stanza omitted)
 	ctrlSup          ControllerSupervisorReporter                  // R-523: in-guest controller supervisor (nil → stanza omitted)
 	guestNet         GuestNetReporter                              // R-54: per-guest network watchdog (nil → stanza omitted)
+	diskTrim         GuestDiskTrimReporter                         // R-444: weekly guest disk trim (nil → stanza omitted)
 	selfUpdate       SelfUpdateReporter                            // D1: agent self-update pending status (nil → false)
 	mgmtPlane        MgmtPlaneReporter                             // G1: management-plane health (nil → stanza omitted)
 	oob              OOBReporter                                   // H1: operator-access health (nil → stanza omitted)
@@ -221,6 +228,12 @@ func (c *Collector) SetGuestNetReporter(g GuestNetReporter) *Collector {
 	return c
 }
 
+// SetGuestDiskTrimReporter wires the R-444 weekly trim job as a report source (nil-safe → stanza omitted).
+func (c *Collector) SetGuestDiskTrimReporter(r GuestDiskTrimReporter) *Collector {
+	c.diskTrim = r
+	return c
+}
+
 // SelfUpdateReporter is the D1 seam the selfupdate commit-manager plugs into (same consumer-side
 // pattern — hub does not import selfupdate). nil (feature not wired) → pending=false on the report.
 type SelfUpdateReporter interface {
@@ -377,6 +390,10 @@ func (c *Collector) Collect(ctx context.Context) (*HostReport, error) {
 	if c.guestNet != nil {
 		report.GuestNet = c.guestNet.GuestNetStatus(ctx)
 	}
+	// R-444: the last weekly trim result per guest (nil reporter = not wired → stanza omitted).
+	if c.diskTrim != nil {
+		report.GuestDiskTrim = c.diskTrim.GuestDiskTrimStatus(ctx)
+	}
 	// D1: agent self-update pending status (nil reporter → pending=false, the steady state).
 	if c.selfUpdate != nil {
 		report.SelfUpdatePending, report.SelfUpdatePendingVersion = c.selfUpdate.SelfUpdatePending()

internal/hub/collect_disktrim_test.go

@@ -0,0 +1,54 @@
+package hub
+
+import (
+	"context"
+	"encoding/json"
+	"testing"
+)
+
+// R-444: the guest_disk_trim stanza must reach a report built through the PRODUCTION collect path, be absent from
+// the wire when the job is not wired, and carry the keys the hub's System page reads.
+
+type fakeDiskTrim struct{ st *GuestDiskTrimStatus }
+
+func (f fakeDiskTrim) GuestDiskTrimStatus(context.Context) *GuestDiskTrimStatus { return f.st }
+
+func TestCollect_GuestDiskTrim(t *testing.T) {
+	px := &fakePx{node: "n", ns: newTestNodeStatus()}
+	c := NewCollector(px, fakeProber{status: "running", detail: "connected"}, fakeObserver{}, nil, nil, nil, "h", "0.150.0", quietLogger())
+	r, err := c.Collect(context.Background())
+	if err != nil {
+		t.Fatalf("Collect: %v", err)
+	}
+	b, _ := json.Marshal(r)
+	var m map[string]any
+	_ = json.Unmarshal(b, &m)
+	if _, ok := m["guest_disk_trim"]; ok {
+		t.Fatalf("guest_disk_trim on the wire with no reporter wired: %s", b)
+	}
+
+	c.SetGuestDiskTrimReporter(fakeDiskTrim{st: &GuestDiskTrimStatus{Schedule: "weekly", Guests: []GuestDiskTrim{{
+		VMID: 9201, LastAttemptAt: "2026-10-07T08:30:00Z", OK: true, BytesTrimmed: 90143313920, Mounts: 2,
+		DurationSeconds: 24.4, LastOKAt: "2026-10-07T08:30:00Z",
+	}}}})
+	r, err = c.Collect(context.Background())
+	if err != nil {
+		t.Fatalf("Collect: %v", err)
+	}
+	b, _ = json.Marshal(r)
+	m = nil
+	_ = json.Unmarshal(b, &m)
+	dt, ok := m["guest_disk_trim"].(map[string]any)
+	if !ok || dt["schedule"] != "weekly" {
+		t.Fatalf("guest_disk_trim missing or wrong on the wire: %s", b)
+	}
+	g := dt["guests"].([]any)[0].(map[string]any)
+	for _, k := range []string{"vmid", "last_attempt_at", "ok", "bytes_trimmed", "mounts", "duration_seconds", "last_ok_at"} {
+		if _, ok := g[k]; !ok {
+			t.Fatalf("guest_disk_trim.guests[0] lacks %q: %v", k, g)
+		}
+	}
+	if g["bytes_trimmed"] != float64(90143313920) || g["ok"] != true {
+		t.Fatalf("values did not survive the round trip: %v", g)
+	}
+}

internal/hub/report.go

@@ -124,6 +124,11 @@ type HostReport struct {
 	// on HostReport would have been the only report block named against that convention.
 	GuestNet *GuestNetStatus `json:"guest_net,omitempty"`
 
+	// GuestDiskTrim is the weekly guest disk trim stanza (R-444, `09` §3 decision 139): the schedule and, per owned
+	// guest, the LAST trim result as persisted by the agent (it survives a restart). Present only when the trim job is
+	// wired; an empty `guests` list means the job runs and no guest has been trimmed yet. No secret.
+	GuestDiskTrim *GuestDiskTrimStatus `json:"guest_disk_trim,omitempty"`
+
 	// LogTail is the agent's on-demand debug-ring tail (v0.83.0 observability) — the agent
 	// mirror of the controller's report log_tails channel. Present ONLY on the heartbeat
 	// right after the control envelope requested it (log_tail_requested); consume-once on
@@ -208,6 +213,27 @@ type GuestNetGuest struct {
 	Message       string `json:"message,omitempty"`
 }
 
+// GuestDiskTrimStatus is the R-444 weekly trim stanza. `schedule` is a plain description of when the job runs (local
+// time of the host); `guests` holds one entry per owned guest that has had at least one trim attempt.
+type GuestDiskTrimStatus struct {
+	Schedule string          `json:"schedule"`
+	Guests   []GuestDiskTrim `json:"guests,omitempty"`
+}
+
+// GuestDiskTrim is one guest's LAST trim attempt. `ok` with `last_attempt_at` is the verdict of that attempt — never
+// read the time alone as success; `last_ok_at` is the last attempt that succeeded ("" = never). `bytes_trimmed` is
+// the sum of the "(N bytes) trimmed" lines `pct fstrim` printed, over `mounts` mount points.
+type GuestDiskTrim struct {
+	VMID            int     `json:"vmid"`
+	LastAttemptAt   string  `json:"last_attempt_at"`
+	OK              bool    `json:"ok"`
+	BytesTrimmed    int64   `json:"bytes_trimmed"`
+	Mounts          int     `json:"mounts"`
+	DurationSeconds float64 `json:"duration_seconds"`
+	LastOKAt        string  `json:"last_ok_at,omitempty"`
+	Error           string  `json:"error,omitempty"`
+}
+
 type PBSDRStatus struct {
 	State          string `json:"state"`
 	StorageID      string `json:"storage_id,omitempty"`
@@ -428,6 +454,17 @@ type SmartSummary struct {
 	ReallocatedSectors   *int `json:"reallocated_sectors"`
 	PendingSectors       *int `json:"pending_sectors"`
 	OfflineUncorrectable *int `json:"offline_uncorrectable"`
+	// R-330 (disk health Phase 2): three more SATA raw counters. omitempty + pointer: absent (an
+	// older agent, an NVMe/USB device, or a drive that does not report the attribute) is OMITTED —
+	// unknown, never a zero (S-39). Wire only: no verdict reads them yet.
+	//   187 Reported_Uncorrect — the failing drive's most telling counter (normalized 1 vs thresh 0,
+	//       raw 1001) while SMART still said PASSED.
+	//   188 Command_Timeout — some vendors PACK several counters into the 48-bit raw value, so the
+	//       number is carried as reported and must not be compared across vendors.
+	//   199 UDMA_CRC_Error_Count — cabling / link errors, not the medium.
+	ReportedUncorrect *int64 `json:"reported_uncorrect,omitempty"`
+	CommandTimeout    *int64 `json:"command_timeout,omitempty"`
+	UDMACRCErrors     *int64 `json:"udma_crc_errors,omitempty"`
 
 	// NVMe attributes.
 	CriticalWarning *int `json:"critical_warning"`

internal/localapi/crashguard.go

@@ -0,0 +1,102 @@
+package localapi
+
+import (
+	"encoding/json"
+	"errors"
+	"io"
+	"io/fs"
+	"net/http"
+	"os"
+	"time"
+)
+
+// GET /host/crash-guard (R-856, `09` §3 decision 143): what the host's crash guard
+// (configs/felhom-crash-guard, `11` §5.9) recorded about the most recent HOST boot. The controller
+// reads it once after it starts: when the host's last boot followed an UNCLEAN stop, its app mails
+// wait ~15 minutes instead of the normal 90 s boot grace.
+//
+// Read-only and Proxmox-free: the agent reads the guard's state file (root-owned, 0644 — the
+// non-root agent can read it) and passes four fields through. Host-wide, token-authed (any valid
+// per-guest token sees the host's view, as GET /host/metrics does).
+//
+// NEVER an error page. A missing file (no guard installed, or no boot recorded yet), an unreadable
+// one, or one that does not parse answers 200 with present:false — the controller reads that as
+// UNKNOWN and keeps its normal boot grace. Pinned by TestR856_CrashGuard*.
+
+// defaultCrashGuardStatePath is where configs/felhom-crash-guard writes its state (STATE_DIR there).
+const defaultCrashGuardStatePath = "/var/lib/felhom-crash-guard/state.json"
+
+// crashGuardStateMax bounds the read; the real file is well under 4 KiB.
+const crashGuardStateMax = 1 << 20
+
+// CrashGuardResponse is the data block of GET /host/crash-guard. Field names are the controller's
+// agentapi.CrashGuardState (felhom-controller internal/agentapi/crashguard.go) — a wire contract,
+// pinned by TestR856_CrashGuardWireMatchesControllerClient.
+type CrashGuardResponse struct {
+	Present         bool   `json:"present"`
+	LastBootAt      string `json:"last_boot_at,omitempty"` // RFC3339 UTC ("2006-01-02T15:04:05Z")
+	LastBootUnclean bool   `json:"last_boot_unclean"`
+	Tripped         bool   `json:"tripped"`
+}
+
+// crashGuardFile is the subset of the guard's state.json the route passes through. Every other key
+// (armed, boot_id, config, unclean_boots, last_trip, ...) is ignored.
+type crashGuardFile struct {
+	LastBootAt      string `json:"last_boot_at"`
+	LastBootUnclean bool   `json:"last_boot_unclean"`
+	Tripped         bool   `json:"tripped"`
+}
+
+// readCrashGuardState reads and parses the guard's state file. ok=false on ANY failure (missing,
+// unreadable, oversized, not a JSON object, a field of the wrong type); reason says which, for the log.
+func readCrashGuardState(path string) (resp CrashGuardResponse, ok bool, reason string) {
+	f, err := os.Open(path)
+	if err != nil {
+		if errors.Is(err, fs.ErrNotExist) {
+			return resp, false, "no state file"
+		}
+		return resp, false, "unreadable: " + err.Error()
+	}
+	defer f.Close()
+	raw, err := io.ReadAll(io.LimitReader(f, crashGuardStateMax+1))
+	if err != nil {
+		return resp, false, "read: " + err.Error()
+	}
+	if len(raw) > crashGuardStateMax {
+		return resp, false, "state file too large"
+	}
+	var st crashGuardFile
+	// Unmarshal into a struct fails on a non-object top level (null decodes, so reject it below).
+	if err := json.Unmarshal(raw, &st); err != nil {
+		return resp, false, "unparseable: " + err.Error()
+	}
+	var probe map[string]json.RawMessage
+	if err := json.Unmarshal(raw, &probe); err != nil || probe == nil {
+		return resp, false, "unparseable: not a JSON object"
+	}
+	resp = CrashGuardResponse{Present: true, LastBootUnclean: st.LastBootUnclean, Tripped: st.Tripped}
+	// Normalise to RFC3339 UTC; an unparseable time passes through as-is (the controller reads an
+	// unparseable boot time as "not this start's boot" → its normal grace).
+	if t, perr := time.Parse(time.RFC3339, st.LastBootAt); perr == nil {
+		resp.LastBootAt = t.UTC().Format(time.RFC3339)
+	} else {
+		resp.LastBootAt = st.LastBootAt
+	}
+	return resp, true, ""
+}
+
+func (s *Server) handleCrashGuard(w http.ResponseWriter, r *http.Request, vmid int) {
+	path := s.crashGuardStatePath
+	if path == "" {
+		path = defaultCrashGuardStatePath
+	}
+	resp, ok, reason := readCrashGuardState(path)
+	if !ok {
+		s.logger.Debug("local-api: /host/crash-guard not present", "vmid", vmid, "reason", reason)
+		writeOK(w, CrashGuardResponse{Present: false})
+		return
+	}
+	s.logger.Debug("local-api: /host/crash-guard served", "vmid", vmid,
+		"last_boot_at", resp.LastBootAt, "last_boot_unclean", resp.LastBootUnclean, "tripped", resp.Tripped)
+	writeOK(w, resp)
+}

internal/localapi/crashguard_test.go

@@ -0,0 +1,205 @@
+package localapi
+
+import (
+	"encoding/json"
+	"io"
+	"log/slog"
+	"net/http"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+// The shape of /var/lib/felhom-crash-guard/state.json as read on demo-hp on 2026-10-06 (values from
+// that read where they matter; lists/objects kept to the same key set).
+const crashGuardFixture = `{
+  "armed": true,
+  "boot_id": "3f1c0f1e-6a0b-4d7e-9b7a-0c2d4e6f8a1b",
+  "config": {"LIMIT": 3, "WINDOW_MINUTES": 60, "PANIC_SECONDS": 10},
+  "kernel_panic": 10,
+  "last_boot_at": "2026-10-05T07:56:41Z",
+  "last_boot_unclean": true,
+  "last_trip": {},
+  "rearmed_at": "2026-10-04T14:02:11Z",
+  "rearmed_by": "operator",
+  "tripped": false,
+  "unclean_boots": ["2026-10-05T07:56:41Z"],
+  "unclean_boots_24h": 1,
+  "unclean_boots_in_window": 1,
+  "updated_at": "2026-10-05T07:57:02Z",
+  "version": 1
+}`
+
+// controllerCrashGuardState is a COPY of the controller's wire type, felhom-controller
+// controller/internal/agentapi/crashguard.go `CrashGuardState` (commit 8b13a5e) — same field names,
+// same tags. If either side renames a key, the contract test below fails.
+type controllerCrashGuardState struct {
+	Present         bool   `json:"present"`
+	LastBootAt      string `json:"last_boot_at,omitempty"`
+	LastBootUnclean bool   `json:"last_boot_unclean"`
+	Tripped         bool   `json:"tripped"`
+}
+
+func newCrashGuardServer(t *testing.T, statePath string) http.Handler {
+	t.Helper()
+	srv, err := NewServer(Options{
+		ListenAddr: "127.0.0.1:0",
+		Guests:     &fakeGuests{},
+		Backups:    &fakeBackups{},
+		Store:      &fakeStore{},
+		Storage:    fakeStorage{},
+		Tokens:     staticTokens{"A": 8200, "B": 9300},
+		Logger:     slog.New(slog.NewTextHandler(io.Discard, nil)),
+	})
+	if err != nil {
+		t.Fatalf("new server: %v", err)
+	}
+	srv.crashGuardStatePath = statePath
+	return srv.Handler()
+}
+
+func writeCrashGuardFixture(t *testing.T, body string) string {
+	t.Helper()
+	p := filepath.Join(t.TempDir(), "state.json")
+	if err := os.WriteFile(p, []byte(body), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	return p
+}
+
+// getCrashGuard calls the route and decodes the envelope with the CONTROLLER's type.
+func getCrashGuard(t *testing.T, h http.Handler, token string) (int, controllerCrashGuardState, string) {
+	t.Helper()
+	w := do(t, h, "GET", "/host/crash-guard", token, "")
+	var env struct {
+		OK   bool                      `json:"ok"`
+		Data controllerCrashGuardState `json:"data"`
+	}
+	if w.Code == http.StatusOK {
+		if err := json.Unmarshal(w.Body.Bytes(), &env); err != nil {
+			t.Fatalf("decode %q: %v", w.Body.String(), err)
+		}
+		if !env.OK {
+			t.Fatalf("ok=false: %s", w.Body.String())
+		}
+	}
+	return w.Code, env.Data, w.Body.String()
+}
+
+// A present state file (demo-hp's shape) passes the three facts through.
+func TestR856_CrashGuardPresentFile(t *testing.T) {
+	h := newCrashGuardServer(t, writeCrashGuardFixture(t, crashGuardFixture))
+	code, st, body := getCrashGuard(t, h, "A")
+	if code != http.StatusOK {
+		t.Fatalf("got %d, want 200 (%s)", code, body)
+	}
+	want := controllerCrashGuardState{Present: true, LastBootAt: "2026-10-05T07:56:41Z", LastBootUnclean: true, Tripped: false}
+	if st != want {
+		t.Fatalf("state = %+v, want %+v", st, want)
+	}
+
+	// A tripped, clean boot reads back as such (both bools are carried, not defaulted).
+	h = newCrashGuardServer(t, writeCrashGuardFixture(t,
+		`{"last_boot_at":"2026-10-05T09:56:41+02:00","last_boot_unclean":false,"tripped":true,"version":1}`))
+	_, st, _ = getCrashGuard(t, h, "B")
+	want = controllerCrashGuardState{Present: true, LastBootAt: "2026-10-05T07:56:41Z", LastBootUnclean: false, Tripped: true}
+	if st != want {
+		t.Fatalf("offset time / tripped: state = %+v, want %+v (time normalised to UTC Z)", st, want)
+	}
+}
+
+// No state file (no guard on this host, or no boot recorded yet) → 200 present:false.
+func TestR856_CrashGuardMissingFile(t *testing.T) {
+	h := newCrashGuardServer(t, filepath.Join(t.TempDir(), "absent", "state.json"))
+	code, st, body := getCrashGuard(t, h, "A")
+	if code != http.StatusOK {
+		t.Fatalf("missing file: got %d, want 200 (%s)", code, body)
+	}
+	if st.Present || st.LastBootUnclean || st.Tripped || st.LastBootAt != "" {
+		t.Fatalf("missing file: state = %+v, want present:false and nothing else", st)
+	}
+}
+
+// A garbled file → 200 present:false, never a 5xx — every shape of garbage.
+func TestR856_CrashGuardGarbageFile(t *testing.T) {
+	for name, body := range map[string]string{
+		"truncated":  crashGuardFixture[:40],
+		"not json":   "this is not json\n",
+		"empty":      "",
+		"null":       "null",
+		"array":      `[{"last_boot_unclean":true}]`,
+		"wrong type": `{"last_boot_at":"2026-10-05T07:56:41Z","last_boot_unclean":"yes","tripped":false}`,
+		"lone brace": "{",
+	} {
+		t.Run(name, func(t *testing.T) {
+			h := newCrashGuardServer(t, writeCrashGuardFixture(t, body))
+			code, st, raw := getCrashGuard(t, h, "A")
+			if code != http.StatusOK {
+				t.Fatalf("got %d, want 200 (%s)", code, raw)
+			}
+			if st.Present || st.LastBootUnclean {
+				t.Fatalf("garbage %q read as %+v, want present:false", name, st)
+			}
+		})
+	}
+	// The path is a directory, not a file: unreadable → present:false, 200.
+	h := newCrashGuardServer(t, t.TempDir())
+	if code, st, raw := getCrashGuard(t, h, "A"); code != http.StatusOK || st.Present {
+		t.Fatalf("directory path: got %d %+v (%s), want 200 present:false", code, st, raw)
+	}
+}
+
+// No / unknown token → 401, like every sibling route; a cross-guest ?vmid= → 403.
+func TestR856_CrashGuardRequiresGuestToken(t *testing.T) {
+	h := newCrashGuardServer(t, writeCrashGuardFixture(t, crashGuardFixture))
+	for _, tok := range []string{"", "bogus"} {
+		w := do(t, h, "GET", "/host/crash-guard", tok, "")
+		if w.Code != http.StatusUnauthorized {
+			t.Fatalf("token %q: got %d, want 401", tok, w.Code)
+		}
+		if json.Valid(w.Body.Bytes()) {
+			var env struct {
+				Data controllerCrashGuardState `json:"data"`
+			}
+			_ = json.Unmarshal(w.Body.Bytes(), &env)
+			if env.Data.Present || env.Data.LastBootUnclean {
+				t.Fatalf("token %q: the refusal leaked the state: %s", tok, w.Body.String())
+			}
+		}
+	}
+	if w := do(t, h, "GET", "/host/crash-guard?vmid=9300", "A", ""); w.Code != http.StatusForbidden {
+		t.Fatalf("cross-guest query: got %d, want 403", w.Code)
+	}
+}
+
+// Wire contract: every key the controller's CrashGuardState decodes is emitted under exactly that
+// name, and the agent emits no key the controller does not know.
+func TestR856_CrashGuardWireMatchesControllerClient(t *testing.T) {
+	h := newCrashGuardServer(t, writeCrashGuardFixture(t, crashGuardFixture))
+	w := do(t, h, "GET", "/host/crash-guard", "A", "")
+	var env struct {
+		OK   bool                       `json:"ok"`
+		Data map[string]json.RawMessage `json:"data"`
+	}
+	if err := json.Unmarshal(w.Body.Bytes(), &env); err != nil || !env.OK {
+		t.Fatalf("envelope: %v %s", err, w.Body.String())
+	}
+	want := []string{"present", "last_boot_at", "last_boot_unclean", "tripped"}
+	for _, k := range want {
+		if _, ok := env.Data[k]; !ok {
+			t.Errorf("agent does not emit %q, which the controller decodes (%s)", k, w.Body.String())
+		}
+	}
+	if len(env.Data) != len(want) {
+		t.Errorf("agent emits %d keys, controller knows %d: %s", len(env.Data), len(want), w.Body.String())
+	}
+	// And the agent's own type agrees with the controller's copy, field for field.
+	var mine CrashGuardResponse
+	var theirs controllerCrashGuardState
+	raw, _ := json.Marshal(env.Data)
+	_ = json.Unmarshal(raw, &mine)
+	_ = json.Unmarshal(raw, &theirs)
+	if (controllerCrashGuardState{mine.Present, mine.LastBootAt, mine.LastBootUnclean, mine.Tripped}) != theirs {
+		t.Errorf("agent %+v vs controller %+v", mine, theirs)
+	}
+}

internal/localapi/r894_saved_backup_test.go

@@ -0,0 +1,144 @@
+package localapi
+
+import (
+	"context"
+	"io"
+	"log/slog"
+	"net/http"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/backup"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+)
+
+// R-894 — after an agent restart, an UNREADABLE storage must fall back to the last success saved on
+// disk, not to "never". Measured 2026-10-05 on demo-hp: a restart at 04:57, the off-site storage
+// unreachable at 06:25, the 7-day tier (last copy 4 days old) read DUE, vzdump failed.
+//
+// Every test here builds a NEW server and a NEW BackupSuccessState from the same file — that is the
+// restart. The in-memory store (fakeStore) is always fresh, as after a real restart.
+
+// r894Server builds a two-tier server whose off-site tier answers the storage listing with lister.
+func r894Server(t *testing.T, path string, pbsSvc BackupService) *Server {
+	t.Helper()
+	srv, err := NewServer(Options{
+		ListenAddr: "127.0.0.1:0", Guests: &fakeGuests{}, Backups: &fakeBackups{}, Store: &fakeStore{},
+		Storage: fakeStorage{targets: []hub.StorageTarget{{Name: "local"}, {Name: "felhom-pbs"}}},
+		Tokens:  staticTokens{"A": 8200},
+		BackupTiers: []BackupTier{
+			{TargetID: "local", Cadence: 24 * time.Hour, Primary: true, Service: &fakeBackups{}},
+			{TargetID: "felhom-pbs", Cadence: 7 * 24 * time.Hour, Service: pbsSvc},
+		},
+		LastKnownBackups: backup.NewBackupSuccessState(path),
+		Logger:           slog.New(slog.NewTextHandler(io.Discard, nil)),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	srv.baseCtx = context.Background()
+	srv.now = func() time.Time { return testNow }
+	return srv
+}
+
+// unreadable is the off-site storage as demo-hp saw it: "Can't connect to 10.77.0.1:8007".
+func unreadable() archiveLister {
+	return archiveLister{fakeBackups: &fakeBackups{}, err: errStorageRead}
+}
+
+// THE R-894 CASE, end to end. Agent 1 takes an off-site backup through POST /backup (the fake
+// runner's success is 12 h before testNow). The agent restarts. The storage cannot be read. The tier
+// must read NOT due, from the copy saved on disk.
+//
+// COMPANION RED-PROOF (observed): delete the `lookup == archiveUnknown && s.lastKnown != nil` block in
+// handleBackupDue → this fails with "after a restart an unreadable storage must fall back to the saved
+// copy (12 h old, 7-day tier) — NOT due; got {… Due:true … AgeState:unknown …}". Restored.
+func TestBackupDue_R894_RestartThenUnreadableStorage_FreshSavedCopyIsNotDue(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "backup-success-state.json")
+
+	// Agent 1: a real backup job through the endpoint the controller calls.
+	first := r894Server(t, path, &fakeBackups{})
+	if rr := do(t, first.Handler(), "POST", "/backup?target=felhom-pbs", "A", ""); rr.Code != http.StatusAccepted {
+		t.Fatalf("POST /backup: %d %s", rr.Code, rr.Body.String())
+	}
+	waitFor(t, func() bool {
+		_, ok := backup.NewBackupSuccessState(path).LastKnownSuccess("felhom-pbs", 8200)
+		return ok
+	})
+
+	// Agent 2: a restart (new server, new state from the same file), and the storage is unreachable.
+	got := dueFor(t, r894Server(t, path, unreadable()).Handler(), "felhom-pbs")
+	if got.Due {
+		t.Fatalf("after a restart an unreadable storage must fall back to the saved copy (12 h old, 7-day tier) — NOT due; got %+v", got)
+	}
+	if got.AgeState != AgeStateKnown || got.AgeSecs == nil || *got.AgeSecs != int64((12*time.Hour).Seconds()) {
+		t.Fatalf("the age must come from the saved copy (12 h, known); got %+v", got)
+	}
+}
+
+// The deliberate rule stays: an unreadable storage must not suppress a backup that IS due. A saved
+// copy older than the cadence reads DUE.
+//
+// COMPANION RED-PROOF (observed): make the fallback answer not-due whenever a saved copy exists
+// (`if fromDisk { …Due:false… }` before the cadence check) → this fails with "a saved copy 9 days old
+// under a 7-day cadence MUST read due". Restored.
+func TestBackupDue_R894_RestartThenUnreadableStorage_OldSavedCopyIsDue(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "backup-success-state.json")
+	st := backup.NewBackupSuccessState(path)
+	if err := st.RecordBackupSuccess("felhom-pbs", backupAt("felhom-pbs", 8200, 9*24*time.Hour, true)); err != nil {
+		t.Fatal(err)
+	}
+	got := dueFor(t, r894Server(t, path, unreadable()).Handler(), "felhom-pbs")
+	if !got.Due {
+		t.Fatalf("a saved copy 9 days old under a 7-day cadence MUST read due; got %+v", got)
+	}
+	if got.AgeState != AgeStateKnown {
+		t.Fatalf("the age is known (from disk); got %+v", got)
+	}
+}
+
+// No saved copy → the pre-R-894 answer, byte for byte: DUE, age UNKNOWN (never ABSENT — the controller
+// fires its window-gate valve only on absent, R-88).
+func TestBackupDue_R894_RestartThenUnreadableStorage_NoSavedCopyIsDueUnknown(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "backup-success-state.json")
+	got := dueFor(t, r894Server(t, path, unreadable()).Handler(), "felhom-pbs")
+	if !got.Due || got.AgeState != AgeStateUnknown || got.AgeSecs != nil {
+		t.Fatalf("no saved copy + unreadable storage must stay DUE with age unknown; got %+v", got)
+	}
+}
+
+// A storage that ANSWERS is the ground truth: an archive absent there makes the tier due even when the
+// file remembers a fresh success (a pruned or deleted copy must be made again).
+//
+// COMPANION RED-PROOF (observed): drop `lookup == archiveUnknown &&` from the fallback condition → this
+// fails with "the storage answered 'no archive' — the saved copy must NOT stand in for it". Restored.
+func TestBackupDue_R894_SavedCopyIgnoredWhenStorageAnswers(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "backup-success-state.json")
+	st := backup.NewBackupSuccessState(path)
+	if err := st.RecordBackupSuccess("felhom-pbs", backupAt("felhom-pbs", 8200, time.Hour, true)); err != nil {
+		t.Fatal(err)
+	}
+	absent := archiveLister{fakeBackups: &fakeBackups{}, found: false}
+	got := dueFor(t, r894Server(t, path, absent).Handler(), "felhom-pbs")
+	if !got.Due {
+		t.Fatalf("the storage answered 'no archive' — the saved copy must NOT stand in for it; got %+v", got)
+	}
+}
+
+// A FAILED backup is never saved: it must not make a tier look fresh after a restart.
+func TestBackupDue_R894_FailedBackupIsNotSaved(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "backup-success-state.json")
+	first := r894Server(t, path, &fakeBackups{failErr: "could not activate storage 'felhom-pbs'"})
+	if rr := do(t, first.Handler(), "POST", "/backup?target=felhom-pbs", "A", ""); rr.Code != http.StatusAccepted {
+		t.Fatalf("POST /backup: %d %s", rr.Code, rr.Body.String())
+	}
+	waitFor(t, func() bool { return len(first.store.Backups(context.Background())) == 1 })
+	if _, ok := backup.NewBackupSuccessState(path).LastKnownSuccess("felhom-pbs", 8200); ok {
+		t.Fatal("a failed backup must never be saved as a success")
+	}
+	got := dueFor(t, r894Server(t, path, unreadable()).Handler(), "felhom-pbs")
+	if !got.Due {
+		t.Fatalf("after a failed backup and a restart the tier must still be due; got %+v", got)
+	}
+}

internal/localapi/server.go

@@ -85,6 +85,13 @@ type BackupStore interface {
 	RestoreTests(ctx context.Context) []hub.RestoreTest
 }
 
+// LastKnownBackupStore (R-894) is the on-disk newest-success-per-tier record. Satisfied by
+// *backup.BackupSuccessState.
+type LastKnownBackupStore interface {
+	RecordBackupSuccess(target string, b hub.Backup) error
+	LastKnownSuccess(target string, vmid int) (time.Time, bool)
+}
+
 // StorageView yields the host's observed storage targets (for mapping a mount's storage id →
 // fast/slow class). Satisfied by *storage.Observer.
 type StorageView interface {
@@ -164,6 +171,10 @@ type Options struct {
 	// PRIMARY tier, inside the backup goroutine and BEFORE the host-wide heavy-op gate is released — so the OS leg
 	// that it starts can never overlap another backup or a restore-test (`11` C10). OPTIONAL — nil → nothing runs.
 	AfterPrimaryBackup func(ctx context.Context, vmid int)
+	// LastKnownBackups (R-894) keeps the newest successful backup per tier ON DISK, so the due-check's
+	// fallback for an UNREADABLE storage after an agent restart is the last known copy, not "never".
+	// nil = the pre-R-894 behaviour (in-memory record only).
+	LastKnownBackups LastKnownBackupStore
 	// Privileged runs the fenced root wrappers (E-2a: felhom-backup-target-apply). OPTIONAL — when
 	// nil, POST /backup/target reports "not configured". Satisfied by *proxmox.ExecRunner.
 	Privileged PrivilegedRunner
@@ -229,7 +240,6 @@ type Options struct {
 	// POST /escrow/recover-offsite-password. OPTIONAL — nil → that route reports "not configured"
 	// (503) instead of failing obscurely. Satisfied by escrow.OffsiteKeyRecoverer.
 	EscrowRecovery EscrowRecoverer
-
 }
 
 // defaultBackupCadence is the fallback /backup/due window when none is configured.
@@ -279,30 +289,34 @@ type Server struct {
 	// is the pre-R-82 shape.
 	tiers []BackupTier
 	// inFlight (R-85) is shared with the restore-test scheduler so the two never run together.
-	inFlight *backup.InFlight
+	inFlight           *backup.InFlight
 	afterPrimaryBackup func(ctx context.Context, vmid int) // the OS leg (agent v0.140.0); nil = none
-	logger   *slog.Logger
+	lastKnown          LastKnownBackupStore                // R-894: on-disk newest success per tier; nil = none
-	now      func() time.Time
+	logger             *slog.Logger
+	now                func() time.Time
 
-	disks           DiskOps             // slice 8C (optional)
+	disks           DiskOps           // slice 8C (optional)
-	diskGate        StorageGate         // slice 8C (optional)
+	diskGate        StorageGate       // slice 8C (optional)
-	guestList       GuestLister         // slice 8C (optional)
+	guestList       GuestLister       // slice 8C (optional)
-	guestAttach     GuestAttacher       // slice 10 P2 (optional)
+	guestAttach     GuestAttacher     // slice 10 P2 (optional)
-	mem             MemoryOps           // v0.90.0 R-24 guest RAM resize (optional)
+	mem             MemoryOps         // v0.90.0 R-24 guest RAM resize (optional)
-	memMu           sync.Mutex          // single-flight around a resize apply (one customer per host)
+	memMu           sync.Mutex        // single-flight around a resize apply (one customer per host)
-	netStorage      NetworkStorageOps   // Part A1: NAS network mounts (optional)
+	netStorage      NetworkStorageOps // Part A1: NAS network mounts (optional)
-	netMountRoot    string              // the user-data namespace root for the network-mount role gate
+	netMountRoot    string            // the user-data namespace root for the network-mount role gate
-	smbCredsDir     string              // where SMB creds files are written (out-of-band, 0600)
+	smbCredsDir     string            // where SMB creds files are written (out-of-band, 0600)
-	escrowStagePath string              // fork-4: 0600 staging file for the pushed restic repo password
+	escrowStagePath string            // fork-4: 0600 staging file for the pushed restic repo password
+	// crashGuardStatePath (R-856) is the host crash guard's state file read by GET /host/crash-guard;
+	// empty = defaultCrashGuardStatePath. A seam: tests point it at a fixture.
+	crashGuardStatePath string
 	// escrowRecovery (R-199, v0.125.0) assembles chain links 6-8: fetch this host's own sealed
 	// identity blob from the hub, unseal it with the customer's recovery code, return ONLY the
 	// offsite repository password. OPTIONAL — nil (no hub client configured) makes
 	// POST /escrow/recover-offsite-password answer 503 rather than pretending.
-	escrowRecovery  EscrowRecoverer
+	escrowRecovery EscrowRecoverer
-	intent          IntentRecorder      // slice 10 P3 (optional)
+	intent         IntentRecorder      // slice 10 P3 (optional)
-	guestBinds      *GuestBindStore     // F9 startup bind re-assert record (optional)
+	guestBinds     *GuestBindStore     // F9 startup bind re-assert record (optional)
-	formatJobs      *FormatJobStore     // F20-BUG3 detached-format job record (optional)
+	formatJobs     *FormatJobStore     // F20-BUG3 detached-format job record (optional)
-	staleLock       StaleLockController // F2-b startup stale-lock recovery (optional)
+	staleLock      StaleLockController // F2-b startup stale-lock recovery (optional)
 	// guestPower (F-REBOOT) is per-guest start-attempt state for the guest-power watchdog.
 	// Guarded by guestPowerMu in guestpower.go; in-memory on purpose (see guestPowerState).
 	guestPower map[int]guestPowerState
@@ -475,6 +489,7 @@ func NewServer(o Options) (*Server, error) {
 	s.tiers = normalizeBackupTiers(o.BackupTiers, o.Backups, cadence)
 	s.inFlight = o.InFlight
 	s.afterPrimaryBackup = o.AfterPrimaryBackup
+	s.lastKnown = o.LastKnownBackups
 	if s.backups == nil && len(s.tiers) > 0 {
 		s.backups = s.tiers[0].Service
 	}
@@ -520,6 +535,9 @@ func (s *Server) Handler() http.Handler {
 	// Host metrics (slice 9): host-wide health + per-storage capacity for the customer's monitoring
 	// view. Host-wide, token-authed, fresh (a live collect, not the 15-min hub snapshot).
 	mux.HandleFunc("GET /host/metrics", s.withGuest(s.handleHostMetrics))
+	// R-856 (`09` §3 decision 143): the host crash guard's record of the last HOST boot — the controller
+	// waits ~15 min with app mails after an unclean one. Read-only; a missing/garbled file = present:false.
+	mux.HandleFunc("GET /host/crash-guard", s.withGuest(s.handleCrashGuard))
 	// Disk management (slice 8C) — self-scoped; format routes through the data-bearing classifier+gate.
 	mux.HandleFunc("GET /disks", s.withGuest(s.handleDisks))
 	mux.HandleFunc("GET /disks/candidates", s.withGuest(s.handleDiskCandidates))
@@ -899,6 +917,13 @@ func (s *Server) handleBackup(w http.ResponseWriter, r *http.Request, vmid int)
 			s.logger.Info("local-api: backup job complete", "vmid", vmid, "target", tier.TargetID, "job", jobID, "archive", b.Archive)
 		}
 		s.store.RecordBackup(b)
+		if b.Success && s.lastKnown != nil {
+			if err := s.lastKnown.RecordBackupSuccess(tier.TargetID, b); err != nil {
+				// Not fatal: the backup exists. Only the fallback after a restart loses this copy.
+				s.logger.Warn("local-api: could not save the backup on disk for the due-check fallback (R-894)",
+					"vmid", vmid, "target", tier.TargetID, "err", err)
+			}
+		}
 		s.finishJob(key, jobID, b)
 		// OS leg (agent v0.140.0): after the night's whole-guest copy exists, still holding the heavy-op gate.
 		if b.Success && tier.Primary && s.afterPrimaryBackup != nil {
@@ -1061,6 +1086,20 @@ func (s *Server) handleBackupDue(w http.ResponseWriter, r *http.Request, vmid in
 		newest, haveNewest = t, true
 		unparseable = false // ground truth supersedes an unreadable in-memory timestamp
 	}
+	// R-894: the storage could not be read → the last success saved ON DISK stands in for the in-memory
+	// record a restart emptied. ONLY on archiveUnknown: a storage that answers is the ground truth, and an
+	// archive absent there must make the tier due even when the file remembers one (a pruned copy).
+	// A saved copy older than the cadence still reads DUE below — an unreadable storage never suppresses
+	// a backup that is due.
+	fromDisk := false
+	if lookup == archiveUnknown && s.lastKnown != nil {
+		if saved, ok := s.lastKnown.LastKnownSuccess(tier.TargetID, vmid); ok && (!haveNewest || saved.After(newest)) {
+			newest, haveNewest, fromDisk = saved, true, true
+			unparseable = false
+			s.logger.Info("local-api: backup storage unreadable — due-check uses the last success saved on disk (R-894)",
+				"vmid", vmid, "target", tier.TargetID, "saved", saved.UTC().Format(time.RFC3339))
+		}
+	}
 	if !haveNewest {
 		// R-88 Part 2: THREE distinct reasons for a nil age, each with its own state. Only ABSENT is a
 		// positive claim of "never backed up"; only that one may license the controller to bypass its
@@ -1083,13 +1122,17 @@ func (s *Server) handleBackupDue(w http.ResponseWriter, r *http.Request, vmid in
 	}
 	age := s.now().Sub(newest)
 	ageSecs := int64(age.Seconds())
+	suffix := ""
+	if fromDisk {
+		suffix = " (storage unreadable — age from the last success saved on disk)"
+	}
 	if age >= tier.Cadence {
 		writeOK(w, BackupDueResponse{VMID: vmid, Due: true, AgeSecs: &ageSecs, AgeState: AgeStateKnown,
-			Reason: "older than cadence", Target: echo})
+			Reason: "older than cadence" + suffix, Target: echo})
 		return
 	}
 	writeOK(w, BackupDueResponse{VMID: vmid, Due: false, AgeSecs: &ageSecs, AgeState: AgeStateKnown,
-		Reason: "within cadence window", Target: echo})
+		Reason: "within cadence window" + suffix, Target: echo})
 }
 
 // BackupTiersResponse is GET /backup/tiers (R-82): the tiers this agent serves, primary first.
@@ -1477,4 +1520,6 @@ func writeStatus(w http.ResponseWriter, code int, ok bool, data any, errMsg stri
 }
 
 // SetAfterPrimaryBackup wires the hook that runs after a successful primary-tier backup (the OS leg, agent v0.140.0).
-func (s *Server) SetAfterPrimaryBackup(fn func(ctx context.Context, vmid int)) { s.afterPrimaryBackup = fn }
+func (s *Server) SetAfterPrimaryBackup(fn func(ctx context.Context, vmid int)) {
+	s.afterPrimaryBackup = fn
+}

internal/osupdate/leg.go

@@ -106,6 +106,9 @@ type WrapperReport struct {
 	LiveRestore         json.RawMessage `json:"live_restore"`
 	Facts               json.RawMessage `json:"facts"`
 	Bundle              json.RawMessage `json:"bundle"` // the config bundle's result (R-840, mode "bundle")
+	// OOMCheck (R-528, `09` decision 157): the docker layer's memory-kill check, {result, oom_killed, oom_event,
+	// exit_code, image, detail}. Carried to the hub UNCHANGED; the agent never reads it.
+	OOMCheck json.RawMessage `json:"oom_check"`
 	// R-868 (v0.144.0): the agent's ids, echoed from the plan, so a report kept on disk can be sent without the
 	// agent process that started the pass. ReleaseID / VMID were always in the report.
 	RunID     string `json:"run_id"`
@@ -143,6 +146,9 @@ type Report struct {
 	DockerEngine        string          `json:"docker_engine,omitempty"` // docker layer: the engine after the step
 	Authority           string          `json:"authority,omitempty"`     // docker layer: ring0 | signed
 	Undo                bool            `json:"undo,omitempty"`          // docker layer: a signed undo (downgrade)
+	// OOMCheck: docker layer — the wrapper's oom_check object, byte-for-byte (R-528; the hub decides approval on it).
+	// Pinned by TestDocker_OOMCheckReachesTheHubUnchanged and TestR868_KeptCopyCarriesTheOOMCheck.
+	OOMCheck json.RawMessage `json:"oom_check,omitempty"`
 
 	unsent string // R-868: the wrapper's kept copy of this pass's report — deleted once the hub has it
 }
@@ -630,6 +636,7 @@ func (l *Leg) runLayer(ctx context.Context, runID, layer string, vmid int, trigg
 	}
 	rep.Upgraded, rep.PassSeconds = wr.Upgraded, wr.PassSeconds
 	rep.DockerEngine, rep.Authority, rep.Undo = wr.DockerEngine, wr.Authority, wr.Undo
+	rep.OOMCheck = rawOrNil(wr.OOMCheck)
 	if rep.Outcome == "" {
 		switch {
 		case rep.Mode == "inventory" && !blk.Enabled:
@@ -707,6 +714,14 @@ func (l *Leg) runLayer(ctx context.Context, runID, layer string, vmid int, trigg
 	return l.finish(ctx, lg, rep)
 }
 
+// rawOrNil: a wrapper field that is absent or JSON null stays out of the hub report (omitempty).
+func rawOrNil(m json.RawMessage) json.RawMessage {
+	if len(m) == 0 || string(m) == "null" {
+		return nil
+	}
+	return m
+}
+
 func onlyDocker(in []Package) []Package {
 	var out []Package
 	for _, p := range in {

internal/osupdate/leg_test.go

@@ -98,9 +98,13 @@ func (f *fakeWrapper) RunStdin(ctx context.Context, _ io.Reader, name string, ar
 	return f.Run(ctx, name, args...)
 }
 
-type fakeHub struct{ reports []Report }
+type fakeHub struct {
+	reports []Report
+	bodies  [][]byte // the exact bytes posted (R-528: the oom_check object must arrive unchanged)
+}
 
 func (h *fakeHub) PostOSReport(_ context.Context, body []byte) error {
+	h.bodies = append(h.bodies, append([]byte(nil), body...))
 	var r Report
 	json.Unmarshal(body, &r)
 	h.reports = append(h.reports, r)
@@ -504,3 +508,42 @@ func TestHealthVerdict_ControllerBlindToDockerFails(t *testing.T) {
 		t.Fatal("an older wrapper (no field) must not fail")
 	}
 }
+
+// R-528 (`09` decision 157): the wrapper's oom_check object reaches the hub's docker report byte-for-byte; the guest
+// and host reports carry none. COMPANION RED-PROOF: drop `rep.OOMCheck = rawOrNil(wr.OOMCheck)` in runLayer → "no
+// oom_check in the docker report".
+const oomCheckWire = `{"detail":"the engine reported the memory kill: OOMKilled=true and the oom event","exit_code":137,"image":"gitea.dooplex.hu/admin/felhom-controller:0.300.0","oom_event":true,"oom_killed":true,"result":"pass"}`
+
+func TestDocker_OOMCheckReachesTheHubUnchanged(t *testing.T) {
+	w := &fakeWrapper{t: t, applyRep: map[string]WrapperReport{LayerDocker: {
+		Upgraded:     []Package{{Name: "docker-ce", Version: "5:29.8.2-1~debian.13~trixie"}},
+		DockerEngine: "29.8.2", Authority: "ring0", OOMCheck: json.RawMessage(oomCheckWire)}}}
+	l, h := newLeg(t, w, &hub.WireOSUpdate{Ring: 0, Enabled: true})
+	l.Run(context.Background(), 9201, "night")
+	found := false
+	for _, b := range h.bodies {
+		var m map[string]json.RawMessage
+		if err := json.Unmarshal(b, &m); err != nil {
+			t.Fatal(err)
+		}
+		var layer string
+		json.Unmarshal(m["layer"], &layer)
+		oc, has := m["oom_check"]
+		if layer != LayerDocker {
+			if has {
+				t.Fatalf("the %s report carries an oom_check: %s", layer, oc)
+			}
+			continue
+		}
+		found = true
+		if !has {
+			t.Fatalf("no oom_check in the docker report: %s", b)
+		}
+		if string(oc) != oomCheckWire {
+			t.Fatalf("oom_check changed on the way:\n got %s\nwant %s", oc, oomCheckWire)
+		}
+	}
+	if !found {
+		t.Fatalf("no docker report posted: %s", calls(w))
+	}
+}

internal/osupdate/unsent.go

@@ -144,6 +144,7 @@ func (l *Leg) reportFromKept(ctx context.Context, wr WrapperReport, path string)
 	}
 	rep.Upgraded, rep.PassSeconds = wr.Upgraded, wr.PassSeconds
 	rep.DockerEngine, rep.Authority, rep.Undo = wr.DockerEngine, wr.Authority, wr.Undo
+	rep.OOMCheck = rawOrNil(wr.OOMCheck)
 	wantEngine := ""
 	for _, u := range wr.Upgraded {
 		if u.Name == "docker-ce" {

internal/osupdate/unsent_test.go

@@ -150,3 +150,24 @@ func must(t *testing.T, err error) {
 		t.Fatal(err)
 	}
 }
+
+// R-528: a kept docker report (the agent was killed) still carries the oom_check object to the hub unchanged.
+// COMPANION RED-PROOF: drop `rep.OOMCheck = rawOrNil(wr.OOMCheck)` in reportFromKept → "oom_check lost".
+func TestR868_KeptCopyCarriesTheOOMCheck(t *testing.T) {
+	w := &fakeWrapper{t: t}
+	l, h := newLeg(t, w, &hub.WireOSUpdate{Ring: 0, Enabled: true})
+	ring := 0
+	kept := WrapperReport{Mode: "apply", Layer: LayerDocker, RunID: "20261007T020000Z", Trigger: "night", Ring: &ring, VMID: 9201,
+		ReleaseID: "ring0-20261007T020000Z", HealthBefore: guestOK(), HealthAfter: guestOK(), DockerEngine: "29.8.2",
+		Upgraded: []Package{{Name: "docker-ce", Version: "5:29.8.2-1~debian.13~trixie"}}, OOMCheck: json.RawMessage(oomCheckWire)}
+	b, _ := json.Marshal(kept)
+	must(t, os.WriteFile(reportFile(l.PlanDir, kept.RunID, LayerDocker, "apply"), b, 0o600))
+	if n := l.SendUnsent(context.Background()); n != 1 || len(h.bodies) != 1 {
+		t.Fatalf("sent %d, bodies %d", n, len(h.bodies))
+	}
+	var m map[string]json.RawMessage
+	must(t, json.Unmarshal(h.bodies[0], &m))
+	if string(m["oom_check"]) != oomCheckWire {
+		t.Fatalf("oom_check lost or changed: %q", m["oom_check"])
+	}
+}

internal/storage/r330_smart_attrs_test.go

@@ -0,0 +1,64 @@
+package storage
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+// R-330 (disk health Phase 2) — attributes 187, 188 and 199 ride the wire.
+//
+// The failing drive of 2026-08-14 carried 187 Reported_Uncorrect at raw 1001 while SMART said PASSED;
+// none of the three reached the controller. The values are carried as RAW counters; an absent
+// attribute is OMITTED from the JSON (unknown), never sent as 0 (S-39).
+
+// The 2026-08-14 shape: PASSED, 187 raw 1001, plus 188/199 and the existing three.
+const r330SATA = `{"smart_status":{"passed":true},"ata_smart_attributes":{"table":[
+ {"id":5,"raw":{"value":0}},
+ {"id":187,"raw":{"value":1001}},
+ {"id":188,"raw":{"value":4295032833}},
+ {"id":197,"raw":{"value":8}},
+ {"id":198,"raw":{"value":8}},
+ {"id":199,"raw":{"value":3}}]}}`
+
+// COMPANION RED-PROOF (observed): delete the three R-330 cases in parseSMART → this fails with
+// "187 Reported_Uncorrect must be carried (raw 1001); got <nil>". Restored.
+func TestParseSMART_R330_CarriesTheThreeCounters(t *testing.T) {
+	s := parseSMART([]byte(r330SATA))
+	if s.ReportedUncorrect == nil || *s.ReportedUncorrect != 1001 {
+		t.Fatalf("187 Reported_Uncorrect must be carried (raw 1001); got %v", s.ReportedUncorrect)
+	}
+	// 188's raw value is vendor-packed on some drives (this one is 0x100010001): carried as reported, not truncated.
+	if s.CommandTimeout == nil || *s.CommandTimeout != 4295032833 {
+		t.Fatalf("188 Command_Timeout must be carried as the full raw value; got %v", s.CommandTimeout)
+	}
+	if s.UDMACRCErrors == nil || *s.UDMACRCErrors != 3 {
+		t.Fatalf("199 UDMA_CRC_Error_Count must be carried (raw 3); got %v", s.UDMACRCErrors)
+	}
+	if s.PendingSectors == nil || *s.PendingSectors != 8 {
+		t.Fatalf("the existing counters must be unchanged; pending=%v", s.PendingSectors)
+	}
+}
+
+// A drive (or an NVMe device) that does not report the attributes leaves them nil, and the JSON
+// OMITS the keys — the receiver reads "unknown", never a measured zero.
+//
+// COMPANION RED-PROOF (observed): drop `,omitempty` from the three tags in hub.SmartSummary → this
+// fails with "an unreported attribute must be omitted, not sent: … reported_uncorrect …". Restored.
+func TestParseSMART_R330_AbsentIsOmittedNotZero(t *testing.T) {
+	for name, raw := range map[string]string{
+		"sata without the three": `{"smart_status":{"passed":true},"ata_smart_attributes":{"table":[{"id":5,"raw":{"value":0}}]}}`,
+		"nvme":                   `{"smart_status":{"passed":true},"nvme_smart_health_information_log":{"critical_warning":0,"media_errors":0,"percentage_used":3}}`,
+	} {
+		s := parseSMART([]byte(raw))
+		if s.ReportedUncorrect != nil || s.CommandTimeout != nil || s.UDMACRCErrors != nil {
+			t.Fatalf("%s: unreported attributes must stay nil; got %v %v %v", name, s.ReportedUncorrect, s.CommandTimeout, s.UDMACRCErrors)
+		}
+		b, _ := json.Marshal(s)
+		for _, k := range []string{"reported_uncorrect", "command_timeout", "udma_crc_errors"} {
+			if strings.Contains(string(b), `"`+k+`"`) {
+				t.Fatalf("%s: an unreported attribute must be omitted, not sent: %s", name, b)
+			}
+		}
+	}
+}

internal/storage/smart.go

@@ -44,6 +44,9 @@ const (
 	ataReallocatedSectorCt = 5
 	ataCurrentPending      = 197
 	ataOfflineUncorrect    = 198
+	ataReportedUncorrect   = 187 // R-330
+	ataCommandTimeout      = 188 // R-330
+	ataUDMACRCErrorCount   = 199 // R-330
 )
 
 // parseSMART maps smartctl JSON to a hub.SmartSummary, handling SATA + NVMe and degrading
@@ -85,6 +88,12 @@ func parseSMART(raw []byte) hub.SmartSummary {
 				s.PendingSectors = intPtr(int(a.Raw.Value))
 			case ataOfflineUncorrect:
 				s.OfflineUncorrectable = intPtr(int(a.Raw.Value))
+			case ataReportedUncorrect:
+				s.ReportedUncorrect = int64Ptr(a.Raw.Value)
+			case ataCommandTimeout:
+				s.CommandTimeout = int64Ptr(a.Raw.Value)
+			case ataUDMACRCErrorCount:
+				s.UDMACRCErrors = int64Ptr(a.Raw.Value)
 			}
 		}
 	}
@@ -142,3 +151,5 @@ func parseThinPoolMetadata(raw []byte) (float64, bool) {
 }
 
 func intPtr(v int) *int { return &v }
+
+func int64Ptr(v int64) *int64 { return &v }

scripts/test_gate_decoys.py

@@ -0,0 +1,367 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""test_gate_decoys.py — can this repo's gates be fooled by a LABEL? (R-421, R-426)
+
+The same instrument as `felhom.eu/scripts/test_gate_decoys.py`: a decoy is the LABEL without the
+FACT, and a gate that passes on the label alone — or refuses the genuine article — is a live hole.
+Every gate is asserted in BOTH directions: the decoy must be convicted, the genuine article passed.
+
+Covered here (the `COVERS` literal is AST-read by `felhom.eu/scripts/decoy_coverage_gate.py`, which
+never imports this file):
+
+  published         check-published-versions.py, against a FAKE Gitea (see below).
+  release-complete  check-release-complete.py, in a scratch clone whose `origin` is a scratch bare
+                    repository, against the same fake Gitea.
+  reuse-refs, instructions, observations
+                    the three SHARED felhom.eu scripts, run against a scratch clone of THIS repo —
+                    so the decoy is planted in the agent's own REUSE.md / CLAUDE.md / REPORT.md and
+                    coverage is per input, not per script.
+
+NEVER THE REAL GITEA. Both network gates read `GITEA_BASE` from the environment (CI already sets it
+to the in-cluster URL); here it points at an `http.server` bound to 127.0.0.1 inside this process,
+and every proxy variable is removed from the child's environment so urllib cannot route around it.
+A test that asked the real registry would pass or fail on whatever was published that day — the
+constant-for-measurement shape — and would reach the network from a hook.
+
+NEVER THE REAL TREE. Every planted file lives in a scratch directory: a workspace that holds a
+clone of this repo beside symlinks to the sibling clones the shared scripts reach across to.
+
+Run from the repo root:  python3 scripts/test_gate_decoys.py
+Exit 0 every decoy judged correctly · 1 a decoy passed or a genuine article was refused.
+"""
+import http.server
+import io
+import json
+import os
+import shutil
+import socketserver
+import subprocess
+import sys
+import tempfile
+import threading
+
+ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+PARENT = os.path.dirname(ROOT)
+# DECOY_SHARED_DIR exists for ONE purpose: the red-proof. It lets a mutated COPY of the shared scripts
+# be judged without editing the felhom.eu clone. Unset, the suite judges the real shared scripts.
+SHARED = os.environ.get("DECOY_SHARED_DIR") or os.path.join(PARENT, "felhom.eu", "scripts")
+
+# ── WHAT THIS FILE COVERS ────────────────────────────────────────────────────────────────────────
+# Read by felhom.eu/scripts/decoy_coverage_gate.py, which AST-parses this literal. A gate named here
+# MUST have a decoy below that has been seen to fail.
+COVERS = {
+    "published": "against a FAKE Gitea: a tag whose package 404s, a tag whose tree lacks the configs, a package one "
+                 "version past the newest tag (published, never tagged), a patch-gap orphan, a missing package that "
+                 "lexical sorting would drop out of the retention window (0.9.x vs 0.10.x); a tags api answering 500 "
+                 "or a non-JSON 200 is INCONCLUSIVE, never a pass - vs a clean registry, a non-semver tag and a version "
+                 "older than the retention window (not asserted, BY DESIGN) (R-426)",
+    "release-complete": "the newest `## vX.Y.Z` with no tag anywhere, a tag parked on an unrelated commit, a tag "
+                        "with no package; a registry 500 is INCONCLUSIVE - vs the genuine release, an `## Unreleased` "
+                        "heading above it, a newer version named only in prose or under `###`, a tag that only "
+                        "origin has (the shallow-CI shape); a LOCAL-only tag passes BY DESIGN (CI's fresh clone "
+                        "and the published gate's converse probe are what see it) (R-426)",
+    "reuse-refs": "a cited .go and a cited .md path that do not exist, planted in THIS repo's REUSE.md - vs the "
+                  "real file (R-426)",
+    "instructions": "a component version literal in THIS repo's CLAUDE.md effective text - vs the same sentence "
+                    "inside an HTML comment (R-426)",
+    "observations": "R-419 in THIS repo's REPORT.md: an Observations note SAYING it carries no marker - vs the "
+                    "two genuine markers (R-426)",
+}
+
+fails = []
+ran = 0
+
+
+def report(name, rc, out, expect_rc, must=()):
+    global ran
+    ran += 1
+    missing = [m for m in must if m not in out]
+    if rc == expect_rc and not missing:
+        print("  ok  %-62s rc=%d (expected %d)" % (name, rc, expect_rc))
+    else:
+        hole = expect_rc != 0 and rc == 0
+        fails.append("%s: rc=%d expected %d%s; missing %s\n%s" % (
+            name, rc, expect_rc, " - LIVE HOLE" if hole else "", missing, out[-900:]))
+
+
+# ── the fake Gitea ───────────────────────────────────────────────────────────────────────────────
+class Fake(object):
+    """What the fake registry serves. Reset per case."""
+
+    def reset(self):
+        self.tags = []            # tag names, as the tags api lists them
+        self.packages = set()     # versions whose generic package downloads
+        self.raw = set()          # versions whose tag tree serves the probe config
+        self.tags_status = 200
+        self.tags_body = None     # override bytes for the tags api
+        self.pkg_status = None    # override status for EVERY package request
+        self.hits = []
+
+
+FAKE = Fake()
+FAKE.reset()
+PKG_PREFIX = "/api/packages/admin/generic/felhom-agent/"
+RAW_PREFIX = "/admin/felhom-agent/raw/tag/v"
+
+
+class Handler(http.server.BaseHTTPRequestHandler):
+    def log_message(self, *a):
+        pass
+
+    def _answer(self, status, body=b""):
+        self.send_response(status)
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        if self.command != "HEAD":
+            self.wfile.write(body)
+
+    def do_GET(self):
+        p = self.path
+        FAKE.hits.append(p)
+        if p.startswith("/api/v1/repos/admin/felhom-agent/tags"):
+            body = FAKE.tags_body if FAKE.tags_body is not None else \
+                json.dumps([{"name": t} for t in FAKE.tags]).encode()
+            return self._answer(FAKE.tags_status, body)
+        if p.startswith(PKG_PREFIX):
+            if FAKE.pkg_status is not None:
+                return self._answer(FAKE.pkg_status)
+            v = p[len(PKG_PREFIX):].split("/", 1)[0]
+            return self._answer(200, b"ELF") if v in FAKE.packages else self._answer(404)
+        if p.startswith(RAW_PREFIX):
+            v = p[len(RAW_PREFIX):].split("/", 1)[0]
+            ok = v in FAKE.raw and p.endswith("/configs/felhom-agent.service")
+            return self._answer(200, b"[Unit]\n") if ok else self._answer(404)
+        return self._answer(404)
+
+    do_HEAD = do_GET
+
+
+class Server(socketserver.ThreadingMixIn, http.server.HTTPServer):
+    daemon_threads = True
+
+
+def child_env(base):
+    env = {k: v for k, v in os.environ.items() if "proxy" not in k.lower()}
+    env["GITEA_BASE"] = base
+    env["NO_PROXY"] = env["no_proxy"] = "127.0.0.1,localhost"
+    return env
+
+
+def run(argv, cwd, env=None):
+    # input="" — a child must never inherit (and block on) this process's stdin
+    p = subprocess.run(argv, cwd=cwd, env=env, capture_output=True, text=True, input="")
+    return p.returncode, p.stdout + p.stderr
+
+
+def sh(argv, cwd):
+    rc, out = run(argv, cwd)
+    if rc != 0:
+        raise SystemExit("setup command failed (%s): %s" % (" ".join(argv), out))
+    return out.strip()
+
+
+# ── published ────────────────────────────────────────────────────────────────────────────────────
+def published_cases(base):
+    gate = os.path.join(ROOT, "scripts", "check-published-versions.py")
+    keep = json.load(io.open(os.path.join(ROOT, "scripts", "retention-policy.json"),
+                             encoding="utf-8"))["generic_versions_kept"]
+    TAGS = ["v0.150.%d" % i for i in range(3)]          # inside any retention window >= 3
+    VERS = [t[1:] for t in TAGS]
+
+    def case(name, setup, expect_rc, must=()):
+        FAKE.reset()
+        FAKE.tags = list(TAGS)
+        FAKE.packages = set(VERS)
+        FAKE.raw = set(VERS)
+        setup()
+        rc, out = run([sys.executable, gate], ROOT, child_env(base))
+        if not FAKE.hits:
+            fails.append("published/%s: the gate never asked the fake Gitea - the seam is not wired" % name)
+        report("published: " + name, rc, out, expect_rc, must)
+
+    case("GENUINE: every tag downloadable and serving its configs", lambda: None, 0,
+         ("ALL RELEASED VERSIONS INSTALLABLE",))
+    case("GENUINE: a non-semver tag is not a release", lambda: FAKE.tags.append("v0.150.2-rc1"), 0,
+         ("ALL RELEASED VERSIONS INSTALLABLE",))
+    case("FACT: a tag whose package 404s", lambda: FAKE.packages.discard("0.150.1"), 1,
+         ("FAIL v0.150.1", "binary NOT downloadable"))
+    case("FACT: a tag whose tree does not serve the configs", lambda: FAKE.raw.discard("0.150.2"), 1,
+         ("FAIL v0.150.2", "does not serve"))
+    case("FACT: published one patch past the newest tag, never tagged",
+         lambda: FAKE.packages.add("0.150.3"), 1, ("PUBLISHED VERSION(S) WITH NO TAG", "v0.150.3"))
+    case("FACT: published in a patch GAP between two tags",
+         lambda: (FAKE.tags.remove("v0.150.1"),), 1, ("v0.150.1 is downloadable", "has no git tag"))
+
+    def lexical():
+        # keep+1 tags: 0.9.0 and 0.10.0..0.10.<keep-1>. By SEMVER the oldest is 0.9.0 (dropped); by
+        # STRING sort "0.10.0" is the smallest and would be the one dropped - so its missing package
+        # is convicted only if the window is cut by semver.
+        FAKE.tags = ["v0.9.0"] + ["v0.10.%d" % i for i in range(keep)]
+        FAKE.packages = set(t[1:] for t in FAKE.tags) - {"0.10.0"}
+        FAKE.raw = set(t[1:] for t in FAKE.tags)
+    case("FACT: a missing package lexical sorting would drop (0.10.0 vs 0.9.0)", lexical, 1,
+         ("FAIL v0.10.0",))
+
+    def retired():
+        FAKE.tags = ["v0.9.0"] + ["v0.10.%d" % i for i in range(keep)]
+        FAKE.packages = set(t[1:] for t in FAKE.tags) - {"0.9.0"}
+        FAKE.raw = set(t[1:] for t in FAKE.tags)
+    case("BY DESIGN: a version older than the retention window is not asserted", retired, 0,
+         ("NOT ASSERTED", "0.9.0"))
+
+    def five_hundred():
+        FAKE.tags_status = 500
+    case("INCONCLUSIVE: the tags api answers 500", five_hundred, 2, ("INCONCLUSIVE",))
+
+    def html():
+        FAKE.tags_body = b"<html>sign in</html>"
+    case("INCONCLUSIVE: the tags api answers a 200 that is not JSON", html, 2, ("INCONCLUSIVE",))
+
+    # an unreachable Gitea: a port nothing listens on
+    s = Server(("127.0.0.1", 0), Handler)
+    dead = "http://127.0.0.1:%d" % s.server_address[1]
+    s.server_close()
+    rc, out = run([sys.executable, gate], ROOT, child_env(dead))
+    report("published: INCONCLUSIVE: Gitea unreachable", rc, out, 2, ("INCONCLUSIVE", "URLs tried"))
+
+
+# ── release-complete ─────────────────────────────────────────────────────────────────────────────
+def release_cases(base, ws):
+    bare = os.path.join(ws, "origin.git")
+    work = os.path.join(ws, "rc-work")
+    sh(["git", "clone", "-q", "--bare", "--no-tags", "file://" + ROOT, bare], ws)
+    sh(["git", "clone", "-q", "--no-tags", "file://" + bare, work], ws)
+    sh(["git", "config", "user.email", "decoy@gate.invalid"], work)
+    sh(["git", "config", "user.name", "decoy"], work)
+    # the WORKING-TREE gate, so the file under test is the one being edited, not HEAD's
+    shutil.copy(os.path.join(ROOT, "scripts", "check-release-complete.py"),
+                os.path.join(work, "scripts", "check-release-complete.py"))
+    sh(["git", "add", "scripts/check-release-complete.py"], work)
+    sh(["git", "commit", "-q", "--allow-empty", "-m", "the gate under test"], work)
+    base_sha = sh(["git", "rev-parse", "HEAD"], work)
+    ch = os.path.join(work, "CHANGELOG.md")
+    original = io.open(ch, encoding="utf-8").read()
+    V = "9.9.9"
+
+    def case(name, top, expect_rc, must=(), tag=None, origin_tag=False, packaged=True, pkg_status=None):
+        FAKE.reset()
+        if packaged:
+            FAKE.packages = {V}
+        FAKE.pkg_status = pkg_status
+        try:
+            io.open(ch, "w", encoding="utf-8").write(top + original)
+            sh(["git", "commit", "-q", "-am", name], work)
+            if tag == "head":
+                sh(["git", "tag", "-a", "v" + V, "-m", "decoy", "HEAD"], work)
+            elif tag == "unrelated":
+                empty = sh(["git", "mktree"], work)   # stdin is "" — the empty tree, written to this repo
+                orphan = sh(["git", "commit-tree", "-m", "unrelated", empty], work)
+                sh(["git", "tag", "-a", "v" + V, "-m", "decoy", orphan], work)
+            if origin_tag:
+                sh(["git", "push", "-q", "origin", "HEAD:refs/tags/v" + V], work)
+            rc, out = run([sys.executable, os.path.join(work, "scripts", "check-release-complete.py")],
+                          work, child_env(base))
+            report("release-complete: " + name, rc, out, expect_rc, must)
+        finally:
+            run(["git", "tag", "-d", "v" + V], work)
+            run(["git", "push", "-q", "origin", ":refs/tags/v" + V], work)
+            sh(["git", "reset", "-q", "--hard", base_sha], work)
+
+    HEAD = "## v%s — 2026-10-06\n\n- decoy release\n\n" % V
+    case("GENUINE: tagged at HEAD and published", HEAD, 0,
+         ("newest CHANGELOG version: v9.9.9", "is tagged, placed and published"), tag="head")
+    case("GENUINE: an `## Unreleased` heading above the release", "## Unreleased\n\n- wip\n\n" + HEAD, 0,
+         ("newest CHANGELOG version: v9.9.9",), tag="head")
+    case("GENUINE: a newer version named only in prose and under ###",
+         "The `## v10.0.0` heading is not written yet.\n### v10.0.0 notes\n\n" + HEAD, 0,
+         ("newest CHANGELOG version: v9.9.9",), tag="head")
+    case("GENUINE: the tag only on origin (the shallow-CI shape)", HEAD, 0,
+         ("exists on origin",), origin_tag=True)
+    case("BY DESIGN: a LOCAL-only tag passes (CI's fresh clone sees only origin)", HEAD, 0,
+         ("an ancestor of HEAD",), tag="head")
+    case("FACT: the newest heading has no tag anywhere", HEAD, 1, ("DOES NOT EXIST",))
+    case("FACT: a tag parked on an unrelated commit", HEAD, 1, ("NOT an ancestor",), tag="unrelated")
+    case("FACT: tagged, never published", HEAD, 1, ("IS NOT PUBLISHED",), tag="head", packaged=False)
+    case("INCONCLUSIVE: the registry answers 500", HEAD, 2, ("INCONCLUSIVE",), tag="head", pkg_status=500)
+    case("FACT beats INCONCLUSIVE: no tag AND the registry answers 500", HEAD, 1, ("DOES NOT EXIST",),
+         pkg_status=500)
+
+
+# ── the shared felhom.eu scripts, against THIS repo's inputs ─────────────────────────────────────
+def shared_cases(ws):
+    """A scratch WORKSPACE: a clone of this repo beside symlinks to the siblings, because the shared
+    scripts reach across (REUSE.md cites hub paths; instructions_gate reads the workspace CLAUDE.md)."""
+    for g in ("reuse_refs_check.py", "instructions_gate.py", "observations_gate.py"):
+        if not os.path.isfile(os.path.join(SHARED, g)):
+            fails.append("shared gate %s is MISSING beside this clone (tried %s) - a failure, never a skip"
+                         % (g, SHARED))
+            return
+    space = os.path.join(ws, "workspace")
+    os.makedirs(space)
+    for entry in sorted(os.listdir(PARENT)):
+        if entry in ("felhom.eu", "felhom-controller", "app-catalog-felhom.eu", "homelab-manifests",
+                     "CLAUDE.md", ".claude-memory"):
+            os.symlink(os.path.join(PARENT, entry), os.path.join(space, entry))
+    repo = os.path.join(space, "felhom-agent")
+    sh(["git", "clone", "-q", "--no-tags", "file://" + ROOT, repo], ws)
+    # the WORKING-TREE inputs the plants go into, so a case judges today's file
+    for f in ("REUSE.md", "CLAUDE.md", "REPORT.md"):
+        shutil.copy(os.path.join(ROOT, f), os.path.join(repo, f))
+
+    def case(name, gate, relpath, extra, expect_rc, must=()):
+        p = os.path.join(repo, relpath)
+        backup = io.open(p, encoding="utf-8").read()
+        try:
+            if extra:
+                io.open(p, "w", encoding="utf-8").write(backup + extra)
+            rc, out = run([sys.executable, os.path.join(SHARED, gate), repo], repo)
+            report(name, rc, out, expect_rc, must)
+        finally:
+            io.open(p, "w", encoding="utf-8").write(backup)
+
+    case("reuse-refs: GENUINE: this repo's REUSE.md", "reuse_refs_check.py", "REUSE.md", "", 0, ("FAILED 0",))
+    case("reuse-refs: FACT: a cited .go path that does not exist", "reuse_refs_check.py", "REUSE.md",
+         u"\n- see `internal/localapi/does_not_exist.go`\n", 1, ("does_not_exist.go",))
+    case("reuse-refs: FACT: a cited .md path that does not exist", "reuse_refs_check.py", "REUSE.md",
+         u"\n- see `docs/99-does-not-exist.md`\n", 1, ("99-does-not-exist.md",))
+    case("instructions: GENUINE: this repo's CLAUDE.md", "instructions_gate.py", "CLAUDE.md", "", 0,
+         ("instructions_gate: OK",))
+    case("instructions: FACT: a version literal in effective text", "instructions_gate.py", "CLAUDE.md",
+         u"\nThe agent runs v0.148.0 today.\n", 1, ("v0.148.0",))
+    case("instructions: GENUINE: the same sentence in an HTML comment", "instructions_gate.py", "CLAUDE.md",
+         u"\n<!--\nThe agent ran v0.148.0 on 2026-10-06.\n-->\n", 0, ("instructions_gate: OK",))
+    case("observations: FACT: R-419, prose SAYING it has no marker", "observations_gate.py", "REPORT.md",
+         u"\n## Observations\n\n1. **A real finding.** It carries no `FILED:` marker and no "
+         u"`NOT-A-FINDING:` marker, deliberately.\n", 1)
+    case("observations: GENUINE: a FILED marker", "observations_gate.py", "REPORT.md",
+         u"\n## Observations\n\n1. **A real finding.** Something broke. **FILED: R-419**\n", 0)
+    case("observations: GENUINE: a NOT-A-FINDING marker", "observations_gate.py", "REPORT.md",
+         u"\n## Observations\n\n1. **A real finding.** Odd. **NOT-A-FINDING: my own typo, corrected in "
+         u"the same minute.**\n", 0)
+
+
+def main():
+    srv = Server(("127.0.0.1", 0), Handler)
+    threading.Thread(target=srv.serve_forever, daemon=True).start()
+    base = "http://127.0.0.1:%d" % srv.server_address[1]
+    ws = tempfile.mkdtemp(prefix="agent-decoys-")
+    print("agent gate decoys — fake Gitea at %s, scratch %s" % (base, ws))
+    try:
+        published_cases(base)
+        release_cases(base, ws)
+        shared_cases(ws)
+    finally:
+        srv.shutdown()
+        srv.server_close()
+        shutil.rmtree(ws, ignore_errors=True)
+    if fails:
+        print()
+        for f in fails:
+            print("FAIL: %s" % f)
+        return 1
+    print("\nagent gate decoys OK — %d case(s), every label judged on its fact (R-421)" % ran)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())