v0.237.0: the Update button takes a backup first, and tells the truth (update arc slice 4 — R-448, R-443, R-439)

POST /api/stacks/{name}/update is now a guarded job answering 202:
cheap refusals (hold — R-439, busy, migration, deploying, memory via the
deploy's own memoryVerdict, a fixed 2 GB disk floor, and no restorable
Tier-2 copy) → backup-first when the proven copy is older than
update.backup_max_age (24h) → safety dump BEFORE the pin moves → pin →
pull (failure puts the pin back) → up → health (.felhom.yml check or 60 s
settle, update.health_timeout 5m). Not healthy → the app is stopped and
HELD (RestoreHold reason update_failed, same store and gate as R-379) and
the page names the backup to restore from; the pin stays. Success is only
ever update_phase=done after health (R-443). UpdateStack is deleted.

The restorable-unit predicate is EXTRACTED to backup.Tier2UnitRestorePoint
and shared with the backups page (row pinned unchanged). The copy is aged
by the last successful Tier-2 copy, not the manifest created_at — measured
on demo-hp that created_at moves only on definition changes.

Crash safety: update-journal.json before each phase; RecoverUpdates before
the boot sweep, ResumeInterruptedUpdates after the guards are wired.

Three unattended start paths ignored a hold and now honour it: the
drive-return gate (restart + boot recreate) and the nightly volume dump.
The nightly capture and Tier-2 run skip held apps so the restore point
survives. No automatic rollback — measured per-app; route back = restore.

Tests A–H across stacks/backup/api/web/cmd; six red-proofs seen to fail.

Claude-Session: https://claude.ai/code/session_0159rPz1ZhFKsS53msqPYxtS

CHANGELOG.md

@@ -1,3 +1,68 @@
+## v0.237.0 — the Update button takes a backup first, and tells the truth (2026-09-13, update arc slice 4 — R-448, R-443, R-439)
+
+**MinAgent: 0.129.0** (unchanged)
+
+**What it replaced.** `UpdateStack` advanced the pin, pulled, ran `up -d` and returned. No copy first,
+no check of memory, disk, a running backup or a held app, and success the moment `up` returned —
+measured in `SPIKE-app-update-2026-09-01` §4 as **HTTP 200 over an app that was already
+crash-looping** (R-443). A held app could be updated at all (R-439). **`UpdateStack` is deleted**; its
+only caller was the API.
+
+**The guarded update** (`internal/stacks/update.go`), a job answering **202** at once, with phases
+`checking → backing-up (only if stale) → safety-dump → pinning → pulling → starting → verifying →
+done | failed` on `GET /api/stacks/{name}` (`updating`, `update_phase`, `update_phase_label`,
+`update_error`, `hold_reason`):
+
+1. **Cheap refusals first, each a 409 with a Hungarian sentence, before the intent is recorded:** held
+   (R-439 — `update` joins `start`/`restart` in the router's hold check), a backup/restore/app-data op
+   or quiesce holding it, a migration, already updating, deploying, memory (the deploy's own check,
+   extracted as `memoryVerdict`, releasing the app's current request), disk (**fixed 2 GB floor** on the
+   Docker data root — the image size is not known without a registry query).
+2. **The precondition is the existing verified backup** (operator ruling 2026-09-02): an openable
+   Tier-2 recovery unit with a PROVEN copy date — `backup.Tier2UnitRestorePoint`, **extracted from the
+   backups page handler, not duplicated** (the page renders identically, pinned). No such copy →
+   refused. Copy older than `update.backup_max_age` (default `24h`) → `RunAppBackupNow` runs this app's
+   own legs (DB dump → volume dump → unit capture → Tier-2) first; it must succeed AND yield a fresh
+   restorable copy, or nothing moves.
+3. **Safety dump before the pin moves** (`WriteUpdateSafetyDump` = R-361's `writeSafetyDump`).
+4. **Pin before pull**; a **pull failure puts the pin and definition BACK** (nothing ran).
+5. **Health, not the exit code:** the app's `.felhom.yml` check through the existing probe, or — with
+   none declared — every container running, none restarting, for 60 s; bounded by
+   `update.health_timeout` (default `5m`). Not healthy → **the app is stopped and HELD**
+   (`settings.RestoreHold`, new `reason: update_failed` + `copy_date`, same store and same gate as
+   R-379), the pin stays on the new version, and the page carries the hold sentence naming the backup
+   to restore from. A successful unit restore lifts an update hold (only that kind).
+
+**Measured before it was designed, and the design changed because of it.** The copy's age is the last
+SUCCESSFUL Tier-2 copy, not the unit manifest's `created_at`: on demo-hp bookstack's mirror held a
+database dump from 2026-09-13T00:30Z under a manifest dated 2026-09-12T02:15:29Z, because a capture
+rewrites the manifest only when the app's DEFINITION changes. Aged by the manifest, a quiet app would
+be "stale" forever and "back up first" would not fix it.
+
+**Crash safety is a journal** (`<data>/update-journal.json`, fsynced, written before each phase).
+`RecoverUpdates` runs before the boot sweep: interrupted before the pin → dropped; while pinning or
+pulling → pin put back; after `up` → the app is marked Updating (the boot sweep and the dead-app alarm
+leave it alone) and `ResumeInterruptedUpdates` re-runs `up` + the health wait once the backup side is
+wired, ending healthy or held.
+
+**"A hold that only one path honours is not a hold" — three unattended paths did not honour one, and
+now do:** the drive-return gate (`intermediary.go` restart + boot recreate), and the nightly volume
+dump (`DumpAppVolumesSafe` ends in `StartStack`). The nightly capture and Tier-2 run also skip a held
+app, so they cannot overwrite the restore point the hold text names.
+
+**Deliberately NOT built:** putting the old version back automatically. Measured per-app
+(`SPIKE-upgrade-test-2026-09-06`) and unpredictable; the route back is the restore. The multi-major
+jump is not gated here (Slice 6); it fails health and is held honestly.
+
+**Tests.** stacks: A–G (success only after health, stale → backup first, backup failure, no unit,
+seven cheap refusals, pull failure pin-back, health failure hold, unsaved hold, three recovery
+shapes), hold text on GetStacks, the exact phase labels. backup: proven copy time (incl. the measured
+bookstack case), hold text, restore clears only update holds, busy, nightly legs skip held apps. api:
+R-439, R-443 (202, never "completed"), no-backup 409 records no intent. web: backup row unchanged by the
+extraction, drive-return skips a held app. cmd: guards wired, recovery before the boot sweep, resume
+after the guards, boot gate + dead-app alarm know about updates. **Six companion red-proofs, each seen
+to fail** — three first ran INERT or did not compile and were fixed before they counted (REPORT.md).
+
 ## v0.236.0 — "delete my data too" deletes the data, or says that it could not (2026-09-13, R-442)
 
 **The defect.** A customer removes an app and ticks *also delete my data*. The box says it worked;

CONTEXT.md

@@ -7,7 +7,17 @@
 >
 > Ask Claude Code: "Please update CONTEXT.md with what we did today"
 
-Last updated: 2026-09-13 (v0.236.0 — R-442: "delete my data too" deletes it or refuses)
+Last updated: 2026-09-13 (v0.237.0 — update arc slice 4: the guarded update)
+
+> **2026-09-13 — v0.237.0 (slice 4: R-448, R-443, R-439).** Update is a guarded 202 job:
+> refusals (hold, busy, memory, disk, no restorable Tier-2 copy) → backup-first if the proven copy is
+> older than `update.backup_max_age` (24h) → safety dump → pin → pull (failure: pin BACK) → up →
+> health (`.felhom.yml` check or 60 s settle, `update.health_timeout` 5m; failure: stop + HOLD,
+> `RestoreHold.Reason=update_failed`, pin stays). `UpdateStack` deleted. **The copy is aged by the last
+> successful Tier-2 copy, NOT the manifest `created_at`** — measured on demo-hp that `created_at` moves
+> only on definition changes. Journal `update-journal.json`; `RecoverUpdates` before the boot sweep.
+> Three unattended paths that ignored a hold now honour it (drive-return gate ×2, nightly volume dump);
+> capture and Tier-2 skip held apps. **No automatic rollback** — measured per-app; route back = restore.
 
 > **2026-09-13 — v0.236.0 (R-442).** Removal resolves the drive from the app's OWN `app.yaml`
 > `HDD_PATH` (the `07` ~L437 rule), never the global `cfg.Paths.HDDPath` (set on no box); a data

REUSE.md

@@ -109,7 +109,7 @@
 | `Manager.PersistUnitRedeployConfig` (R-47, v0.153.0) | controller/internal/stacks/deploy.go | `(name, env map[string]string) error` | the PERSIST half of `RedeployFromEnv` — app.yaml + locked fields + in-memory flags, **starts nothing** | **TRAP: the restore paths must use THIS, never `RedeployFromEnv`.** RedeployFromEnv ends in a full `up -d`, which before the replay IS the H4 race. RedeployFromEnv is now literally this + the unchanged up-and-report tail |
 | `Manager.StartStackServices` (R-47, v0.153.0) | controller/internal/stacks/manager.go | `(name string, services []string) error` | scoped `compose up -d <svc>...` — the DB-only window a dump is replayed in | **REFUSES an empty list** (argument-less `up -d` is a FULL start — the one silent fall-through that would reintroduce the race). No `logPostStartStatus`: the app containers are absent on purpose. Never `RestartStack` here — it is a full up in disguise |
 | `appbackup.DBServiceNames` / `dbTypeForImage` (R-47, v0.153.0) | controller/internal/appbackup/dbservices.go | `(composePath string) ([]string, error)` | naming the compose SERVICE(s) holding a database, sorted | yaml.v3 `services:` MAP parse — **never a line scan** (immich's top-level `immich_ml_cache:` / `immich_postgres_data:` volume keys look exactly like services). `dbTypeForImage` is shared with `DiscoverDatabases`, which is what makes "a dump exists ⇒ a service can be named" hold. An error means CANNOT-TELL, never "no database" — callers refuse when a dump exists |
-| `Manager.StartStack/StopStack/RestartStack/UpdateStack` | controller/internal/stacks/manager.go | `(name string) error` | Lifecycle | Protected stacks refuse stop; all funnel through composeExec. **NOT writers of desired state (R-166)** — 14 call sites, only 2 are the customer; recording intent here would make a nightly backup indistinguishable from the customer pressing Stop. Use `SetDesiredState` at the intent point instead |
+| `Manager.StartStack/StopStack/RestartStack` | controller/internal/stacks/manager.go | `(name string) error` | Lifecycle | **`UpdateStack` was DELETED in v0.237.0 — use `StartGuardedUpdate` (update.go).** Protected stacks refuse stop; all funnel through composeExec. **NOT writers of desired state (R-166)** — 14 call sites, only 2 are the customer; recording intent here would make a nightly backup indistinguishable from the customer pressing Stop. Use `SetDesiredState` at the intent point instead |
 | `Manager.SetDesiredState` / `DesiredStateOf` / `BackfillDesiredState` (R-166, v0.189.0) | controller/internal/stacks/desiredstate.go | `(name, desired string) error` / `(Stack) string` / `() int` | THE customer-intent record — `app.yaml` `desired_state`, tri-state `""`/`running`/`stopped` | **ONE OWNER: the customer's action.** Writers are the API action switch, `DeployStack`, `UpdateOptionalConfig`'s redeploy branch, and the `.fab` restore adapter — nothing else, ever. **`""` (absent) means UNKNOWN, never "running"**: every pre-v0.189.0 app.yaml reads absent, so treating it as running would start every deliberately-stopped app on upgrade. Write intent BEFORE the act and REFUSE the act if it fails (§8.2). Backfill is **running-only** — never infer `stopped` from zero containers, that inference IS the defect |
 | `Manager.DriveLive` (R-171, v0.190.0) | controller/internal/stacks/deploy.go | `(hddPath string) bool` | is an app's data drive a live mountpoint RIGHT NOW | Wraps the **same** `isMountPoint` seam the userdata belt uses (`manager.go`) — never write a second liveness check, the two would drift invisibly. The system/local path is legitimately not a mountpoint and returns true |
 | `bootrecon.StartGate` (R-171, v0.190.0) | controller/internal/bootrecon/bootrecon.go | `MayStart(stack) (bool, reason)` | THE one question the boot sweep asks before starting anything | **Fail-safe: cannot determine ⇒ return FALSE.** One seam for all three holders (absent drive · quiesce · an in-flight app-data operation) because they differ only in the reason string. Implemented in `main.go` (`bootDriveGate`) reusing `quiesce.SuppressedStacks()`, `AppStopGuard.HeldStacks()` and `Manager.DriveLive` — never re-derive any of them. Held apps go to `Result.HeldByDrive`, **never** `StillDown` (that is the dead-app alarm's bucket) |
@@ -120,7 +120,12 @@
 | `Manager.DeleteStack` / `RemoveStack` | controller/internal/stacks/delete.go | `(name, removeHDDData[, backupPaths])` | THE guarded removal paths | Orphan/protected/deploying/running checks + ProtectedHDDPaths filter before any RemoveAll. **R-442 (v0.236.0): the drive is the app's OWN `app.yaml` `HDD_PATH` (`appHDDPath`), never `cfg.Paths.HDDPath`; a data removal that cannot be resolved returns a typed `*RemoveRefusedError` BEFORE `compose down` — handlers `errors.As` it to 409 + `Message`** |
 | `resolveContainerState` / `aggregateState` | controller/internal/stacks/manager.go | `(dockerState, dockerStatus)` / `([]ContainerInfo)` | State classification | `.State` says "running" even when unhealthy — `.Status` parse is the fix |
 | `Manager.recordInstalledImages` (v0.233.0) | controller/internal/stacks/installed.go | `(name, stackDir string, env []string)` | writing down what each compose SERVICE is ACTUALLY running, into `app.yaml.installed_images` | Called after a successful compose up from `StartStack`/`RestartStack`/`UpdateStack`/`runComposeDeploy`. **Reads the CONTAINER, never `docker-compose.yml`** — that file is the value the syncer has already moved (spike §3: 25 minutes of disagreement). **A failed write NEVER refuses the action** — the deliberate OPPOSITE of `SetDesiredState`: intent refused, observation logged at ERROR. **NOT from `StartStackServices`** (the R-47 DB-only window would overwrite a complete record with a partial one). Skips the write when ref+digest are unchanged, and carries `at` forward so it means "running since". Its OWN seam (`installedExecFn`) with a **context + 30 s timeout** — the two existing exec helpers have neither |
-| `Manager.SetPin` / `AdoptPins` / `RenderPlanFor` / `AppliedComposePath` (v0.235.0) | controller/internal/stacks/pin.go | `SetPin(name, stackDir, pin, composeSrc) error` | THE version freeze — `app.yaml.pinned_images` + the stored `applied-compose.yml` | **`PinnedImages` is INTENT, `InstalledImages` is an OBSERVATION — never feed one from the other** (the R-166 category error, one field over). Four writers only: deploy, `UpdateStack` (via `advancePinToCatalog`, which advances the pin and re-renders BEFORE the pull, and REFUSES the update if the pin cannot be written), the restore adapter (this is what closes R-441), and `AdoptPins`. `AdoptPins` reuses `observationCoversTemplate` — do NOT write a second completeness rule — and skips loudly rather than inventing a pin. Absent pin = pre-v0.235.0 behaviour |
+| `Manager.SetPin` / `AdoptPins` / `RenderPlanFor` / `AppliedComposePath` (v0.235.0) | controller/internal/stacks/pin.go | `SetPin(name, stackDir, pin, composeSrc) error` | THE version freeze — `app.yaml.pinned_images` + the stored `applied-compose.yml` | **`PinnedImages` is INTENT, `InstalledImages` is an OBSERVATION — never feed one from the other** (the R-166 category error, one field over). Four writers only: deploy, the guarded update (via `advancePinToCatalog`, which advances the pin and re-renders BEFORE the pull, and REFUSES the update if the pin cannot be written; a failed pull puts the pin BACK via `SetPin`), the restore adapter (this is what closes R-441), and `AdoptPins`. `AdoptPins` reuses `observationCoversTemplate` — do NOT write a second completeness rule — and skips loudly rather than inventing a pin. Absent pin = pre-v0.235.0 behaviour |
+| `Manager.UpdatePreflight` / `StartGuardedUpdate` / `RecoverUpdates` / `ResumeInterruptedUpdates` (v0.237.0) | controller/internal/stacks/update.go | `UpdatePreflight(name) *UpdateRefusal`; `StartGuardedUpdate(name) error` | THE update — refusals, then a 202 job with phases on `Stack.Updating/UpdatePhase/UpdateError` | **Never report an update complete before health is known (R-443).** Every cheap refusal runs BEFORE the intent write. Safety dump BEFORE the pin moves; pin BEFORE pull; pull failure → pin back; health failure → stop + HOLD, pin stays. Journal-before-mutate (`update-journal.json`); `RecoverUpdates` MUST run before the boot sweep and `ResumeInterruptedUpdates` AFTER `SetUpdateGuards`. Seams: `updateComposeFn`, `updateHealthFn`, `updateMemoryFn`, `updateDiskFreeFn`, `updateNowFn` (R-457: the age check and the test read ONE clock). Unwired guards ⇒ every update refused |
+| `stacks.UpdateGuards` + `updateGuardsAdapter` (v0.237.0) | controller/internal/stacks/update.go, controller/cmd/controller/main.go | `HoldFor`, `Busy`, `RestorePoint`, `BackupNow`, `SafetyDump`, `HoldAfterFailedUpdate` | the ONLY bridge from the update job to the backup side (stacks cannot import backup) | Wired by `stackMgr.SetUpdateGuards` — pinned by `TestSlice4_UpdateGuardsAreWiredAtStartup`. Add a guard HERE, never by importing backup into stacks |
+| `backup.Manager.Tier2UnitRestorePoint` + `Tier2RestorePoint.ProvenCopyTime` (v0.237.0) | controller/internal/backup/update_guard.go | `(stack) (Tier2RestorePoint, error)` | "can this app be restored from Tier 2, and from when" — the predicate that gates BOTH the „Teljes visszaállítás" action and an update | **ONE predicate, two callers** (extracted from `buildAppBackupRows`, not copied). `CopyDate` is what the page NAMES (the package date, R-403); `ProvenCopyTime` is how OLD the data is — the last successful copy, because the manifest's `created_at` moves only when the DEFINITION changes (measured: a fresh dump under a 22-h-older manifest). Do not age a copy by `CopyDate` |
+| `backup.Manager.HoldAfterFailedUpdate` / `RunAppBackupNow` / `WriteUpdateSafetyDump` / `UpdateBusy` (v0.237.0) | controller/internal/backup/update_guard.go | see file | the update's hold, per-app backup-now, safety dump, busy check | The hold is `settings.RestoreHold` with `Reason: update_failed` — SAME store and gate as R-379, never a second map. `RunAppBackupNow` composes the nightly legs for ONE app (admission, DB dump, volume dump, capture, Tier-2) — do not write a second backup orchestration. A successful unit restore lifts an UPDATE hold only |
+| `stacks.Manager.memoryVerdict` (v0.237.0) | controller/internal/stacks/deploy.go | `(newReq, newLimit, releasedReq, releasedLimit int) (refusal, warning string)` | the deploy's memory check, shared with the update | An update RELEASES the app's current request first. Deploy passes `0, 0` and is byte-identical in wording and log line |
 | `Syncer.SetRenderPlanFn` + `renderSource` (v0.235.0) | controller/internal/sync/sync.go | `func(appName string) stacks.RenderPlan` | the catalog render table | **NIL-SAFE: no seam = copy verbatim = the old product.** Catalog images == pin → verbatim (fixes flow + self-healing, both deliberately kept); differ → the WHOLE stored definition, **never a ref substitution into a newer template** (`wger 2.6`). `.felhom.yml` always verbatim (R-458). The syncer must NEVER read app.yaml. Re-reads the applied file before writing it — a test caught it writing an empty compose over a live app |
 | `Stack.CatalogImages` vs `Stack.TemplateImages` (v0.235.0) | controller/internal/stacks/manager.go | both `map[string]string` | badge input vs "what the next `up -d` gives this app" | **THE TRAP: same type, same shape, opposite meaning after the freeze.** `TemplateImages` reads the LIVE (possibly frozen) compose file; `CatalogImages` reads the syncer's clone. `web.compareInstalledToTemplate` MUST use `CatalogImages` or it answers „Naprakész" on exactly the apps that are behind, with every test green. Red-proved |
 | `Manager.BackfillInstalledImages` (v0.234.0) | controller/internal/stacks/installed.go | `() int` | seeding `installed_images` for apps that have NO record — call ONCE at startup | Beside `BackfillDesiredState` in `cmd/controller/main.go`, after it and BEFORE the boot reconciler (pinned by an AST-walking test that asserts the ORDER). **READS only** — starts nothing, writes no compose file. **Never overwrites an existing record** (an app that has one is not even observed). **REFUSES a partial observation** (`observationCoversTemplate`): `web.compareInstalledToTemplate` reads a service-count mismatch as BEHIND, so seeding a degraded app from what is visible renders „Frissítés elérhető" over an app that is current. The bring-up paths may write a partial because they follow a SUCCESSFUL `up -d` where a gap is real news; a backfill meets any state and must be stricter |

controller/README.md

@@ -472,7 +472,7 @@ through them: `EffectiveLifecycle()`, `CanInstall()`, `IsAbandoned()`.
 Two additions, and **neither changes how an update behaves**. Slice 1 is a record; slice 2 is a label.
 
 **Slice 1 — `app.yaml` gains `installed_images`.** After every successful `compose up` from
-`StartStack`, `RestartStack`, `UpdateStack` and the deploy path, `Manager.recordInstalledImages`
+`StartStack`, `RestartStack`, the guarded update (v0.237.0; `UpdateStack` before it) and the deploy path, `Manager.recordInstalledImages`
 (`internal/stacks/installed.go`) reads what each container is ACTUALLY running and writes it down,
 **keyed by compose SERVICE name**:
 
@@ -527,8 +527,8 @@ the current template pins and returns a `*MetaBadge` rendered by the existing `m
   page. **Known limitation:** for the 23 floating pins (`postgres:16-alpine`, `mariadb:11.6`, …) the
   reference can be identical while the image behind it has moved, so those apps can read „Naprakész"
   when they may not be. Digest-level comparison needs a registry query and is deferred.
-- **Information only.** The badge is wired to no action; `Frissítés`/`Újraindítás`/`Leállítás` are
-  byte-identical to before (`TestScenarioE_TheUpdateButtonIsUntouched`).
+- **Information only.** The badge is wired to no action. (What the `Frissítés` button itself does
+  changed in v0.237.0 — see "The guarded update" below.)
 
 Reasoning and the seven-slice plan: `felhom.eu/documentation/architecture/09-update-architecture.md`.
 
@@ -539,7 +539,7 @@ deliberate Update moves it. Everything else in a template — health checks, mem
 fields — still arrives on the 15-minute cycle, and a broken definition still repairs itself.
 
 - **`app.yaml` gains `pinned_images`** (service → ref): what the app is SUPPOSED to run. **Not**
-  `installed_images`, which is an observation. Written only by the deploy path, `UpdateStack`, a
+  `installed_images`, which is an observation. Written only by the deploy path, the guarded update, a
   restore, and the one-time `AdoptPins`. **Absent = unpinned = pre-v0.235.0 behaviour.**
 - **`applied-compose.yml`** in the stack dir stores the exact definition the pin came from. The
   syncer copies only `docker-compose.yml` and `.felhom.yml`, so that name is safe.
@@ -555,6 +555,50 @@ fields — still arrives on the 15-minute cycle, and a broken definition still r
 
 Reasoning: `felhom.eu/documentation/architecture/09-update-architecture.md` §3, §5.
 
+#### The guarded update (v0.237.0 — update arc slice 4)
+
+**`POST /api/stacks/{name}/update` no longer updates on the spot.** It refuses what it must, starts a
+job, and answers **202**. The page polls `GET /api/stacks/{name}`.
+
+| field | meaning |
+|---|---|
+| `updating` | a guarded update is in progress |
+| `update_phase` / `update_phase_label` | `checking`, `backing-up`, `safety-dump`, `pinning`, `pulling`, `starting`, `verifying`, `done`, `failed` — and the Hungarian label for each |
+| `update_error` | the customer sentence when the update did not complete |
+| `hold_reason` | the hold's sentence while the app is held (failed update OR failed restore) |
+
+**Refused with 409 before anything moves:** held; a backup, restore, app-data op or quiesce holding
+the app; a migration; already updating; deploying; not enough memory for the NEW template's request
+(the deploy's own `memoryVerdict`, releasing the app's current request); less than **2 GB** free on
+the Docker data root (a fixed floor — image sizes are not known without a registry query); and **no
+restorable backup** (no openable Tier-2 recovery unit with a proven copy).
+
+**The sequence.** A proven copy older than `update.backup_max_age` is refreshed first
+(`RunAppBackupNow`: this app's DB dump, volume dump, unit capture, Tier-2 copy). Then a database
+safety dump, then the pin moves, then pull, `up`, and the health wait (`.felhom.yml` check, or 60 s of
+every container running for an app with none; bounded by `update.health_timeout`). **A failed pull
+puts the pin back. An app that does not become healthy is stopped and HELD** — the pin stays on the
+new version, and the hold sentence names the backup it can be restored from. A successful unit
+restore lifts an update hold.
+
+**Config (`controller.yaml`):**
+
+```yaml
+update:
+  backup_max_age: 24h   # a proven Tier-2 copy older than this is refreshed before the update
+  health_timeout: 5m    # how long the new version has to become healthy before the app is held
+```
+
+**Crash safety:** `<data>/update-journal.json` is written before every phase; `RecoverUpdates` (before
+the boot sweep) puts a pin back or marks an interrupted update for `ResumeInterruptedUpdates`.
+
+**Every unattended start path honours a hold:** the boot sweep and the app-stop guard (as before), and
+since v0.237.0 the drive-return gate and the nightly volume dump. The nightly capture and Tier-2 run
+skip a held app so its restore point is not overwritten.
+
+**Not done, deliberately:** the old version is never put back automatically — whether that works is
+per-app and was measured unpredictable. Reasoning: `felhom.eu/documentation/architecture/09-update-architecture.md` §6.
+
 #### App Info Pages
 
 Each app can define rich metadata in `.felhom.yml`:

controller/cmd/controller/main.go

@@ -476,9 +476,20 @@ func main() {
 	// R-171: hand the boot sweep the settings it needs to answer "is this app's drive live?" BEFORE
 	// the goroutine starts — an unwired gate is silently the pre-v0.190.0 behaviour that started apps
 	// onto absent drives. TestMainWiresBootDriveGate walks this file's AST for the assignment.
+	// --- Slice 4: guarded-update crash recovery (Scenario G) ---
+	// BEFORE the boot reconciler, and the order is load-bearing: an update interrupted after `up` is
+	// marked Updating here, and bootDriveGate refuses an Updating app — started after the sweep, the
+	// sweep could bring up a half-updated app with no record of why. Pin-backs for updates interrupted
+	// before anything ran happen here too. The resumed health wait itself is launched further down,
+	// once the backup side is wired, because a resumed update that fails must be able to HOLD.
+	if resumed := stackMgr.RecoverUpdates(); len(resumed) > 0 {
+		logger.Printf("[WARN] [update] %d interrupted update(s) will resume after the backup side is wired: %v", len(resumed), resumed)
+	}
+
 	bootDriveSettings = sett
 	bootQuiesceLoop = quiesceLoop
 	bootAppStopGuard = appStopGuard
+	bootStackMgr = stackMgr
 	go runBootReconcile(ctx, stackMgr, logger)
 
 	// --- Start CPU collector ---
@@ -537,6 +548,16 @@ func main() {
 		backupMgr.SetSharesReconciler(stackMgr.ReconcileSamba)
 	}
 
+	// --- Slice 4: the guarded update's backup side ---
+	// Wired unconditionally: with backup disabled the adapter answers "no restore point" and every
+	// update is refused with the no-backup sentence — the precondition cannot be met, which is true.
+	// An UNWIRED manager also refuses (fail closed); TestSlice4_UpdateGuardsAreWiredAtStartup walks
+	// this file for the call, because a seam built and never wired has shipped here seven times.
+	stackMgr.SetUpdateGuards(&updateGuardsAdapter{b: backupMgr, q: quiesceLoop})
+	if n := stackMgr.ResumeInterruptedUpdates(ctx); n > 0 {
+		logger.Printf("[WARN] [update] resumed %d interrupted update(s)", n)
+	}
+
 	// SLICE 2: the offsite apply-bridge is launched further down, AFTER the self-updater is constructed
 	// (R-71a: the bridge's settle-gate reads the updater's floor/update-running state to defer the
 	// consume past a managed day-0 floor-update). See "offsite apply-bridge" below.
@@ -1906,6 +1927,7 @@ var (
 	bootDriveSettings *settings.Settings
 	bootQuiesceLoop   *quiesce.Loop
 	bootAppStopGuard  *backup.AppStopGuard
+	bootStackMgr      *stacks.Manager
 )
 
 // bootDriveGate answers bootrecon.StartGate for the real controller. It enforces §8.2: an app that
@@ -1943,6 +1965,11 @@ func (g bootDriveGate) MayStart(stackName string) (bool, string) {
 	if bootQuiesceLoop.SuppressedStacks()[stackName] {
 		return false, "a whole-guest backup (quiesce) is holding it — the quiesce loop restarts its own stacks"
 	}
+	// 1b. slice 4: a guarded update is moving this app (including one RecoverUpdates marked for
+	//     resumption). The update job owns the bring-up and ends in healthy or held.
+	if bootStackMgr != nil && bootStackMgr.IsUpdating(stackName) {
+		return false, "a guarded update is in progress — the update brings it up and verifies it"
+	}
 	// 2. an app-data operation in flight
 	for _, held := range bootAppStopGuard.HeldStacks() {
 		if held == stackName {
@@ -1985,6 +2012,9 @@ func (g driveStartGate) MayStart(stackName string) (bool, string) {
 	// app-stop guard's Recover reaches it directly. A hold only one path honours is not a hold.
 	if g.sett != nil {
 		if h, ok := g.sett.GetRestoreHold(stackName); ok {
+			if h.Reason == settings.HoldReasonUpdateFailed {
+				return false, "held after a failed update (" + h.At + ") — restore it from its backup to start it"
+			}
 			return false, "held after a failed restore whose rollback also failed (" + h.At + ") — clear the hold to start it"
 		}
 	}
@@ -2236,7 +2266,11 @@ func scanDeployedAppRunStates(mgr *stacks.Manager, q *quiesce.Loop, g *backup.Ap
 	// `g` covers the per-app operations — the nightly volume dump, an off-site reconstitution and a
 	// .fab export. Both are nil-safe, and the union is taken here rather than inside classifyRunStates
 	// so that pure function keeps its single `quiesced` parameter and its existing tests.
-	return classifyRunStates(mgr.GetStacks(), unionSuppressed(q.SuppressedStacks(), g.SuppressedStacks()), q.FailedRestarts(), time.Now())
+	// Slice 4: a THIRD mechanism moves an app on purpose — the guarded update recreates it and waits
+	// for health, and it ends in healthy or HELD. Counting it as dead mid-update would be R-330's false
+	// alarm one mechanism over.
+	suppressed := unionSuppressed(unionSuppressed(q.SuppressedStacks(), g.SuppressedStacks()), mgr.UpdatingStacks())
+	return classifyRunStates(mgr.GetStacks(), suppressed, q.FailedRestarts(), time.Now())
 }
 
 // unionSuppressed merges the suppression sets of the two mechanisms that stop apps on purpose.
@@ -3313,3 +3347,61 @@ func integrityOKMsg(res backup.IntegrityResult) string {
 	}
 	return msg + ")"
 }
+
+// updateGuardsAdapter implements stacks.UpdateGuards over the backup manager (slice 4). The stacks
+// package cannot import backup, so this is the one place the two meet. Nil-safe on b: a box with
+// backup disabled has no restore point, and the update is refused for that true reason.
+type updateGuardsAdapter struct {
+	b *backup.Manager
+	q *quiesce.Loop
+}
+
+func (a *updateGuardsAdapter) HoldFor(name string) (bool, string) {
+	if a.b == nil {
+		return false, ""
+	}
+	return a.b.RestoreHoldFor(name)
+}
+
+func (a *updateGuardsAdapter) Busy(name string) (bool, string) {
+	if a.q.SuppressedStacks()[name] {
+		return true, "a whole-guest backup (quiesce) is holding it"
+	}
+	if a.b == nil {
+		return false, ""
+	}
+	return a.b.UpdateBusy(name)
+}
+
+func (a *updateGuardsAdapter) RestorePoint(name string) (stacks.UpdateRestorePoint, error) {
+	if a.b == nil {
+		return stacks.UpdateRestorePoint{}, fmt.Errorf("backup is not enabled on this box")
+	}
+	rp, err := a.b.Tier2UnitRestorePoint(name)
+	if err != nil {
+		return stacks.UpdateRestorePoint{}, err
+	}
+	at, proven := rp.ProvenCopyTime()
+	return stacks.UpdateRestorePoint{Restorable: rp.Restorable, Proven: proven, ProvenAt: at}, nil
+}
+
+func (a *updateGuardsAdapter) BackupNow(ctx context.Context, name string) error {
+	if a.b == nil {
+		return fmt.Errorf("backup is not enabled on this box")
+	}
+	return a.b.RunAppBackupNow(ctx, name)
+}
+
+func (a *updateGuardsAdapter) SafetyDump(ctx context.Context, name string) ([]string, error) {
+	if a.b == nil {
+		return nil, fmt.Errorf("backup is not enabled on this box")
+	}
+	return a.b.WriteUpdateSafetyDump(ctx, name)
+}
+
+func (a *updateGuardsAdapter) HoldAfterFailedUpdate(name string, at, provenCopyAt time.Time) error {
+	if a.b == nil {
+		return fmt.Errorf("backup is not enabled on this box — the hold cannot be recorded")
+	}
+	return a.b.HoldAfterFailedUpdate(name, at, provenCopyAt)
+}

controller/cmd/controller/slice4_wiring_test.go

@@ -0,0 +1,104 @@
+package main
+
+import (
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"strings"
+	"testing"
+
+	"gitea.dooplex.hu/admin/felhom-controller/internal/settings"
+)
+
+// Update arc slice 4 — the seams are WIRED, in the order that makes them true. A seam built and never
+// wired has shipped in this project seven times; each assertion here fails if a line moves or goes.
+
+func slice4CallLines(t *testing.T) (map[string][]int, *ast.File, *token.FileSet) {
+	t.Helper()
+	fset := token.NewFileSet()
+	f, err := parser.ParseFile(fset, "main.go", nil, 0)
+	if err != nil {
+		t.Fatal(err)
+	}
+	lines := map[string][]int{}
+	ast.Inspect(f, func(n ast.Node) bool {
+		if call, ok := n.(*ast.CallExpr); ok {
+			if sel, ok := call.Fun.(*ast.SelectorExpr); ok {
+				lines[sel.Sel.Name] = append(lines[sel.Sel.Name], fset.Position(call.Pos()).Line)
+			} else if id, ok := call.Fun.(*ast.Ident); ok {
+				lines[id.Name] = append(lines[id.Name], fset.Position(call.Pos()).Line)
+			}
+		}
+		return true
+	})
+	return lines, f, fset
+}
+
+func TestSlice4_UpdateGuardsAreWiredAtStartup(t *testing.T) {
+	lines, _, _ := slice4CallLines(t)
+	if len(lines["SetUpdateGuards"]) == 0 {
+		t.Fatal("SetUpdateGuards is never called — every update would be refused as unwired")
+	}
+	if len(lines["RecoverUpdates"]) == 0 || len(lines["runBootReconcile"]) == 0 {
+		t.Fatal("RecoverUpdates or the boot reconciler is missing from main.go")
+	}
+	if lines["RecoverUpdates"][0] > lines["runBootReconcile"][0] {
+		t.Errorf("RecoverUpdates (line %d) must run BEFORE the boot reconciler (line %d), or the sweep can start a half-updated app",
+			lines["RecoverUpdates"][0], lines["runBootReconcile"][0])
+	}
+	if len(lines["ResumeInterruptedUpdates"]) == 0 || lines["ResumeInterruptedUpdates"][0] < lines["SetUpdateGuards"][0] {
+		t.Error("ResumeInterruptedUpdates must run AFTER SetUpdateGuards — a resumed update that fails must be able to HOLD")
+	}
+}
+
+func funcBodySource(t *testing.T, f *ast.File, fset *token.FileSet, recv, name string) string {
+	t.Helper()
+	for _, d := range f.Decls {
+		fn, ok := d.(*ast.FuncDecl)
+		if !ok || fn.Name.Name != name || fn.Body == nil {
+			continue
+		}
+		if recv != "" {
+			if fn.Recv == nil {
+				continue
+			}
+			id, ok := fn.Recv.List[0].Type.(*ast.Ident)
+			if !ok || id.Name != recv {
+				continue
+			}
+		}
+		var names []string
+		ast.Inspect(fn.Body, func(n ast.Node) bool {
+			if sel, ok := n.(*ast.SelectorExpr); ok {
+				names = append(names, sel.Sel.Name)
+			}
+			return true
+		})
+		return strings.Join(names, " ")
+	}
+	t.Fatalf("func %s.%s not found", recv, name)
+	return ""
+}
+
+func TestSlice4_BootSweepAndDeadAppAlarmKnowAboutUpdates(t *testing.T) {
+	_, f, fset := slice4CallLines(t)
+	if !strings.Contains(funcBodySource(t, f, fset, "bootDriveGate", "MayStart"), "IsUpdating") {
+		t.Error("the boot sweep must refuse an app a guarded update is moving")
+	}
+	if !strings.Contains(funcBodySource(t, f, fset, "", "scanDeployedAppRunStates"), "UpdatingStacks") {
+		t.Error("the dead-app alarm must not count an app the update itself is recreating")
+	}
+}
+
+// The shared start gate names WHICH hold it is honouring — an operator reading the boot log must not
+// be told a restore failed when an update did.
+func TestSlice4_DriveStartGate_NamesAnUpdateHold(t *testing.T) {
+	sett := holdTestSettings(t)
+	if err := sett.SetRestoreHold(settings.RestoreHold{Stack: "bookstack", At: "2026-09-13T08:00:00Z", Reason: settings.HoldReasonUpdateFailed}); err != nil {
+		t.Fatal(err)
+	}
+	ok, why := driveStartGate{sett: sett}.MayStart("bookstack")
+	if ok || !strings.Contains(why, "held after a failed update") {
+		t.Errorf("ok=%v why=%q", ok, why)
+	}
+}

controller/internal/api/router.go

@@ -579,7 +579,11 @@ func (r *Router) actionStack(w http.ResponseWriter, action, name string) {
 	// R-379/R-380: an app held after a failed restore + failed rollback must not start from the
 	// customer's button either. Checked BEFORE the drive gate because it applies to driveless apps,
 	// which is the class the hold exists for.
-	if action == "start" || action == "restart" {
+	//
+	// R-439 (slice 4): `update` is in this list. It was not until v0.237.0, so a held app could be
+	// updated — the one action most likely to make a held app's data worse. Pinned by
+	// TestR439_UpdateOfAHeldAppIsRefused.
+	if action == "start" || action == "restart" || action == "update" {
 		if held, why := r.restoreHoldFor(name); held {
 			writeJSON(w, http.StatusConflict, apiResponse{OK: false, Error: why})
 			return
@@ -618,6 +622,20 @@ func (r *Router) actionStack(w http.ResponseWriter, action, name string) {
 		}
 	}
 
+	// Slice 4: every cheap refusal of an update — busy, already updating, deploying, memory, disk, and
+	// "no backup to return to" — BEFORE the intent below is recorded, so an update that was never going
+	// to happen records nothing (§8.2). Each is a 409 with the Hungarian sentence.
+	if action == "update" {
+		if ref := r.stackMgr.UpdatePreflight(name); ref != nil {
+			status := http.StatusConflict
+			if ref.Reason == "not_found" {
+				status = http.StatusNotFound
+			}
+			writeJSON(w, status, apiResponse{OK: false, Error: ref.Message})
+			return
+		}
+	}
+
 	// R-166: THE CUSTOMER-INTENT POINT. This switch is where a human's decision about whether their
 	// app should be running enters the system, and until v0.189.0 that decision was recorded nowhere
 	// — so the box had to infer it from container counts, and inferred wrong for a power cut and for
@@ -647,12 +665,18 @@ func (r *Router) actionStack(w http.ResponseWriter, action, name string) {
 	case "restart":
 		err = r.stackMgr.RestartStack(name)
 	case "update":
-		err = r.stackMgr.UpdateStack(name)
+		// Slice 4: the GUARDED update. It returns as soon as the job has started; the result is only
+		// ever known from GET /api/stacks/{name} (updating / update_phase / update_error).
+		err = r.stackMgr.StartGuardedUpdate(name)
 	}
 
 	if err != nil {
 		r.logger.Printf("[ERROR] [api] %s failed for %s: %v", action, name, err)
 		status := http.StatusInternalServerError
+		var ref *stacks.UpdateRefusal
+		if errors.As(err, &ref) {
+			status = http.StatusConflict
+		}
 		if strings.Contains(err.Error(), "protected") {
 			status = http.StatusForbidden
 		}
@@ -663,6 +687,15 @@ func (r *Router) actionStack(w http.ResponseWriter, action, name string) {
 		return
 	}
 
+	// R-443 (slice 4): an update is NEVER reported completed here. The spike measured this line saying
+	// "update completed" over an app that was already crash-looping. It answers 202 — accepted, not
+	// finished — and "completed" exists only as update_phase=done on GET /api/stacks/{name}, which is
+	// written after the app's health is known. Pinned by TestR443_UpdateIsNeverReportedCompleteSynchronously.
+	if action == "update" {
+		writeJSON(w, http.StatusAccepted, apiResponse{OK: true, Message: "Frissítés elindult – az állapot a kártyán követhető",
+			Data: map[string]interface{}{"accepted": true, "completed": false}})
+		return
+	}
 	writeJSON(w, http.StatusOK, apiResponse{OK: true, Message: "Stack " + name + " " + action + " completed"})
 
 	// Trigger integration lifecycle hooks after successful action

controller/internal/api/slice4_update_test.go

@@ -0,0 +1,165 @@
+package api
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-controller/internal/backup"
+	"gitea.dooplex.hu/admin/felhom-controller/internal/config"
+	"gitea.dooplex.hu/admin/felhom-controller/internal/settings"
+	"gitea.dooplex.hu/admin/felhom-controller/internal/stacks"
+)
+
+// Update arc slice 4 through the PRODUCTION handler: actionStack → the real stacks.Manager
+// (NewManager + ScanStacks) and the real backup.Manager over real settings. No docker is reached:
+// every path here refuses, or fails at the pin (the app's catalog template is absent on purpose).
+
+type apiFakeGuards struct {
+	b  *backup.Manager
+	rp stacks.UpdateRestorePoint
+	// blindToHolds makes the manager-side preflight NOT see holds, so a test can prove the ROUTER's
+	// own hold check refuses — the two layers are each pinned separately (the preflight's by
+	// TestSlice4_D_CheapRefusals/held). Without it, removing either layer passes inertly, because the
+	// other refuses with the same sentence (observed on the first run of red-proof 4, 2026-09-13).
+	blindToHolds bool
+}
+
+func (g *apiFakeGuards) HoldFor(n string) (bool, string) {
+	if g.blindToHolds {
+		return false, ""
+	}
+	return g.b.RestoreHoldFor(n)
+}
+func (g *apiFakeGuards) Busy(string) (bool, string) { return false, "" }
+func (g *apiFakeGuards) RestorePoint(string) (stacks.UpdateRestorePoint, error) {
+	return g.rp, nil
+}
+func (g *apiFakeGuards) BackupNow(context.Context, string) error { return nil }
+func (g *apiFakeGuards) SafetyDump(context.Context, string) ([]string, error) {
+	return nil, nil
+}
+func (g *apiFakeGuards) HoldAfterFailedUpdate(string, time.Time, time.Time) error { return nil }
+
+const slice4AppYAML = "deployed: true\nenv: {}\npinned_images:\n    app: nginx:1.27\n"
+
+func newSlice4Router(t *testing.T) (*Router, *settings.Settings, *apiFakeGuards, string) {
+	t.Helper()
+	root := t.TempDir()
+	dir := filepath.Join(root, "stacks", "app")
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dir, "docker-compose.yml"), []byte("services:\n  app:\n    image: nginx:1.27\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dir, "app.yaml"), []byte(slice4AppYAML), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg := &config.Config{}
+	cfg.Paths.StacksDir = filepath.Join(root, "stacks")
+	cfg.Paths.DataDir = filepath.Join(root, "data")
+	cfg.Paths.SystemDataPath = filepath.Join(root, "sys")
+	cfg.Stacks.ComposeCommand = "docker compose"
+	lg := log.New(io.Discard, "", 0)
+	m, err := stacks.NewManager(cfg, lg)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := m.ScanStacks(); err != nil {
+		t.Fatal(err)
+	}
+	sett, err := settings.Load(filepath.Join(root, "settings.json"), lg)
+	if err != nil {
+		t.Fatal(err)
+	}
+	b := backup.NewManager(cfg, sett, lg)
+	g := &apiFakeGuards{b: b, rp: stacks.UpdateRestorePoint{Restorable: true, Proven: true, ProvenAt: time.Now().Add(-time.Hour)}}
+	m.SetUpdateGuards(g)
+	return &Router{cfg: cfg, stackMgr: m, backupMgr: b, logger: lg}, sett, g, dir
+}
+
+func postUpdate(t *testing.T, r *Router) (int, apiResponse) {
+	t.Helper()
+	w := httptest.NewRecorder()
+	r.actionStack(w, "update", "app")
+	var resp apiResponse
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("non-JSON body %q: %v", w.Body.String(), err)
+	}
+	return w.Code, resp
+}
+
+// TestR439_UpdateOfAHeldAppIsRefused — R-439 closed.
+//
+// COMPANION RED-PROOF 4 (REPORT.md): remove `|| action == "update"` from actionStack's hold check. The
+// preflight then refuses on its own grounds with a DIFFERENT sentence, and this test fails on the
+// message — which is what proves the router line is the one doing it.
+func TestR439_UpdateOfAHeldAppIsRefused(t *testing.T) {
+	r, sett, g, dir := newSlice4Router(t)
+	g.rp = stacks.UpdateRestorePoint{} // the preflight's own refusal would say "no backup" — not the hold
+	g.blindToHolds = true              // only the router's line can produce the hold's sentence
+	if err := sett.SetRestoreHold(settings.RestoreHold{Stack: "app", At: "2026-09-13T08:00:00Z", Reason: settings.HoldReasonUpdateFailed, CopyDate: "2026-09-13T01:30:00Z"}); err != nil {
+		t.Fatal(err)
+	}
+	before, _ := os.ReadFile(filepath.Join(dir, "app.yaml"))
+	code, resp := postUpdate(t, r)
+	_, holdText := r.backupMgr.RestoreHoldFor("app")
+	if code != http.StatusConflict || resp.OK || resp.Error != holdText {
+		t.Fatalf("a HELD app's update must be refused with the hold's own sentence: code=%d ok=%v error=%q", code, resp.OK, resp.Error)
+	}
+	after, _ := os.ReadFile(filepath.Join(dir, "app.yaml"))
+	if string(before) != string(after) {
+		t.Error("a refused update must record no intent — app.yaml changed")
+	}
+}
+
+func TestSlice4_Router_NoBackupIs409AndRecordsNothing(t *testing.T) {
+	r, _, g, dir := newSlice4Router(t)
+	g.rp = stacks.UpdateRestorePoint{Restorable: false}
+	before, _ := os.ReadFile(filepath.Join(dir, "app.yaml"))
+	code, resp := postUpdate(t, r)
+	if code != http.StatusConflict || resp.Error != fmt.Sprintf(stacks.MsgUpdateNoBackupFmt, "app") {
+		t.Fatalf("code=%d error=%q", code, resp.Error)
+	}
+	if after, _ := os.ReadFile(filepath.Join(dir, "app.yaml")); string(before) != string(after) {
+		t.Error("the preflight refusal must come BEFORE the intent write")
+	}
+}
+
+// TestR443_UpdateIsNeverReportedCompleteSynchronously — R-443 closed. The handler answers 202 with
+// completed:false; the job then runs (and here fails at the pin, the catalog being absent), and the
+// outcome exists ONLY on GET /api/stacks/{name}.
+func TestR443_UpdateIsNeverReportedCompleteSynchronously(t *testing.T) {
+	r, _, _, _ := newSlice4Router(t)
+	code, resp := postUpdate(t, r)
+	if code != http.StatusAccepted {
+		t.Fatalf("an accepted update must answer 202, got %d (%+v)", code, resp)
+	}
+	data, _ := resp.Data.(map[string]interface{})
+	if data["completed"] != false || data["accepted"] != true {
+		t.Errorf("the body must say accepted and NOT completed, got %v", resp.Data)
+	}
+	if resp.Message == "Stack app update completed" {
+		t.Error("the synchronous response claimed completion — R-443")
+	}
+	deadline := time.Now().Add(5 * time.Second)
+	for time.Now().Before(deadline) {
+		if st, ok := r.stackMgr.GetStack("app"); ok && !st.Updating {
+			if st.UpdatePhase != stacks.UpdatePhaseFailed || st.UpdateError != stacks.MsgUpdatePinFailed {
+				t.Errorf("the job's truth must be on the stack: phase=%q err=%q", st.UpdatePhase, st.UpdateError)
+			}
+			return
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+	t.Fatal("the job never finished")
+}

controller/internal/backup/backup.go

@@ -680,6 +680,15 @@ func (m *Manager) runVolumeDumps() (summary []string, dumped int, allOK bool) {
 		if m.cfg != nil && m.cfg.IsProtectedStack(stack.Name) {
 			continue
 		}
+		// Slice 4 / R-379: a HELD app is deliberately stopped, and DumpAppVolumesSafe ends in
+		// StartStack — so without this line the nightly backup restarted every held app, every night.
+		// "A hold that only one path honours is not a hold." Checked BEFORE the volume check so a held
+		// app is never stopped or started by this leg at all.
+		if m.isHeld(stack.Name) {
+			m.logger.Printf("[WARN] [backup] Skipping volume dump for %s — the app is HELD stopped (its restore point is preserved)", stack.Name)
+			summary = append(summary, fmt.Sprintf("SKIP %s volumes (held)", stack.Name))
+			continue
+		}
 		// Volume check FIRST — a volume-less stack must not be stopped at all (see gate-order note).
 		if len(m.stackProvider.GetDockerVolumes(stack.Name)) == 0 {
 			if m.isDebug() {

controller/internal/backup/offbox_reconstitute.go

@@ -337,6 +337,15 @@ func (m *Manager) RestoreHoldFor(stack string) (bool, string) {
 	if !ok {
 		return false, ""
 	}
+	// Slice 4: one storage, two reasons. An update hold names the copy it can be restored from; a
+	// restore hold names nothing, because the restore it refers to already consumed the copy.
+	if h.Reason == settings.HoldReasonUpdateFailed {
+		copyDate := "legutóbbi"
+		if h.CopyDate != "" {
+			copyDate = fmtHoldTime(h.CopyDate)
+		}
+		return true, fmt.Sprintf(UpdateHoldFmt, stack, fmtHoldTime(h.At), copyDate)
+	}
 	when := h.At
 	if t, err := time.Parse(time.RFC3339, h.At); err == nil {
 		when = t.Format("2006-01-02 15:04")

controller/internal/backup/recovery_unit.go

@@ -385,6 +385,14 @@ func (m *Manager) captureAllRecoveryUnits() {
 		if m.settings != nil && (m.settings.IsDisconnected(drivePath) || m.settings.IsDecommissioned(drivePath)) {
 			continue // drive not writable — skip, the existing unit stays as-is
 		}
+		// Slice 4: a HELD app's unit is its RESTORE POINT, and the hold text names that copy's date.
+		// Re-capturing it would write the definition the app is held ON (after a failed update: the new
+		// version that would not start) into the unit, and the next Tier-2 run would mirror it over the
+		// copy the customer was told to restore from. The unit stays exactly as it was until the hold is
+		// lifted.
+		if m.isHeld(stack.Name) {
+			continue
+		}
 		m.noteAttempted(stack.Name)
 		// The reserve, checked BEFORE anything is written. Per app, and the loop continues.
 		if !m.admitApp(stack.Name) {

controller/internal/backup/restore_unit.go

@@ -390,5 +390,8 @@ func (m *Manager) RestoreFromRecoveryUnitAt(stackName, unitDir string) (UnitRest
 	}
 	m.logger.Printf("[INFO] [backup] Restore-from-unit completed: %s — %d volume(s) of %d listed, %d database(s) of %d listed",
 		stackName, res.VolumesReplayed, res.ManifestVolumes, res.DBsReplayed, res.ManifestDBs)
+	// Slice 4: the app is back on its unit's definition and data and was started — the route back an
+	// update hold names. Lift that hold (and only that kind; see clearUpdateHoldAfterRestore).
+	m.clearUpdateHoldAfterRestore(stackName)
 	return res, nil
 }

controller/internal/backup/slice4_update_guard_test.go

@@ -0,0 +1,168 @@
+package backup
+
+import (
+	"fmt"
+	"io"
+	"log"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-controller/internal/settings"
+)
+
+// Update arc slice 4 — the backup side of the guarded update.
+
+// The measured case (demo-hp 2026-09-13, bookstack): the mirror's manifest said 2026-09-12T02:15:29Z
+// while its database dump was written 2026-09-13T00:30Z and the Tier-2 run succeeded at 01:30Z. The
+// update's age must be the proven COPY time; the manifest date would call a fresh copy stale forever.
+func TestSlice4_ProvenCopyTime_IsTheLastSuccessNotTheManifestDate(t *testing.T) {
+	rp := restorePointFromCoverage(Tier2Coverage{
+		UnitRestorable: true, UnitPackageDate: "2026-09-12T02:15:29Z",
+		CopyLastRun: "2026-09-13T01:30:00Z", CopyLastSuccess: "2026-09-13T01:30:00Z",
+	})
+	at, ok := rp.ProvenCopyTime()
+	if !ok || !at.Equal(time.Date(2026, 9, 13, 1, 30, 0, 0, time.UTC)) {
+		t.Fatalf("proven copy time = %v ok=%v, want the last successful copy", at, ok)
+	}
+	// The page still names the PACKAGE date (R-403) — the extraction changes nothing it shows.
+	if rp.CopyDate != "2026-09-12T02:15:29Z" || !rp.CopyDateProven || rp.PackagePreserved {
+		t.Errorf("restore point = %+v", rp)
+	}
+}
+
+func TestSlice4_ProvenCopyTime_PreservedPackageUsesThePackageDate(t *testing.T) {
+	rp := restorePointFromCoverage(Tier2Coverage{
+		UnitRestorable: true, UnitPackageDate: "2026-09-01T02:00:00Z", UnitLegPreserved: true,
+		CopyLastSuccess: "2026-09-13T01:30:00Z",
+	})
+	if at, ok := rp.ProvenCopyTime(); !ok || !at.Equal(time.Date(2026, 9, 1, 2, 0, 0, 0, time.UTC)) {
+		t.Errorf("a PRESERVED package is as old as the package, got %v ok=%v", at, ok)
+	}
+}
+
+func TestSlice4_ProvenCopyTime_NoProvenOrNoUnitIsNoRestorePoint(t *testing.T) {
+	for _, cov := range []Tier2Coverage{
+		{UnitRestorable: true, CopyLastRun: "2026-09-13T01:30:00Z"},      // attempt, never a success (R-101)
+		{UnitRestorable: false, CopyLastSuccess: "2026-09-13T01:30:00Z"}, // a copy with no openable unit
+		{UnitRestorable: true, CopyLastSuccess: "not-a-date"},            // unparseable is unknown, never "now"
+	} {
+		if _, ok := restorePointFromCoverage(cov).ProvenCopyTime(); ok {
+			t.Errorf("%+v must not yield a proven copy time", cov)
+		}
+	}
+}
+
+func slice4Settings(t *testing.T) *settings.Settings {
+	t.Helper()
+	s, err := settings.Load(filepath.Join(t.TempDir(), "settings.json"), log.New(io.Discard, "", 0))
+	if err != nil {
+		t.Fatal(err)
+	}
+	return s
+}
+
+func TestSlice4_UpdateHoldTextNamesTheTimeAndTheCopy(t *testing.T) {
+	sett := slice4Settings(t)
+	m := &Manager{logger: log.New(io.Discard, "", 0), settings: sett}
+	at := time.Date(2026, 9, 13, 8, 0, 0, 0, time.UTC)
+	copyAt := time.Date(2026, 9, 13, 1, 30, 0, 0, time.UTC)
+	if err := m.HoldAfterFailedUpdate("bookstack", at, copyAt); err != nil {
+		t.Fatal(err)
+	}
+	h, ok := sett.GetRestoreHold("bookstack")
+	if !ok || h.Reason != settings.HoldReasonUpdateFailed || h.CopyDate != "2026-09-13T01:30:00Z" {
+		t.Fatalf("hold = %+v ok=%v", h, ok)
+	}
+	held, why := m.RestoreHoldFor("bookstack")
+	// Budapest is UTC+2 in September: 08:00Z → 10:00, 01:30Z → 03:30.
+	want := fmt.Sprintf(UpdateHoldFmt, "bookstack", "2026-09-13 10:00", "2026-09-13 03:30")
+	if !held || why != want {
+		t.Errorf("hold text =\n%q\nwant\n%q", why, want)
+	}
+}
+
+func TestSlice4_RestoreHoldTextIsUnchanged(t *testing.T) {
+	sett := slice4Settings(t)
+	m := &Manager{logger: log.New(io.Discard, "", 0), settings: sett}
+	if err := sett.SetRestoreHold(settings.RestoreHold{Stack: "docmost", At: "2026-08-22T14:00:00Z"}); err != nil {
+		t.Fatal(err)
+	}
+	_, why := m.RestoreHoldFor("docmost")
+	if !strings.Contains(why, "visszaállítása") || !strings.Contains(why, "Vedd fel velünk a kapcsolatot") || strings.Contains(why, "frissítése") {
+		t.Errorf("an R-379 restore hold must keep its own sentence, got %q", why)
+	}
+}
+
+func TestSlice4_ASuccessfulRestoreClearsOnlyAnUpdateHold(t *testing.T) {
+	sett := slice4Settings(t)
+	m := &Manager{logger: log.New(io.Discard, "", 0), settings: sett}
+	_ = m.HoldAfterFailedUpdate("upd", time.Now(), time.Now())
+	_ = sett.SetRestoreHold(settings.RestoreHold{Stack: "rst", At: "2026-08-22T14:00:00Z"})
+	m.clearUpdateHoldAfterRestore("upd")
+	m.clearUpdateHoldAfterRestore("rst")
+	if _, ok := sett.GetRestoreHold("upd"); ok {
+		t.Error("a restore is the route back from a failed update — its hold must be lifted")
+	}
+	if _, ok := sett.GetRestoreHold("rst"); !ok {
+		t.Error("an R-379 restore hold stays operator-cleared")
+	}
+}
+
+func TestSlice4_UpdateBusy(t *testing.T) {
+	m := &Manager{logger: log.New(io.Discard, "", 0)}
+	if busy, _ := m.UpdateBusy("app"); busy {
+		t.Fatal("an idle manager is not busy")
+	}
+	if err := m.acquireRunning(); err != nil {
+		t.Fatal(err)
+	}
+	if busy, _ := m.UpdateBusy("app"); !busy {
+		t.Error("a running backup/restore must make an update wait")
+	}
+	m.releaseRunning()
+	m.BeginRestoreOp("tier2-unit-restore", "other")
+	if busy, _ := m.UpdateBusy("app"); !busy {
+		t.Error("a restore op in flight must make an update wait")
+	}
+}
+
+// "A hold that only one path honours is not a hold." The nightly legs are unattended start paths
+// (DumpAppVolumesSafe ends in StartStack) and writers of the restore point the hold text names.
+//
+// COMPANION RED-PROOF (REPORT.md): delete the isHeld skip from runVolumeDumps — the held app is then
+// stopped (and restarted) by the nightly backup, and this test fails.
+func TestSlice4_NightlyLegsLeaveAHeldAppAlone(t *testing.T) {
+	h := newAdmissionHarness(t, "held", "free")
+	h.m.settings = slice4Settings(t)
+	if err := h.m.HoldAfterFailedUpdate("held", time.Now(), time.Now()); err != nil {
+		t.Fatal(err)
+	}
+	h.m.runVolumeDumps()
+	for _, n := range append(append([]string{}, h.volDumped...), h.prov.stopped...) {
+		if n == "held" {
+			t.Fatalf("the nightly volume dump touched a HELD app (dumped=%v stopped=%v)", h.volDumped, h.prov.stopped)
+		}
+	}
+	if len(h.volDumped) != 1 || h.volDumped[0] != "free" {
+		t.Errorf("positive control: the unheld app must still be dumped, got %v", h.volDumped)
+	}
+	h.m.captureAllRecoveryUnits()
+	for _, n := range h.prov.infoHits {
+		if n == "held" {
+			t.Error("the capture must not rewrite a HELD app's restore point")
+		}
+	}
+	var mirrored []string
+	h.m.perAppTier2 = func(name string) error { mirrored = append(mirrored, name); return nil }
+	h.m.RunAllTier2()
+	for _, n := range mirrored {
+		if n == "held" {
+			t.Error("Tier 2 must not mirror over a HELD app's copy")
+		}
+	}
+	if len(mirrored) != 1 {
+		t.Errorf("positive control: the unheld app must still be mirrored, got %v", mirrored)
+	}
+}

controller/internal/backup/tier2.go

@@ -457,6 +457,12 @@ func (m *Manager) RunAllTier2() {
 			m.settings.IsDecommissioned(m.GetAppDrivePath(stack.Name))) {
 			continue
 		}
+		// Slice 4: never mirror over a HELD app's copy — it is the restore point the hold text names.
+		// See the matching skip in captureAllRecoveryUnits.
+		if m.isHeld(stack.Name) {
+			m.logger.Printf("[WARN] [backup] Tier 2 skipped for %s — the app is HELD; its copy is the restore point and is preserved", stack.Name)
+			continue
+		}
 		runOne := m.perAppTier2
 		if runOne == nil {
 			runOne = m.RunTier2

controller/internal/backup/update_guard.go

@@ -0,0 +1,325 @@
+package backup
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"path/filepath"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-controller/internal/settings"
+)
+
+// ── The backup side of the guarded update (update arc slice 4, controller v0.237.0) ─────────────
+//
+// 09-update-architecture.md §3 decision 1 (operator ruling 2026-09-02): the safety copy for an update
+// is a VERIFIED RECENT BACKUP as a PRECONDITION — not a new copy mechanism invented for the update
+// path. So everything in this file composes machinery that already exists and is proven live:
+// the Tier-2 unit restore's own predicate (R-102/R-103), the nightly legs (DB dump, volume dump,
+// unit capture, Tier-2 mirror), the pre-restore safety dump (R-361) and the R-379 hold.
+//
+// The stacks package cannot import this one, so the update job reaches all of it through the
+// stacks.UpdateGuards interface, implemented by an adapter in cmd/controller/main.go.
+
+// Tier2RestorePoint is the answer to "could this app be restored from its Tier-2 copy, and from
+// when?" — the predicate the destructive „Teljes visszaállítás" action is gated on.
+//
+// EXTRACTED, NOT DUPLICATED (slice 4). Until v0.237.0 this computation lived inline in the backups
+// page handler (buildAppBackupRows). The update path needs exactly the same question answered, and
+// a second copy of a predicate is how this project's two copies of `namespaceRoot` came to differ
+// (R-203). So there is one function, and the page and the update both call it.
+type Tier2RestorePoint struct {
+	// Restorable — the copy holds an OPENABLE recovery unit (Tier2Coverage.CanRestoreUnit).
+	Restorable bool
+	// CopyDate — the date the unit restore NAMES: the package's own manifest date, falling back to the
+	// copy date (Tier2Coverage.UnitRestoreDate, R-403). RFC3339 as recorded, "" when unknown.
+	CopyDate string
+	// CopyDateProven — a copy actually SUCCEEDED (LastSuccess is set), never merely an attempt (R-101).
+	CopyDateProven bool
+	// PackagePreserved — the newest run PRESERVED an older package instead of refreshing it (R-403).
+	PackagePreserved bool
+	// CopyLastSuccess — the RFC3339 time of the last Tier-2 copy that succeeded.
+	CopyLastSuccess string
+}
+
+// restorePointFromCoverage is the pure half of the predicate.
+func restorePointFromCoverage(cov Tier2Coverage) Tier2RestorePoint {
+	pkgDate, preserved := cov.UnitRestoreDate()
+	return Tier2RestorePoint{
+		Restorable:       cov.CanRestoreUnit(),
+		CopyDate:         pkgDate,
+		CopyDateProven:   cov.CopyLastSuccess != "",
+		PackagePreserved: preserved,
+		CopyLastSuccess:  cov.CopyLastSuccess,
+	}
+}
+
+// Tier2UnitRestorePoint resolves the app's recorded Tier-2 copy and returns the restore point. The
+// error is the same refusal Tier2RestoreCoverage raises (no copy, drive gone, pre-v2 layout).
+func (m *Manager) Tier2UnitRestorePoint(stackName string) (Tier2RestorePoint, error) {
+	cov, err := m.Tier2RestoreCoverage(stackName)
+	if err != nil {
+		return Tier2RestorePoint{}, err
+	}
+	return restorePointFromCoverage(cov), nil
+}
+
+// ProvenCopyTime returns WHEN the data this copy would restore was last proven copied, and false when
+// there is no proven, restorable copy at all.
+//
+// WHY NOT CopyDate, measured rather than assumed. CopyDate is the unit MANIFEST's created_at, and a
+// capture rewrites the manifest only when the app's DEFINITION changes (compose, app.yaml, controller
+// version) — a nightly DB dump keeps the same file name, so it does not move it. Measured on demo-hp
+// 2026-09-13: bookstack's Tier-2 mirror held `bookstack-mariadb.sql` written 2026-09-13T00:30Z while
+// its manifest still read 2026-09-12T02:15:29Z. Judging "recent" by that date would call a fresh copy
+// stale — and, worse, a "back up first" run would not move it either on a quiet app, so the update
+// would be refused forever.
+//
+// So the age is the last SUCCESSFUL copy (LastSuccess), which the Tier-2 run records only when it
+// actually mirrored the unit — EXCEPT when the run preserved an older package (R-403), in which case
+// the package date is the honest one, because that is what the copy really holds.
+func (p Tier2RestorePoint) ProvenCopyTime() (time.Time, bool) {
+	if !p.Restorable || !p.CopyDateProven {
+		return time.Time{}, false
+	}
+	src := p.CopyLastSuccess
+	if p.PackagePreserved {
+		src = p.CopyDate
+	}
+	t, err := time.Parse(time.RFC3339, src)
+	if err != nil {
+		return time.Time{}, false
+	}
+	return t, true
+}
+
+// UpdateBusy reports whether something else is ALREADY touching this app's data, which refuses an
+// update before anything moves (slice 4 Scenario D). The reason is operator-English; the customer
+// sentence is chosen by the caller.
+//
+// IsRunning is box-wide, deliberately: the backup/restore single-flight is box-wide, and an update's
+// "back up first" leg needs that same flag — an update started beside a running backup would either
+// wait on it invisibly or fail half-way.
+func (m *Manager) UpdateBusy(stackName string) (bool, string) {
+	if m == nil {
+		return false, ""
+	}
+	if m.IsRunning() {
+		return true, "a backup or restore is running (single-flight held)"
+	}
+	if st := m.RestoreStatus(); st.Running {
+		return true, fmt.Sprintf("restore op %q is running for %q", st.Op, st.Stack)
+	}
+	for _, held := range m.appStop.HeldStacks() {
+		if held == stackName {
+			return true, "an app-data operation (volume dump / export / reconstitute) is holding it"
+		}
+	}
+	return false, ""
+}
+
+// ErrUpdateBackupNoUnit is returned when a "back up first" run completed but the app still has no
+// openable Tier-2 unit — typically Tier 2 is switched off for the app or has no second target.
+var ErrUpdateBackupNoUnit = errors.New("a frissítés előtti mentés lefutott, de nem jött létre visszaállítható másolat")
+
+// RunAppBackupNow runs THIS app's backup legs now, in the nightly order, and then its Tier-2 copy:
+// database dump(s) → volume dump (if the app has named volumes) → recovery-unit capture → Tier-2
+// mirror. It is the "back up first" of slice 4 Scenario B.
+//
+// Composed, not reinvented: every leg is the one runDBDumpsInternal and RunAllTier2 already run,
+// including the R-181 reserve (admitApp) before the first write and the R-166 app-stop marker inside
+// DumpAppVolumesSafe. What differs is only the scope — one app instead of all of them — because an
+// update must not bounce every other app on the box to back up one.
+func (m *Manager) RunAppBackupNow(ctx context.Context, stackName string) error {
+	if m.stackProvider == nil {
+		return fmt.Errorf("stack provider not configured")
+	}
+	if m.migrationActive() {
+		return fmt.Errorf("adatáthelyezés folyamatban — a mentés most nem indítható")
+	}
+	if err := m.acquireRunning(); err != nil {
+		return err
+	}
+	m.logger.Printf("[INFO] [backup] update pre-backup for %s: starting (DB dump → volume dump → unit capture → Tier 2)", stackName)
+	start := time.Now()
+	legErr := func() error {
+		defer m.releaseRunning()
+		defer m.beginAdmissionRun()()
+
+		drivePath := m.GetAppDrivePath(stackName)
+		if drivePath == "" || !filepath.IsAbs(drivePath) {
+			return fmt.Errorf("az alkalmazás meghajtója nem határozható meg")
+		}
+		if m.settings != nil && (m.settings.IsDisconnected(drivePath) || m.settings.IsDecommissioned(drivePath)) {
+			return fmt.Errorf("az alkalmazás meghajtója nem elérhető (%s)", drivePath)
+		}
+		if !m.admitApp(stackName) {
+			return fmt.Errorf("nincs elég szabad hely a mentéshez a(z) %s meghajtón", drivePath)
+		}
+		nsRoot := m.namespaceRoot(drivePath)
+
+		discover := m.discoverDBs
+		if discover == nil {
+			discover = func(ctx context.Context) ([]DiscoveredDB, error) {
+				return DiscoverDatabases(ctx, m.logger, m.isDebug(), m.knownStackNames())
+			}
+		}
+		dbs, err := discover(ctx)
+		if err != nil {
+			return fmt.Errorf("adatbázis-felderítés sikertelen: %w", err)
+		}
+		dumped := 0
+		for _, db := range dbs {
+			if db.StackName != stackName {
+				continue
+			}
+			res := DumpOne(ctx, db, AppDBDumpPath(nsRoot, stackName), m.logger, m.isDebug())
+			if res.Error != nil {
+				return fmt.Errorf("adatbázis-mentés sikertelen (%s): %w", db.ContainerName, res.Error)
+			}
+			dumped++
+			m.logger.Printf("[INFO] [backup] update pre-backup for %s: database dump OK (%s, %s)", stackName, db.ContainerName, humanizeBytes(res.Size))
+		}
+
+		if len(m.stackProvider.GetDockerVolumes(stackName)) > 0 {
+			dump := m.dumpVolumesSafe
+			if dump == nil {
+				dump = m.DumpAppVolumesSafe
+			}
+			if err := dump(stackName); err != nil {
+				return fmt.Errorf("kötetmentés sikertelen: %w", err)
+			}
+			m.logger.Printf("[INFO] [backup] update pre-backup for %s: volume dump OK", stackName)
+		}
+
+		if err := m.CaptureRecoveryUnit(stackName); err != nil {
+			return fmt.Errorf("a mentési egység rögzítése sikertelen: %w", err)
+		}
+		m.logger.Printf("[INFO] [backup] update pre-backup for %s: recovery unit captured (%d database dump(s))", stackName, dumped)
+		return nil
+	}()
+	if legErr != nil {
+		m.logger.Printf("[ERROR] [backup] update pre-backup for %s FAILED after %s: %v", stackName, time.Since(start).Round(time.Millisecond), legErr)
+		return legErr
+	}
+
+	runOne := m.perAppTier2
+	if runOne == nil {
+		runOne = m.RunTier2
+	}
+	if err := runOne(stackName); err != nil {
+		m.logger.Printf("[ERROR] [backup] update pre-backup for %s: Tier 2 copy FAILED: %v", stackName, err)
+		return fmt.Errorf("a másodlagos másolat elkészítése sikertelen: %w", err)
+	}
+	m.logger.Printf("[INFO] [backup] update pre-backup for %s: complete in %s", stackName, time.Since(start).Round(time.Millisecond))
+	return nil
+}
+
+// WriteUpdateSafetyDump takes the last-minute database copy an update makes just before it moves the
+// pin: "the state the customer was in a minute ago". It is writeSafetyDump (R-361) unchanged — the
+// same `pre-restore-` undo naming, the same pruning to three, the same never-the-canonical-name rule
+// — so it is also picked up by the same exclusions (it never enters a manifest's db_dumps).
+//
+// Returns the paths written; an app with no database returns (nil, nil), which is a no-op and never a
+// failure (measured in writeSafetyDump: `len(mine) == 0` returns an empty set).
+func (m *Manager) WriteUpdateSafetyDump(ctx context.Context, stackName string) ([]string, error) {
+	nsRoot := m.AppNamespaceRoot(stackName)
+	if nsRoot == "" {
+		return nil, fmt.Errorf("az alkalmazás mentési helye nem határozható meg")
+	}
+	set, err := m.writeSafetyDump(ctx, stackName, nsRoot)
+	if err != nil {
+		return nil, err
+	}
+	var paths []string
+	for _, f := range set.Files {
+		paths = append(paths, f.Path)
+	}
+	if len(paths) == 0 {
+		m.logger.Printf("[INFO] [backup] update safety dump for %s: the app has no database — nothing to copy (no-op)", stackName)
+	} else {
+		m.logger.Printf("[INFO] [backup] update safety dump for %s: %d file(s) %v", stackName, len(paths), paths)
+	}
+	return paths, nil
+}
+
+// UpdateHoldFmt is the customer sentence for an app held after a failed update. Arguments: the app,
+// the time of the failure, and the PROVEN date of the copy it can be restored from. One named string so
+// a test asserts it verbatim instead of retyping Hungarian (R-364).
+const UpdateHoldFmt = "A(z) %s frissítése %s-kor nem sikerült, és az alkalmazás nem indult el az új verzióval. " +
+	"Az alkalmazás biztonsági okból leállítva marad, hogy az adatai ne sérüljenek. " +
+	"Visszaállítható a(z) %s-i biztonsági mentésből a Mentések oldalon."
+
+// holdTimeZone is where the customer-facing hold sentence renders its times. The same zone the web
+// layer renders the Mentések page's copy dates in (web.getTimezone), so the date in the hold text and
+// the date on the page it points at are the same string.
+func holdTimeZone() *time.Location {
+	if loc, err := time.LoadLocation("Europe/Budapest"); err == nil {
+		return loc
+	}
+	return time.UTC
+}
+
+func fmtHoldTime(rfc3339 string) string {
+	t, err := time.Parse(time.RFC3339, rfc3339)
+	if err != nil {
+		return rfc3339
+	}
+	return t.In(holdTimeZone()).Format("2006-01-02 15:04")
+}
+
+// HoldAfterFailedUpdate records that an app is held stopped because its new version did not come up
+// healthy. Same storage and same gate as the R-379 hold — every start path that already refuses a
+// restore hold refuses this one without being touched.
+//
+// It returns the error rather than only logging it, unlike holdAppAfterFailedRollback: the update job
+// has just stopped the app on the strength of this record, and an unrecorded hold is a stopped app
+// that the next restart button will quietly start again. The caller logs it at ERROR and keeps the
+// failure on the page.
+func (m *Manager) HoldAfterFailedUpdate(stackName string, at time.Time, copyDate time.Time) error {
+	if m == nil || m.settings == nil {
+		return fmt.Errorf("no settings wired — the update hold for %s cannot be persisted", stackName)
+	}
+	h := settings.RestoreHold{
+		Stack:  stackName,
+		At:     at.UTC().Format(time.RFC3339),
+		Reason: settings.HoldReasonUpdateFailed,
+	}
+	if !copyDate.IsZero() {
+		h.CopyDate = copyDate.UTC().Format(time.RFC3339)
+	}
+	if err := m.settings.SetRestoreHold(h); err != nil {
+		return fmt.Errorf("persisting the update hold for %s: %w", stackName, err)
+	}
+	m.logger.Printf("[WARN] [backup] %s is HELD STOPPED after a failed update (restore point: %s)", stackName, h.CopyDate)
+	return nil
+}
+
+// isHeld reports whether an app carries ANY hold. Used by the nightly legs to leave a held app alone.
+func (m *Manager) isHeld(stackName string) bool {
+	held, _ := m.RestoreHoldFor(stackName)
+	return held
+}
+
+// clearUpdateHoldAfterRestore lifts an UPDATE hold once a person has restored the app successfully.
+//
+// "A person clears it by restoring" — slice 4 Part 3. The restore just put the app back on the
+// definition and data of its recovery unit and started it, which is the exact route back the hold
+// text names; leaving the hold in place would refuse the next restart of an app that is now fine.
+//
+// A RESTORE hold (R-379) is deliberately NOT cleared here: that hold means a previous restore already
+// left the database in an unknown state, and it stays operator-cleared (`-clear-restore-hold`).
+func (m *Manager) clearUpdateHoldAfterRestore(stackName string) {
+	if m.settings == nil {
+		return
+	}
+	h, ok := m.settings.GetRestoreHold(stackName)
+	if !ok || h.Reason != settings.HoldReasonUpdateFailed {
+		return
+	}
+	if _, err := m.settings.ClearRestoreHold(stackName); err != nil {
+		m.logger.Printf("[ERROR] [backup] %s was restored, but its update hold could not be cleared: %v", stackName, err)
+		return
+	}
+	m.logger.Printf("[INFO] [backup] %s: restore completed — the update hold (set %s) is CLEARED", stackName, h.At)
+}

controller/internal/config/config.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"os"
 	"strings"
+	"time"
 
 	"gopkg.in/yaml.v3"
 )
@@ -46,6 +47,49 @@ type Config struct {
 	Quiesce        QuiesceConfig        `yaml:"quiesce"`
 	MailRelay      MailRelayConfig      `yaml:"mail_relay"`
 	Offsite        OffsiteConfig        `yaml:"offsite"`
+	Update         UpdateConfig         `yaml:"update"`
+}
+
+// UpdateConfig tunes the guarded app update (update arc slice 4, v0.237.0).
+//
+// Both are OPERATOR knobs, not constants: "recent" and "healthy in time" are judgements the operator
+// ruled defaults for (2026-09-13) and may want to move per box without a release.
+type UpdateConfig struct {
+	// BackupMaxAge is how old the app's proven Tier-2 copy may be before an update makes a fresh one
+	// first. Default "24h".
+	BackupMaxAge string `yaml:"backup_max_age"`
+	// HealthTimeout bounds the wait for the new version to become healthy before the app is held.
+	// Default "5m".
+	HealthTimeout string `yaml:"health_timeout"`
+}
+
+// DefaultUpdateBackupMaxAge and DefaultUpdateHealthTimeout are the values used when the config is
+// empty or unparseable. An unparseable value falls back rather than failing the whole config load:
+// a typo in a tuning knob must not stop a controller from starting.
+const (
+	DefaultUpdateBackupMaxAge  = 24 * time.Hour
+	DefaultUpdateHealthTimeout = 5 * time.Minute
+)
+
+// BackupMaxAgeDuration parses BackupMaxAge, falling back to the default on empty/invalid/non-positive.
+func (u UpdateConfig) BackupMaxAgeDuration() time.Duration {
+	return parsePositiveDuration(u.BackupMaxAge, DefaultUpdateBackupMaxAge)
+}
+
+// HealthTimeoutDuration parses HealthTimeout, falling back to the default on empty/invalid/non-positive.
+func (u UpdateConfig) HealthTimeoutDuration() time.Duration {
+	return parsePositiveDuration(u.HealthTimeout, DefaultUpdateHealthTimeout)
+}
+
+func parsePositiveDuration(s string, def time.Duration) time.Duration {
+	if s == "" {
+		return def
+	}
+	d, err := time.ParseDuration(s)
+	if err != nil || d <= 0 {
+		return def
+	}
+	return d
 }
 
 // MailRelayConfig tunes the in-controller SMTP shim (app email → shim → hub → Resend).
@@ -384,6 +428,8 @@ func applyDefaults(cfg *Config) {
 	d(&cfg.MailRelay.TLSListen, ":2465")
 	d(&cfg.MailRelay.PlainNoTLSListen, ":2526")
 	d(&cfg.MailRelay.ShimHost, "felhom-controller")
+	d(&cfg.Update.BackupMaxAge, "24h")
+	d(&cfg.Update.HealthTimeout, "5m")
 	if len(cfg.MailRelay.FromDomains) == 0 {
 		cfg.MailRelay.FromDomains = []string{"felhom.eu"}
 	}

controller/internal/settings/settings.go

@@ -1585,8 +1585,27 @@ type RestoreHold struct {
 	ReplayError string `json:"replay_error,omitempty"`   // what the restore hit
 	RollbackErr string `json:"rollback_error,omitempty"` // what the rollback then hit
 	SafetyDump  string `json:"safety_dump,omitempty"`    // basename of the undo copy that could not be applied
+
+	// Reason (update arc slice 4, v0.237.0) says WHICH operation put the hold in place. Empty means
+	// HoldReasonRestoreFailed — every hold written before this field existed was a restore hold, so
+	// the zero value keeps their meaning and no migration is needed.
+	//
+	// ONE STORAGE, ONE GATE, TWO REASONS — deliberately not a second map. Every start path already
+	// consults GetRestoreHold (the customer's button, the boot sweep, the app-stop guard's Recover),
+	// and "a hold that only one path honours is not a hold". A second map would need every one of
+	// those paths found and changed again, and the one that got missed would be the next R-439.
+	Reason string `json:"reason,omitempty"`
+	// CopyDate is the RFC3339 time of the proven backup the customer is told they can restore from.
+	// Only set for HoldReasonUpdateFailed.
+	CopyDate string `json:"copy_date,omitempty"`
 }
 
+// Hold reasons. See RestoreHold.Reason.
+const (
+	HoldReasonRestoreFailed = ""              // R-379/R-380: a restore AND its rollback failed
+	HoldReasonUpdateFailed  = "update_failed" // slice 4: the new version did not come up healthy
+)
+
 // SetRestoreHold records a hold. Modelled on SetDisconnected: a condition, plus what it is holding.
 func (s *Settings) SetRestoreHold(h RestoreHold) error {
 	s.mu.Lock()

controller/internal/stacks/deploy.go

@@ -4,6 +4,7 @@ import (
 	"crypto/rand"
 	"encoding/base64"
 	"encoding/hex"
+	"errors"
 	"fmt"
 	"log"
 	"math/big"
@@ -236,52 +237,12 @@ func (m *Manager) DeployStack(req DeployRequest) (string, error) {
 	}
 
 	// --- Memory validation ---
-	var deployWarning string
-	reservedMB := m.cfg.System.ReservedMemoryMB
-	totalMB, usedMB, memErr := system.GetMemoryMB()
-	// F1: the controller container cannot read the guest's RAM cap from /proc (no lxcfs) or its own
-	// cgroup (the cap is on the LXC ancestor). Prefer the guest cap from the Docker daemon (runs in the
-	// LXC). And use the controller's OWN committed-memory accounting for "used" — accurate and cheap —
-	// rather than host /proc RSS, which is unobservable-per-guest and would otherwise make this guard
-	// either never fire (host total) or always fire (host used > guest cap).
-	if gt, ok := system.GuestMemTotalMB(); ok && gt > 0 {
-		totalMB = gt
-		memErr = nil
-	}
-	if committedReqMB, _ := m.CommittedMemory(); committedReqMB > 0 || memErr == nil {
-		usedMB = committedReqMB
-	}
-	if memErr != nil {
-		m.logger.Printf("[WARN] [stacks] Cannot read system memory: %v — skipping memory check", memErr)
-	} else {
-		usableMB := totalMB - reservedMB
-		newReqMB := ParseMemoryMB(meta.Resources.MemRequest)
-
-		m.logger.Printf("[INFO] [stacks] Memory check: total=%dMB, reserved=%dMB, usable=%dMB, committed_used=%dMB, new_req=%dMB, remaining=%dMB",
-			totalMB, reservedMB, usableMB, usedMB, newReqMB, usableMB-usedMB-newReqMB)
-
-		// Hard block: committed + new request exceeds usable memory
-		if newReqMB > 0 && usedMB+newReqMB > usableMB {
-			clearDeploying()
-			return "", fmt.Errorf(
-				"Nincs elég memória az alkalmazás telepítéséhez. "+
-					"Szükséges: %d MB, Elérhető: %d MB "+
-					"(összesen: %d MB, ebből %d MB használt, %d MB rendszer számára fenntartva)",
-				newReqMB,
-				usableMB-usedMB,
-				totalMB,
-				usedMB,
-				reservedMB,
-			)
-		}
-
-		// Soft warning: limits exceed total (overcommit)
-		_, currentLimitMB := m.CommittedMemory()
-		newLimitMB := ParseMemoryMB(meta.Resources.MemLimit)
-		if newLimitMB > 0 && currentLimitMB+newLimitMB > totalMB {
-			deployWarning = "Az alkalmazások csúcsterhelése meghaladhatja a rendelkezésre álló memóriát. " +
-				"Normál használat mellett ez nem okoz problémát."
-		}
+	// Slice 4: the block moved into memoryVerdict so the guarded update applies the SAME check with
+	// the SAME wording. Behaviour here is unchanged — same inputs, same log line, same refusal text.
+	refusal, deployWarning := m.memoryVerdict(ParseMemoryMB(meta.Resources.MemRequest), ParseMemoryMB(meta.Resources.MemLimit), 0, 0)
+	if refusal != "" {
+		clearDeploying()
+		return "", errors.New(refusal)
 	}
 
 	// Debug: log received values (redact passwords/secrets)
@@ -1197,3 +1158,62 @@ func randomAlphanumeric(length int) (string, error) {
 	}
 	return string(result), nil
 }
+
+// memoryVerdict is the deploy's memory check, extracted so the guarded update uses it unchanged
+// (slice 4). releasedReqMB/releasedLimitMB are what the act FREES before it takes the new amount — an
+// update replaces the app's own current request, so counting both would refuse an update that fits.
+// A deploy releases nothing and passes 0, 0.
+//
+// Returns the refusal (the deploy's own Hungarian wording, "" = admitted) and the soft overcommit
+// warning. An unreadable memory reading admits with a WARN, exactly as the deploy always has.
+func (m *Manager) memoryVerdict(newReqMB, newLimitMB, releasedReqMB, releasedLimitMB int) (refusal, warning string) {
+	reservedMB := m.cfg.System.ReservedMemoryMB
+	totalMB, usedMB, memErr := system.GetMemoryMB()
+	// F1: the controller container cannot read the guest's RAM cap from /proc (no lxcfs) or its own
+	// cgroup (the cap is on the LXC ancestor). Prefer the guest cap from the Docker daemon (runs in the
+	// LXC). And use the controller's OWN committed-memory accounting for "used" — accurate and cheap —
+	// rather than host /proc RSS, which is unobservable-per-guest and would otherwise make this guard
+	// either never fire (host total) or always fire (host used > guest cap).
+	if gt, ok := system.GuestMemTotalMB(); ok && gt > 0 {
+		totalMB = gt
+		memErr = nil
+	}
+	if committedReqMB, _ := m.CommittedMemory(); committedReqMB > 0 || memErr == nil {
+		usedMB = committedReqMB
+	}
+	if memErr != nil {
+		m.logger.Printf("[WARN] [stacks] Cannot read system memory: %v — skipping memory check", memErr)
+		return "", ""
+	}
+	usedMB -= releasedReqMB
+	if usedMB < 0 {
+		usedMB = 0
+	}
+	usableMB := totalMB - reservedMB
+
+	m.logger.Printf("[INFO] [stacks] Memory check: total=%dMB, reserved=%dMB, usable=%dMB, committed_used=%dMB, new_req=%dMB, remaining=%dMB",
+		totalMB, reservedMB, usableMB, usedMB, newReqMB, usableMB-usedMB-newReqMB)
+
+	// Hard block: committed + new request exceeds usable memory
+	if newReqMB > 0 && usedMB+newReqMB > usableMB {
+		return fmt.Sprintf(
+			"Nincs elég memória az alkalmazás telepítéséhez. "+
+				"Szükséges: %d MB, Elérhető: %d MB "+
+				"(összesen: %d MB, ebből %d MB használt, %d MB rendszer számára fenntartva)",
+			newReqMB,
+			usableMB-usedMB,
+			totalMB,
+			usedMB,
+			reservedMB,
+		), ""
+	}
+
+	// Soft warning: limits exceed total (overcommit)
+	_, currentLimitMB := m.CommittedMemory()
+	currentLimitMB -= releasedLimitMB
+	if newLimitMB > 0 && currentLimitMB+newLimitMB > totalMB {
+		warning = "Az alkalmazások csúcsterhelése meghaladhatja a rendelkezésre álló memóriát. " +
+			"Normál használat mellett ez nem okoz problémát."
+	}
+	return "", warning
+}

controller/internal/stacks/installed_test.go

@@ -407,7 +407,9 @@ func TestGroupE_RestartStackReachesTheRecorder(t *testing.T) {
 func TestGroupE_EveryBringUpPathCallsTheRecorder(t *testing.T) {
 	callers := map[string]bool{} // enclosing func name -> calls recordInstalledImages
 	fset := token.NewFileSet()
-	for _, src := range []string{"manager.go", "deploy.go"} {
+	// update.go since v0.237.0: the guarded update replaced UpdateStack, and it records in
+	// verifyAndConclude — only after the app's health is known (slice 4).
+	for _, src := range []string{"manager.go", "deploy.go", "update.go"} {
 		f, err := parser.ParseFile(fset, src, nil, 0)
 		if err != nil {
 			t.Fatal(err)
@@ -433,7 +435,7 @@ func TestGroupE_EveryBringUpPathCallsTheRecorder(t *testing.T) {
 			}
 		}
 	}
-	for _, want := range []string{"StartStack", "RestartStack", "UpdateStack", "runComposeDeploy"} {
+	for _, want := range []string{"StartStack", "RestartStack", "verifyAndConclude", "runComposeDeploy"} {
 		if !callers[want] {
 			t.Errorf("%s does not call recordInstalledImages — a bring-up path that records nothing leaves a stale record standing", want)
 		}

controller/internal/stacks/manager.go

@@ -2,6 +2,7 @@ package stacks
 
 import (
 	"bytes"
+	"context"
 	"fmt"
 	"log"
 	"os"
@@ -130,17 +131,28 @@ type HealthCheckDetail struct {
 
 // Stack represents a docker compose stack on disk.
 type Stack struct {
-	Name        string             `json:"name"`
-	Meta        Metadata           `json:"meta"`
-	ComposePath string             `json:"compose_path"`
-	State       ContainerState     `json:"state"`
-	Deployed    bool               `json:"deployed"` // Has app.yaml with deployed=true
-	Protected   bool               `json:"protected"`
-	Orphaned    bool               `json:"orphaned"` // Deployed but no catalog template
-	Containers  []ContainerInfo    `json:"containers"`
-	AppConfig   *AppConfig         `json:"app_config,omitempty"`
-	Deploying   bool               `json:"deploying"`              // compose up in progress
-	DeployError string             `json:"deploy_error,omitempty"` // last async deploy error
+	Name        string          `json:"name"`
+	Meta        Metadata        `json:"meta"`
+	ComposePath string          `json:"compose_path"`
+	State       ContainerState  `json:"state"`
+	Deployed    bool            `json:"deployed"` // Has app.yaml with deployed=true
+	Protected   bool            `json:"protected"`
+	Orphaned    bool            `json:"orphaned"` // Deployed but no catalog template
+	Containers  []ContainerInfo `json:"containers"`
+	AppConfig   *AppConfig      `json:"app_config,omitempty"`
+	Deploying   bool            `json:"deploying"`              // compose up in progress
+	DeployError string          `json:"deploy_error,omitempty"` // last async deploy error
+	// Updating / UpdatePhase / UpdatePhaseLabel / UpdateError (update arc slice 4, v0.237.0) are the
+	// guarded update's in-memory progress, the same shape as Deploying/DeployError: the API answers
+	// 202 at once and the page polls GET /api/stacks/{name}. See update.go.
+	Updating         bool   `json:"updating"`
+	UpdatePhase      string `json:"update_phase,omitempty"`
+	UpdatePhaseLabel string `json:"update_phase_label,omitempty"`
+	UpdateError      string `json:"update_error,omitempty"`
+	// HoldReason is the customer sentence of a hold in force on this app (a failed update or a failed
+	// restore), "" when none. Filled on every read from the ONE hold store, never cached, so the page
+	// and the API cannot show a hold the gate has already lifted — or miss one it enforces.
+	HoldReason  string             `json:"hold_reason,omitempty"`
 	HealthProbe *HealthProbeResult `json:"health_probe,omitempty"` // controller-side probe result
 	LastUpdated time.Time          `json:"last_updated"`
 	// RestartingSince (C9-F2) is when this stack was FIRST observed in StateRestarting during the
@@ -198,6 +210,16 @@ type Manager struct {
 	restartPolicyCache map[string]string
 	// execFn replaces execCommand's process boundary in tests; nil in production.
 	execFn func(name string, args ...string) (string, error)
+
+	// --- guarded update (slice 4, update.go) ---
+	updateGuards     UpdateGuards // init-only, SetUpdateGuards; nil ⇒ every update is REFUSED
+	updateComposeFn  func(dir string, env []string, args ...string) (string, error)
+	updateHealthFn   func(ctx context.Context, name string, timeout time.Duration) (bool, string)
+	updateMemoryFn   func(newReqMB, newLimitMB, releasedReqMB, releasedLimitMB int) (refusal, warning string)
+	updateDiskFreeFn func() (freeGiB float64, ok bool)
+	updateNowFn      func() time.Time
+	updateJournalMu  sync.Mutex
+	updateResume     []string // apps whose update was interrupted after `up`; resumed once guards exist
 	// inspectRestartPolicyFn is the docker-inspect seam for the above; nil in production
 	// (dockerRestartPolicy). Tests inject a scripted lookup and never touch docker.
 	inspectRestartPolicyFn func(containerName string) (string, error)
@@ -960,12 +982,15 @@ func aggregateState(containers []ContainerInfo, policyOf restartPolicyLookup) Co
 
 func (m *Manager) GetStacks() []Stack {
 	m.mu.RLock()
-	defer m.mu.RUnlock()
-
 	result := make([]Stack, 0, len(m.stacks))
 	for _, s := range m.stacks {
 		result = append(result, deepCopyStack(s))
 	}
+	g := m.updateGuards
+	m.mu.RUnlock()
+	for i := range result {
+		fillHoldReason(g, &result[i])
+	}
 
 	// Sort alphabetically by display name for consistent UI ordering
 	sort.Slice(result, func(i, j int) bool {
@@ -984,6 +1009,7 @@ func (m *Manager) GetStack(name string) (*Stack, bool) {
 		return nil, false
 	}
 	cp := deepCopyStack(s)
+	fillHoldReason(m.updateGuards, &cp)
 	return &cp, true
 }
 
@@ -1224,54 +1250,11 @@ func (m *Manager) RestartStack(name string) error {
 	return m.RefreshStatus()
 }
 
-func (m *Manager) UpdateStack(name string) error {
-	stack, ok := m.GetStack(name)
-	if !ok {
-		return fmt.Errorf("stack %q not found", name)
-	}
-
-	m.logger.Printf("[INFO] [stacks] Updating stack: %s", name)
-	start := time.Now()
-	dir := filepath.Dir(stack.ComposePath)
-
-	// v0.235.0 — ADVANCE THE PIN FIRST, AND RE-RENDER BEFORE THE PULL.
-	//
-	// This is the ONE act entitled to move a version; the freeze exists so that nothing else can.
-	// The ordering is load-bearing, not stylistic: `compose pull` and `up -d` act on the file on
-	// disk, so the catalog's current definition has to BE that file before either runs. Setting the
-	// pin afterwards would pull the frozen version and change nothing, while reporting success — and
-	// a button that lies is worse than a button that refuses.
-	//
-	// A FAILED PIN WRITE REFUSES THE UPDATE, deliberately the opposite of recordInstalledImages.
-	// That field is an observation and a failed write is a bookkeeping gap; this one is INTENT, and
-	// an update whose intent could not be recorded leaves the box running a version it has no record
-	// of choosing — the exact ambiguity R-166 closed for desired_state, one field over.
-	if err := m.advancePinToCatalog(name, dir); err != nil {
-		m.logger.Printf("[ERROR] [stacks] Stack %s update refused: %v", name, err)
-		return fmt.Errorf("updating stack %s: %w", name, err)
-	}
-
-	env := m.stackEnv(dir)
-
-	if m.isDebug() {
-		m.checkLocalImages(name, dir)
-	}
-
-	if _, err := m.composeExecCustomEnv(dir, env, "pull"); err != nil {
-		m.logger.Printf("[ERROR] [stacks] Stack %s update (pull) failed after %.1fs: %v", name, time.Since(start).Seconds(), err)
-		return fmt.Errorf("pulling images for %s: %w", name, err)
-	}
-
-	if _, err := m.composeExecCustomEnv(dir, env, "up", "-d", "--remove-orphans"); err != nil {
-		m.logger.Printf("[ERROR] [stacks] Stack %s update (up) failed after %.1fs: %v", name, time.Since(start).Seconds(), err)
-		return fmt.Errorf("recreating %s: %w", name, err)
-	}
-
-	m.logger.Printf("[INFO] [stacks] Stack %s updated successfully (took %.1fs)", name, time.Since(start).Seconds())
-	m.recordInstalledImages(name, dir, env)
-	m.logPostStartStatus(name, dir, env)
-	return m.RefreshStatus()
-}
+// UpdateStack was REMOVED in v0.237.0 (update arc slice 4). It advanced the pin, pulled, ran `up -d`
+// and reported success on the compose exit code — no copy first, no refusals, and HTTP 200 over a
+// crash loop (R-443). Its only caller was the API, which now runs StartGuardedUpdate (update.go).
+// Deleting it rather than leaving it is deliberate: an unguarded update path that still compiles is
+// one caller away from being the next R-439.
 
 func (m *Manager) GetLogs(name string, lines int) (string, error) {
 	stack, ok := m.GetStack(name)

controller/internal/stacks/update.go

@@ -0,0 +1,802 @@
+package stacks
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-controller/internal/system"
+)
+
+// ── The guarded update (update arc slice 4, controller v0.237.0) ─────────────────────────────────
+//
+// WHAT IT REPLACED. `UpdateStack` advanced the pin, pulled, ran `up -d` and returned — no copy first,
+// no check of memory, disk, a running backup or a held app, and a success the moment `up` returned.
+// SPIKE-app-update-2026-09-01 §4 measured that as HTTP 200 over an app that was already crash-looping
+// (R-443), and R-439 is that a held app could be updated at all.
+//
+// THE SEQUENCE, and the order is the point:
+//
+//	checking → backing-up (only if the proven copy is too old) → safety-dump → pinning → pulling
+//	→ starting → verifying → done | failed
+//
+//  1. The PRECONDITION is the app's existing verified backup (operator ruling 2026-09-02,
+//     09-update-architecture §3 decision 1): an openable Tier-2 unit with a PROVEN copy date. The
+//     same predicate that permits the destructive „Teljes visszaállítás" permits the update — the
+//     route back IS that restore, so an update without it has no route back.
+//  2. The safety dump is taken BEFORE the pin moves: it is "the state the customer was in a minute ago",
+//     and a minute later the migration may have run.
+//  3. The pin moves BEFORE the pull (v0.235.0's reason: pull and up act on the file on disk).
+//  4. A PULL failure puts the pin BACK — nothing ran, so reverting is safe and honest (Scenario E).
+//  5. A HEALTH failure leaves the pin where it is — the new version's migration may have run, and a
+//     pin claiming the old version would be a record of something untrue (Scenario F). The app is
+//     HELD STOPPED and the customer is told which backup it can be restored from.
+//
+// WHAT IT DELIBERATELY DOES NOT DO: put the old version back by itself. SPIKE-upgrade-test-2026-09-06
+// measured that whether the old image starts on migrated data depends on the app (PrivateBin yes,
+// Docmost and Nextcloud no) and cannot be predicted. The route back is the restore.
+//
+// CRASH SAFETY IS A JOURNAL, NOT A DEFER. A SIGKILL runs no deferred function (Campaign 8 fault 10),
+// so every phase is written to `update-journal.json` BEFORE it starts, and RecoverUpdates reads it at
+// the next startup (Scenario G) — the AppStopGuard pattern.
+
+// Update phases, as recorded in the journal and served as Stack.UpdatePhase.
+const (
+	UpdatePhaseChecking   = "checking"
+	UpdatePhaseBackingUp  = "backing-up"
+	UpdatePhaseSafetyDump = "safety-dump"
+	UpdatePhasePinning    = "pinning"
+	UpdatePhasePulling    = "pulling"
+	UpdatePhaseStarting   = "starting"
+	UpdatePhaseVerifying  = "verifying"
+	UpdatePhaseDone       = "done"
+	UpdatePhaseFailed     = "failed"
+)
+
+// updatePhaseLabels are the customer labels (slice 4 Part 4, exact). `pinning` has no row in the
+// specification — it is instantaneous and is the first step of fetching the new version, so it shares
+// the pull's label rather than inventing a sentence nobody would read.
+var updatePhaseLabels = map[string]string{
+	UpdatePhaseChecking:   "Ellenőrzés…",
+	UpdatePhaseBackingUp:  "Biztonsági mentés készül a frissítés előtt…",
+	UpdatePhaseSafetyDump: "Adatbázis pillanatkép…",
+	UpdatePhasePinning:    "Új verzió letöltése…",
+	UpdatePhasePulling:    "Új verzió letöltése…",
+	UpdatePhaseStarting:   "Indítás az új verzióval…",
+	UpdatePhaseVerifying:  "Működés ellenőrzése…",
+	UpdatePhaseDone:       "Frissítve",
+	UpdatePhaseFailed:     "A frissítés nem sikerült",
+}
+
+// UpdatePhaseLabel returns the customer label for a phase, "" for an unknown one.
+func UpdatePhaseLabel(phase string) string { return updatePhaseLabels[phase] }
+
+// Customer sentences. Named so tests compare against the constant, never a retyped literal (R-364).
+const (
+	MsgUpdateNoGuards      = "A frissítés nem indítható: a frissítés előtti biztonsági ellenőrzés nem érhető el ezen a szerveren."
+	MsgUpdateNotDeployed   = "Az alkalmazás nincs telepítve, ezért nem frissíthető."
+	MsgUpdateDeployingFmt  = "A(z) %s telepítése még folyamatban van — a frissítés utána indítható."
+	MsgUpdateAlreadyFmt    = "A(z) %s frissítése már folyamatban van."
+	MsgUpdateBusy          = "A frissítés most nem indítható: mentés/visszaállítás folyamatban. Próbáld újra, ha befejeződött."
+	MsgUpdateMigrating     = "A frissítés most nem indítható: adatáthelyezés folyamatban."
+	MsgUpdateNoBackupFmt   = "A(z) %s nem frissíthető, mert nincs olyan biztonsági mentése, amelyből vissza lehetne állítani. Kapcsold be a 2. mentést az alkalmazás mentési beállításainál a Mentések oldalon, és várd meg az első sikeres másolatot — utána a frissítés elindítható."
+	MsgUpdateDiskFmt       = "Nincs elég szabad hely a frissítéshez: %.1f GB szabad, az új verzió letöltéséhez legalább %.0f GB szükséges."
+	MsgUpdateBackupFailFmt = "A frissítés nem indult el, mert a frissítés előtti biztonsági mentés nem sikerült: %v. Az alkalmazás változatlanul fut tovább."
+	MsgUpdateBackupNoUnit  = "A frissítés nem indult el: a frissítés előtti mentés lefutott, de nem jött létre friss, visszaállítható másolat. Az alkalmazás változatlanul fut tovább."
+	MsgUpdateDumpFailFmt   = "A frissítés nem indult el, mert az adatbázis pillanatkép nem készült el: %v. Az alkalmazás változatlanul fut tovább."
+	MsgUpdatePinFailed     = "A frissítés nem indult el: az új verzió leírása nem olvasható be. Az alkalmazás változatlanul fut tovább."
+	MsgUpdateJournalFailed = "A frissítés nem indult el: a frissítés naplója nem menthető. Az alkalmazás változatlanul fut tovább."
+	MsgUpdatePullFailed    = "Az új verzió letöltése nem sikerült, ezért a frissítés elmaradt. Az alkalmazás a korábbi verzióval fut tovább."
+	MsgUpdateInterrupted   = "A frissítés megszakadt, mert a vezérlő újraindult, mielőtt az új verzió elindult volna. Az alkalmazás a korábbi verzióval fut tovább."
+	MsgUpdateHoldUnsaved   = "A frissítés nem sikerült, az alkalmazás le lett állítva, de a leállítás rögzítése nem sikerült. Ne indítsd újra — vedd fel velünk a kapcsolatot."
+)
+
+// updateDiskFloorGiB is the free space the Docker data root must have before a pull. A FIXED FLOOR,
+// stated as such: the new image set's size is not known without a registry query (the catalog
+// records tags, not sizes, and §8.1 of 09 already declines registry calls on the customer box), so
+// the rule is "not less than 2 GB", not "enough for these images".
+const updateDiskFloorGiB = 2.0
+
+// updateSettleWindow is the rule for an app with no .felhom.yml health check: every container
+// running, none restarting, for this long.
+const updateSettleWindow = 60 * time.Second
+
+// updatePollEvery is how often the health wait re-reads the stack.
+const updatePollEvery = 5 * time.Second
+
+// UpdateRestorePoint is the precondition answer, reduced to what the update needs.
+type UpdateRestorePoint struct {
+	Restorable bool      // an openable recovery unit exists in the Tier-2 copy
+	Proven     bool      // a copy actually succeeded (never an attempt clock)
+	ProvenAt   time.Time // when the data in that copy was last proven copied
+}
+
+// UpdateGuards is everything the update needs from the backup side. The stacks package cannot import
+// backup, so cmd/controller/main.go wires an adapter (TestSlice4_UpdateGuardsAreWiredAtStartup).
+type UpdateGuards interface {
+	HoldFor(name string) (bool, string)
+	Busy(name string) (bool, string)
+	RestorePoint(name string) (UpdateRestorePoint, error)
+	BackupNow(ctx context.Context, name string) error
+	SafetyDump(ctx context.Context, name string) ([]string, error)
+	HoldAfterFailedUpdate(name string, at, provenCopyAt time.Time) error
+}
+
+// SetUpdateGuards wires the backup side. INIT-ONLY. Unwired, every update is refused (fail closed):
+// an update that cannot see the backup cannot promise a route back.
+func (m *Manager) SetUpdateGuards(g UpdateGuards) {
+	m.mu.Lock()
+	m.updateGuards = g
+	m.mu.Unlock()
+}
+
+func (m *Manager) guards() UpdateGuards {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+	return m.updateGuards
+}
+
+func fillHoldReason(g UpdateGuards, st *Stack) {
+	if g == nil || st == nil || !st.Deployed {
+		return
+	}
+	if held, why := g.HoldFor(st.Name); held {
+		st.HoldReason = why
+	}
+}
+
+// UpdateRefusal is a refusal taken before anything moved. Reason is a stable key for logs and tests;
+// Message is the customer sentence.
+type UpdateRefusal struct {
+	Reason  string
+	Message string
+}
+
+func (r *UpdateRefusal) Error() string { return r.Message }
+
+func (m *Manager) now() time.Time {
+	if m.updateNowFn != nil {
+		return m.updateNowFn()
+	}
+	return time.Now()
+}
+
+func (m *Manager) refuseUpdate(name, reason, msg, detail string) *UpdateRefusal {
+	m.logger.Printf("[ERROR] [stacks] update %s REFUSED (%s): %s", name, reason, detail)
+	return &UpdateRefusal{Reason: reason, Message: msg}
+}
+
+// UpdatePreflight runs every CHEAP refusal (slice 4 Part 1 + the precondition's existence), in order,
+// and returns the first. Nothing is moved and nothing is recorded by it. The router calls it before
+// recording the customer's intent, so an update that was never going to happen records nothing.
+func (m *Manager) UpdatePreflight(name string) *UpdateRefusal {
+	st, ok := m.GetStack(name)
+	if !ok {
+		return m.refuseUpdate(name, "not_found", fmt.Sprintf("stack %q not found", name), "no such stack")
+	}
+	if !st.Deployed {
+		return m.refuseUpdate(name, "not_deployed", MsgUpdateNotDeployed, "not deployed")
+	}
+	g := m.guards()
+	if g == nil {
+		return m.refuseUpdate(name, "guards_unwired", MsgUpdateNoGuards, "no UpdateGuards wired — fail closed")
+	}
+	if st.Deploying {
+		return m.refuseUpdate(name, "deploying", fmt.Sprintf(MsgUpdateDeployingFmt, name), "a deploy is in progress")
+	}
+	if st.Updating {
+		return m.refuseUpdate(name, "updating", fmt.Sprintf(MsgUpdateAlreadyFmt, name), "an update is already in progress")
+	}
+	if held, why := g.HoldFor(name); held {
+		return m.refuseUpdate(name, "held", why, "the app is held")
+	}
+	if busy, why := g.Busy(name); busy {
+		return m.refuseUpdate(name, "busy", MsgUpdateBusy, why)
+	}
+	if m.IsMigrating() {
+		return m.refuseUpdate(name, "migrating", MsgUpdateMigrating, "a data migration is running")
+	}
+	rp, err := g.RestorePoint(name)
+	if err != nil || !rp.Restorable || !rp.Proven {
+		return m.refuseUpdate(name, "no_backup", fmt.Sprintf(MsgUpdateNoBackupFmt, name),
+			fmt.Sprintf("no restorable proven Tier-2 unit (restorable=%v proven=%v err=%v)", rp.Restorable, rp.Proven, err))
+	}
+	if ref := m.updateMemoryRefusal(name, st); ref != nil {
+		return ref
+	}
+	free, known := m.updateDiskFree()
+	switch {
+	case !known:
+		m.logger.Printf("[WARN] [stacks] update %s: free space on the Docker data root is unreadable — proceeding without the %.0f GB floor", name, updateDiskFloorGiB)
+	case free < updateDiskFloorGiB:
+		return m.refuseUpdate(name, "disk", fmt.Sprintf(MsgUpdateDiskFmt, free, updateDiskFloorGiB),
+			fmt.Sprintf("%.2f GiB free on the Docker data root, floor %.0f GiB (fixed floor — image size unknown)", free, updateDiskFloorGiB))
+	}
+	return nil
+}
+
+// updateMemoryRefusal applies the deploy's memory check to the NEW template's request, releasing the
+// app's CURRENT request first (an update replaces it). An unknown new request proceeds with a WARN.
+func (m *Manager) updateMemoryRefusal(name string, st *Stack) *UpdateRefusal {
+	catPath := m.CatalogTemplatePath(name, ".felhom.yml")
+	if _, err := os.Stat(catPath); err != nil {
+		m.logger.Printf("[WARN] [stacks] update %s: the new template's memory request is unknown (%v) — proceeding without the memory check", name, err)
+		return nil
+	}
+	newMeta := LoadMetadata(filepath.Dir(catPath))
+	newReq, newLim := ParseMemoryMB(newMeta.Resources.MemRequest), ParseMemoryMB(newMeta.Resources.MemLimit)
+	if newReq == 0 {
+		m.logger.Printf("[WARN] [stacks] update %s: the new template declares no memory request — proceeding without the memory check", name)
+		return nil
+	}
+	oldReq, oldLim := ParseMemoryMB(st.Meta.Resources.MemRequest), ParseMemoryMB(st.Meta.Resources.MemLimit)
+	verdict := m.updateMemoryFn
+	if verdict == nil {
+		verdict = m.memoryVerdict
+	}
+	if refusal, _ := verdict(newReq, newLim, oldReq, oldLim); refusal != "" {
+		return m.refuseUpdate(name, "memory", refusal, fmt.Sprintf("new_req=%dMB replacing %dMB does not fit", newReq, oldReq))
+	}
+	return nil
+}
+
+func (m *Manager) updateDiskFree() (float64, bool) {
+	if m.updateDiskFreeFn != nil {
+		return m.updateDiskFreeFn()
+	}
+	du := system.GetDiskUsage(system.DockerVolumePath)
+	if du == nil {
+		return 0, false
+	}
+	return du.AvailGB, true
+}
+
+// StartGuardedUpdate re-checks the cheap refusals, claims the Updating flag atomically and launches the
+// job. It returns as soon as the job has STARTED — the result arrives on GET /api/stacks/{name}.
+func (m *Manager) StartGuardedUpdate(name string) error {
+	if ref := m.UpdatePreflight(name); ref != nil {
+		return ref
+	}
+	m.mu.Lock()
+	s, ok := m.stacks[name]
+	if !ok {
+		m.mu.Unlock()
+		return &UpdateRefusal{Reason: "not_found", Message: fmt.Sprintf("stack %q not found", name)}
+	}
+	// A second press between the preflight and here is the race this lock closes.
+	if s.Updating || s.Deploying {
+		m.mu.Unlock()
+		return m.refuseUpdate(name, "updating", fmt.Sprintf(MsgUpdateAlreadyFmt, name), "lost the race for the Updating flag")
+	}
+	s.Updating, s.UpdateError = true, ""
+	s.UpdatePhase, s.UpdatePhaseLabel = UpdatePhaseChecking, UpdatePhaseLabel(UpdatePhaseChecking)
+	m.mu.Unlock()
+
+	m.logger.Printf("[INFO] [stacks] update %s: accepted — guarded update started", name)
+	go m.runGuardedUpdate(context.Background(), name)
+	return nil
+}
+
+// IsUpdating reports whether a guarded update is in progress for the app.
+func (m *Manager) IsUpdating(name string) bool {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+	s, ok := m.stacks[name]
+	return ok && s.Updating
+}
+
+// UpdatingStacks is the set of apps an update is currently moving — for the dead-app alarm, which
+// must not count an app the update itself is recreating (R-330's class, a third mechanism).
+func (m *Manager) UpdatingStacks() map[string]bool {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+	var out map[string]bool
+	for name, s := range m.stacks {
+		if s.Updating {
+			if out == nil {
+				out = map[string]bool{}
+			}
+			out[name] = true
+		}
+	}
+	return out
+}
+
+func (m *Manager) setUpdatePhase(name, phase string) {
+	m.mu.Lock()
+	if s, ok := m.stacks[name]; ok {
+		s.UpdatePhase, s.UpdatePhaseLabel = phase, UpdatePhaseLabel(phase)
+	}
+	m.mu.Unlock()
+}
+
+// finishUpdate is the ONE place Updating goes false. msg is the customer sentence on failure.
+func (m *Manager) finishUpdate(name, phase, msg string) {
+	m.mu.Lock()
+	if s, ok := m.stacks[name]; ok {
+		s.Updating = false
+		s.UpdatePhase, s.UpdatePhaseLabel = phase, UpdatePhaseLabel(phase)
+		s.UpdateError = msg
+	}
+	m.mu.Unlock()
+}
+
+func (m *Manager) updateCompose(dir string, env []string, args ...string) (string, error) {
+	if m.updateComposeFn != nil {
+		return m.updateComposeFn(dir, env, args...)
+	}
+	return m.composeExecCustomEnv(dir, env, args...)
+}
+
+func (m *Manager) updateHealth(ctx context.Context, name string, timeout time.Duration) (bool, string) {
+	if m.updateHealthFn != nil {
+		return m.updateHealthFn(ctx, name, timeout)
+	}
+	return m.waitUpdateHealthy(ctx, name, timeout)
+}
+
+func (m *Manager) healthTimeout() time.Duration {
+	if m.cfg == nil {
+		return 5 * time.Minute
+	}
+	return m.cfg.Update.HealthTimeoutDuration()
+}
+
+func (m *Manager) backupMaxAge() time.Duration {
+	if m.cfg == nil {
+		return 24 * time.Hour
+	}
+	return m.cfg.Update.BackupMaxAgeDuration()
+}
+
+// pre-update copies, kept in the stack dir so they travel with it. Neither name is one the syncer
+// copies (it copies exactly docker-compose.yml and .felhom.yml).
+const (
+	preUpdateComposeFile = "pre-update-compose.yml"
+	preUpdateAppliedFile = "pre-update-applied.yml"
+)
+
+func (m *Manager) runGuardedUpdate(ctx context.Context, name string) {
+	start := m.now()
+	st, ok := m.GetStack(name)
+	if !ok {
+		m.finishUpdate(name, UpdatePhaseFailed, fmt.Sprintf("stack %q not found", name))
+		return
+	}
+	dir := filepath.Dir(st.ComposePath)
+	g := m.guards()
+	entry := updateJournalEntry{StartedAt: start}
+	fail := func(msg, detail string) {
+		m.logger.Printf("[ERROR] [stacks] update %s FAILED in phase %s after %s — nothing was moved: %s", name, entry.Phase, m.now().Sub(start).Round(time.Millisecond), detail)
+		m.clearJournal(name)
+		m.finishUpdate(name, UpdatePhaseFailed, msg)
+	}
+
+	if !m.enterUpdatePhase(name, &entry, UpdatePhaseChecking) {
+		m.finishUpdate(name, UpdatePhaseFailed, MsgUpdateJournalFailed)
+		return
+	}
+	if g == nil {
+		fail(MsgUpdateNoGuards, "no UpdateGuards wired")
+		return
+	}
+	rp, err := g.RestorePoint(name)
+	if err != nil || !rp.Restorable || !rp.Proven {
+		fail(fmt.Sprintf(MsgUpdateNoBackupFmt, name), fmt.Sprintf("precondition vanished: restorable=%v proven=%v err=%v", rp.Restorable, rp.Proven, err))
+		return
+	}
+
+	maxAge := m.backupMaxAge()
+	if age := start.Sub(rp.ProvenAt); age > maxAge {
+		m.logger.Printf("[INFO] [stacks] update %s: the proven copy is %s old (limit %s) — backing up first", name, age.Round(time.Minute), maxAge)
+		if !m.enterUpdatePhase(name, &entry, UpdatePhaseBackingUp) {
+			fail(MsgUpdateJournalFailed, "journal write failed")
+			return
+		}
+		if err := g.BackupNow(ctx, name); err != nil {
+			fail(fmt.Sprintf(MsgUpdateBackupFailFmt, err), "pre-update backup: "+err.Error())
+			return
+		}
+		rp, err = g.RestorePoint(name)
+		if err != nil || !rp.Restorable || !rp.Proven || m.now().Sub(rp.ProvenAt) > maxAge {
+			fail(MsgUpdateBackupNoUnit, fmt.Sprintf("after the backup: restorable=%v proven=%v at=%s err=%v", rp.Restorable, rp.Proven, rp.ProvenAt.Format(time.RFC3339), err))
+			return
+		}
+	} else {
+		m.logger.Printf("[INFO] [stacks] update %s: precondition met — proven copy from %s (%s old, limit %s)", name, rp.ProvenAt.UTC().Format(time.RFC3339), age.Round(time.Minute), maxAge)
+	}
+	entry.ProvenCopyAt = rp.ProvenAt.UTC().Format(time.RFC3339)
+
+	// SAFETY DUMP BEFORE THE PIN MOVES — "a minute ago", before any migration can have run.
+	if !m.enterUpdatePhase(name, &entry, UpdatePhaseSafetyDump) {
+		fail(MsgUpdateJournalFailed, "journal write failed")
+		return
+	}
+	paths, err := g.SafetyDump(ctx, name)
+	if err != nil {
+		fail(fmt.Sprintf(MsgUpdateDumpFailFmt, err), "safety dump: "+err.Error())
+		return
+	}
+	m.logger.Printf("[INFO] [stacks] update %s: safety dump done (%d file(s)) %v", name, len(paths), paths)
+
+	// PINNING — the previous definition is copied aside and journaled BEFORE the pin moves, so a crash
+	// at any later instant can put it back (Scenario G).
+	prevLive, err := os.ReadFile(st.ComposePath)
+	if err != nil {
+		fail(MsgUpdatePinFailed, "reading the live compose file: "+err.Error())
+		return
+	}
+	if err := os.WriteFile(filepath.Join(dir, preUpdateComposeFile), prevLive, 0o644); err != nil {
+		fail(MsgUpdateJournalFailed, "saving the pre-update compose copy: "+err.Error())
+		return
+	}
+	entry.PrevCompose = filepath.Join(dir, preUpdateComposeFile)
+	if applied, aerr := LoadAppliedDefinition(dir); aerr == nil {
+		if err := os.WriteFile(filepath.Join(dir, preUpdateAppliedFile), applied, 0o644); err == nil {
+			entry.PrevApplied = filepath.Join(dir, preUpdateAppliedFile)
+		}
+	}
+	if cfg := LoadAppConfig(dir); cfg != nil && len(cfg.PinnedImages) > 0 {
+		entry.PrevPin = map[string]string{}
+		for k, v := range cfg.PinnedImages {
+			entry.PrevPin[k] = v
+		}
+	}
+	if !m.enterUpdatePhase(name, &entry, UpdatePhasePinning) {
+		m.removePreUpdateCopies(dir)
+		fail(MsgUpdateJournalFailed, "journal write failed")
+		return
+	}
+	if err := m.advancePinToCatalog(name, dir); err != nil {
+		m.pinBack(name, dir, entry)
+		fail(MsgUpdatePinFailed, "advancing the pin: "+err.Error())
+		return
+	}
+
+	env := m.stackEnv(dir)
+	if !m.enterUpdatePhase(name, &entry, UpdatePhasePulling) {
+		m.pinBack(name, dir, entry)
+		fail(MsgUpdateJournalFailed, "journal write failed")
+		return
+	}
+	if _, err := m.updateCompose(dir, env, "pull"); err != nil {
+		// Scenario E: NOTHING RAN. The containers are the old ones and still running, so the honest
+		// state is the old pin and the old file — put both back.
+		m.pinBack(name, dir, entry)
+		m.logger.Printf("[ERROR] [stacks] update %s: pull failed — pin and definition PUT BACK; the app was not touched. Docker said: %v", name, err)
+		fail(MsgUpdatePullFailed, "pull failed: "+err.Error())
+		return
+	}
+
+	if !m.enterUpdatePhase(name, &entry, UpdatePhaseStarting) {
+		m.failAndHold(ctx, name, dir, env, rp.ProvenAt, "journal write failed before up")
+		return
+	}
+	if _, err := m.updateCompose(dir, env, "up", "-d", "--remove-orphans"); err != nil {
+		// Containers may already have been recreated on the new image — something may have run.
+		m.failAndHold(ctx, name, dir, env, rp.ProvenAt, "compose up failed: "+err.Error())
+		return
+	}
+	m.verifyAndConclude(ctx, name, dir, env, rp.ProvenAt, start, &entry)
+}
+
+// verifyAndConclude is the TRUTH half (R-443): success is declared only after the app's health is
+// known, and a failure holds the app.
+func (m *Manager) verifyAndConclude(ctx context.Context, name, dir string, env []string, provenAt, start time.Time, entry *updateJournalEntry) {
+	if !m.enterUpdatePhase(name, entry, UpdatePhaseVerifying) {
+		m.logger.Printf("[ERROR] [stacks] update %s: could not journal the verifying phase — verifying anyway", name)
+	}
+	timeout := m.healthTimeout()
+	waitStart := m.now()
+	healthy, detail := m.updateHealth(ctx, name, timeout)
+	if !healthy {
+		m.failAndHold(ctx, name, dir, env, provenAt, "not healthy: "+detail)
+		return
+	}
+	m.logger.Printf("[INFO] [stacks] update %s: healthy after %s (%s)", name, m.now().Sub(waitStart).Round(time.Second), detail)
+	m.recordInstalledImages(name, dir, env)
+	_ = m.RefreshStatus()
+	m.clearJournal(name)
+	m.removePreUpdateCopies(dir)
+	m.finishUpdate(name, UpdatePhaseDone, "")
+	m.logger.Printf("[INFO] [stacks] update %s: DONE in %s", name, m.now().Sub(start).Round(time.Second))
+}
+
+// failAndHold is Scenario F: stop the app, record the hold, tell the customer the route back.
+func (m *Manager) failAndHold(ctx context.Context, name, dir string, env []string, provenAt time.Time, why string) {
+	m.logger.Printf("[ERROR] [stacks] update %s FAILED after the new version was started: %s — stopping and HOLDING the app; the pin stays on the new version (its migration may have run)", name, why)
+	if _, err := m.updateCompose(dir, env, "down"); err != nil {
+		m.logger.Printf("[ERROR] [stacks] update %s: stopping the failed app also failed: %v", name, err)
+	}
+	msg := MsgUpdateHoldUnsaved
+	if g := m.guards(); g == nil {
+		m.logger.Printf("[ERROR] [stacks] update %s: no UpdateGuards — the hold CANNOT be recorded", name)
+	} else if err := g.HoldAfterFailedUpdate(name, m.now(), provenAt); err != nil {
+		m.logger.Printf("[ERROR] [stacks] update %s: %v", name, err)
+	} else if _, why := g.HoldFor(name); why != "" {
+		msg = why
+	}
+	_ = m.RefreshStatus()
+	m.clearJournal(name)
+	m.removePreUpdateCopies(dir)
+	m.finishUpdate(name, UpdatePhaseFailed, msg)
+}
+
+// pinBack restores the pin, the stored definition and the live file from the journaled copies.
+func (m *Manager) pinBack(name, dir string, entry updateJournalEntry) {
+	prevLive, lerr := os.ReadFile(entry.PrevCompose)
+	if lerr != nil {
+		m.logger.Printf("[ERROR] [stacks] update %s: cannot read the pre-update compose copy (%v) — the definition could NOT be put back", name, lerr)
+	}
+	if len(entry.PrevPin) > 0 {
+		applied := prevLive
+		if entry.PrevApplied != "" {
+			if b, err := os.ReadFile(entry.PrevApplied); err == nil {
+				applied = b
+			}
+		}
+		if err := m.SetPin(name, dir, entry.PrevPin, applied); err != nil {
+			m.logger.Printf("[ERROR] [stacks] update %s: putting the pin back failed: %v", name, err)
+		}
+	}
+	if lerr == nil {
+		if err := os.WriteFile(ComposePathIn(dir), prevLive, 0o644); err != nil {
+			m.logger.Printf("[ERROR] [stacks] update %s: re-rendering the previous definition failed: %v", name, err)
+		}
+	}
+	m.removePreUpdateCopies(dir)
+	m.logger.Printf("[INFO] [stacks] update %s: pin and definition PUT BACK to the pre-update version (%s)", name, summarisePin(entry.PrevPin))
+}
+
+func (m *Manager) removePreUpdateCopies(dir string) {
+	_ = os.Remove(filepath.Join(dir, preUpdateComposeFile))
+	_ = os.Remove(filepath.Join(dir, preUpdateAppliedFile))
+}
+
+// waitUpdateHealthy is the production health wait: the app's own .felhom.yml health check through the
+// existing probe, or — for an app with none — every container running and none restarting for
+// updateSettleWindow. NEVER the compose exit code, and never logPostStartStatus's delayed log line.
+func (m *Manager) waitUpdateHealthy(ctx context.Context, name string, timeout time.Duration) (bool, string) {
+	deadline := m.now().Add(timeout)
+	var runningSince time.Time
+	last := "no observation yet"
+	for {
+		_ = m.RefreshStatus()
+		st, ok := m.GetStack(name)
+		switch {
+		case !ok:
+			last = "stack vanished"
+			runningSince = time.Time{}
+		case st.State == StateRunning:
+			if hc := st.Meta.HealthCheck; hc != nil && len(hc.Checks) > 0 {
+				if c := findProbeContainer(name, st.Containers); c != "" {
+					res := m.runChecks(probeTarget{stackName: name, containerName: c, checks: hc.Checks})
+					m.mu.Lock()
+					if s, ok := m.stacks[name]; ok {
+						s.HealthProbe = res
+					}
+					m.mu.Unlock()
+					if res.Healthy {
+						return true, "the app's health check passed"
+					}
+					last = "health check failing"
+				} else {
+					last = "no probe container"
+				}
+			} else {
+				if runningSince.IsZero() {
+					runningSince = m.now()
+				}
+				if m.now().Sub(runningSince) >= updateSettleWindow {
+					return true, fmt.Sprintf("all containers running, none restarting, for %s (no health check declared)", updateSettleWindow)
+				}
+				last = "running, settling"
+			}
+		default:
+			runningSince = time.Time{}
+			last = "state " + string(st.State)
+		}
+		if !m.now().Before(deadline) {
+			return false, fmt.Sprintf("not healthy within %s (last: %s)", timeout, last)
+		}
+		select {
+		case <-ctx.Done():
+			return false, "cancelled: " + ctx.Err().Error()
+		case <-time.After(updatePollEvery):
+		}
+	}
+}
+
+// ── the journal ──────────────────────────────────────────────────────────────────────────────────
+
+type updateJournalEntry struct {
+	Phase        string            `json:"phase"`
+	StartedAt    time.Time         `json:"started_at"`
+	PrevPin      map[string]string `json:"prev_pin,omitempty"`
+	PrevCompose  string            `json:"prev_compose,omitempty"`
+	PrevApplied  string            `json:"prev_applied,omitempty"`
+	ProvenCopyAt string            `json:"proven_copy_at,omitempty"`
+}
+
+type updateJournal struct {
+	Updates map[string]updateJournalEntry `json:"updates"`
+}
+
+func (m *Manager) updateJournalPath() string {
+	return filepath.Join(m.cfg.Paths.DataDir, "update-journal.json")
+}
+
+func (m *Manager) readUpdateJournal() updateJournal {
+	j := updateJournal{Updates: map[string]updateJournalEntry{}}
+	data, err := os.ReadFile(m.updateJournalPath())
+	if err != nil {
+		return j
+	}
+	if err := json.Unmarshal(data, &j); err != nil {
+		m.logger.Printf("[WARN] [stacks] update journal at %s is corrupt (%v) — quarantining", m.updateJournalPath(), err)
+		_ = os.Rename(m.updateJournalPath(), fmt.Sprintf("%s.corrupt-%d", m.updateJournalPath(), time.Now().Unix()))
+		return updateJournal{Updates: map[string]updateJournalEntry{}}
+	}
+	if j.Updates == nil {
+		j.Updates = map[string]updateJournalEntry{}
+	}
+	return j
+}
+
+// writeUpdateJournal is atomic and fsynced (the AppStopGuard shape): the point is surviving a power cut.
+func (m *Manager) writeUpdateJournal(j updateJournal) error {
+	p := m.updateJournalPath()
+	if len(j.Updates) == 0 {
+		if err := os.Remove(p); err != nil && !os.IsNotExist(err) {
+			return err
+		}
+		return nil
+	}
+	data, err := json.MarshalIndent(j, "", "  ")
+	if err != nil {
+		return err
+	}
+	if err := os.MkdirAll(filepath.Dir(p), 0o755); err != nil {
+		return err
+	}
+	tmp := p + ".tmp"
+	f, err := os.OpenFile(tmp, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o600)
+	if err != nil {
+		return err
+	}
+	if _, err := f.Write(data); err != nil {
+		f.Close()
+		os.Remove(tmp)
+		return err
+	}
+	if err := f.Sync(); err != nil {
+		f.Close()
+		os.Remove(tmp)
+		return err
+	}
+	if err := f.Close(); err != nil {
+		os.Remove(tmp)
+		return err
+	}
+	return os.Rename(tmp, p)
+}
+
+// enterUpdatePhase journals the phase BEFORE it starts and mirrors it for the UI. False means the
+// journal could not be written — the caller must not perform a mutation it could not record.
+func (m *Manager) enterUpdatePhase(name string, entry *updateJournalEntry, phase string) bool {
+	entry.Phase = phase
+	m.updateJournalMu.Lock()
+	j := m.readUpdateJournal()
+	j.Updates[name] = *entry
+	err := m.writeUpdateJournal(j)
+	m.updateJournalMu.Unlock()
+	m.setUpdatePhase(name, phase)
+	if err != nil {
+		m.logger.Printf("[ERROR] [stacks] update %s: journal write for phase %s failed: %v", name, phase, err)
+		return false
+	}
+	m.logger.Printf("[INFO] [stacks] update %s: phase %s", name, phase)
+	return true
+}
+
+func (m *Manager) clearJournal(name string) {
+	m.updateJournalMu.Lock()
+	defer m.updateJournalMu.Unlock()
+	j := m.readUpdateJournal()
+	delete(j.Updates, name)
+	if err := m.writeUpdateJournal(j); err != nil {
+		m.logger.Printf("[ERROR] [stacks] update %s: clearing the journal entry failed: %v", name, err)
+	}
+}
+
+// RecoverUpdates reads the journal at startup (Scenario G). Call it BEFORE the boot reconciler.
+//
+//   - interrupted BEFORE the pin moved (checking, backing-up, safety-dump): nothing moved — the entry is
+//     dropped and the app carries the "interrupted" sentence;
+//   - interrupted while pinning or pulling: nothing RAN — the pin and definition are put back (as E);
+//   - interrupted while starting or verifying: something may have run — the app is marked Updating
+//     (so the boot sweep and the dead-app alarm leave it alone) and queued for ResumeInterruptedUpdates,
+//     which re-runs `up -d` and the health wait, ending in A or F.
+//
+// It needs no backup wiring, because nothing here holds an app — that is left to the resumed job.
+func (m *Manager) RecoverUpdates() []string {
+	m.updateJournalMu.Lock()
+	j := m.readUpdateJournal()
+	m.updateJournalMu.Unlock()
+	if len(j.Updates) == 0 {
+		return nil
+	}
+	var resumed []string
+	for name, e := range j.Updates {
+		st, ok := m.GetStack(name)
+		if !ok {
+			m.logger.Printf("[WARN] [stacks] update recovery: %s is in the journal (phase %s) but no longer exists — dropping the entry", name, e.Phase)
+			m.clearJournal(name)
+			continue
+		}
+		dir := filepath.Dir(st.ComposePath)
+		switch e.Phase {
+		case UpdatePhaseChecking, UpdatePhaseBackingUp, UpdatePhaseSafetyDump:
+			m.logger.Printf("[WARN] [stacks] update recovery: %s was interrupted in %s (started %s) — nothing had moved; dropping it", name, e.Phase, e.StartedAt.Format(time.RFC3339))
+			m.clearJournal(name)
+			m.finishUpdate(name, UpdatePhaseFailed, MsgUpdateInterrupted)
+		case UpdatePhasePinning, UpdatePhasePulling:
+			m.logger.Printf("[WARN] [stacks] update recovery: %s was interrupted in %s (started %s) — nothing had run; putting the pin back", name, e.Phase, e.StartedAt.Format(time.RFC3339))
+			m.pinBack(name, dir, e)
+			m.clearJournal(name)
+			m.finishUpdate(name, UpdatePhaseFailed, MsgUpdateInterrupted)
+		case UpdatePhaseStarting, UpdatePhaseVerifying:
+			m.logger.Printf("[WARN] [stacks] update recovery: %s was interrupted in %s (started %s) — the new version may have run; marking it Updating and RESUMING the health wait", name, e.Phase, e.StartedAt.Format(time.RFC3339))
+			m.mu.Lock()
+			if s, ok := m.stacks[name]; ok {
+				s.Updating, s.UpdateError = true, ""
+				s.UpdatePhase, s.UpdatePhaseLabel = UpdatePhaseVerifying, UpdatePhaseLabel(UpdatePhaseVerifying)
+			}
+			m.updateResume = append(m.updateResume, name)
+			m.mu.Unlock()
+			resumed = append(resumed, name)
+		default:
+			m.logger.Printf("[WARN] [stacks] update recovery: %s has unknown phase %q — dropping the entry", name, e.Phase)
+			m.clearJournal(name)
+		}
+	}
+	return resumed
+}
+
+// ResumeInterruptedUpdates continues the updates RecoverUpdates queued, once the backup side is wired
+// (a resumed update that fails must be able to HOLD). Returns how many were resumed.
+func (m *Manager) ResumeInterruptedUpdates(ctx context.Context) int {
+	m.mu.Lock()
+	names := m.updateResume
+	m.updateResume = nil
+	m.mu.Unlock()
+	for _, name := range names {
+		st, ok := m.GetStack(name)
+		if !ok {
+			m.finishUpdate(name, UpdatePhaseFailed, MsgUpdateInterrupted)
+			continue
+		}
+		m.updateJournalMu.Lock()
+		e, ok := m.readUpdateJournal().Updates[name]
+		m.updateJournalMu.Unlock()
+		if !ok {
+			m.finishUpdate(name, UpdatePhaseFailed, MsgUpdateInterrupted)
+			continue
+		}
+		provenAt, _ := time.Parse(time.RFC3339, e.ProvenCopyAt)
+		dir := filepath.Dir(st.ComposePath)
+		go func(name, dir string, e updateJournalEntry, provenAt time.Time) {
+			env := m.stackEnv(dir)
+			m.logger.Printf("[INFO] [stacks] update %s: resuming after a controller restart — `up -d` then the health wait", name)
+			if _, err := m.updateCompose(dir, env, "up", "-d", "--remove-orphans"); err != nil {
+				m.failAndHold(ctx, name, dir, env, provenAt, "resumed compose up failed: "+err.Error())
+				return
+			}
+			m.verifyAndConclude(ctx, name, dir, env, provenAt, e.StartedAt, &e)
+		}(name, dir, e, provenAt)
+	}
+	return len(names)
+}

controller/internal/stacks/update_test.go

@@ -0,0 +1,569 @@
+package stacks
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+// Update arc slice 4 — the guarded update. Scenarios A–G of the task, through the real job
+// (runGuardedUpdate) with the four process boundaries injected: compose, the health wait, the backup
+// side (UpdateGuards) and the clock. Every assertion reads the EFFECT back — the pin in app.yaml, the
+// bytes of the live compose file, the journal on disk, Updating/UpdatePhase/UpdateError — never "no error".
+
+var slice4T0 = time.Date(2026, 9, 13, 10, 0, 0, 0, time.UTC)
+
+type fakeGuards struct {
+	mu            sync.Mutex
+	calls         []string
+	held          bool
+	holdWhy       string
+	busy          bool
+	rp            UpdateRestorePoint
+	rpErr         error
+	rpAfterBackup *UpdateRestorePoint
+	backupErr     error
+	dumpErr       error
+	holdErr       error
+	holdProvenAt  time.Time
+	pinAtDump     string
+	stackDir      string
+}
+
+func (f *fakeGuards) note(c string) { f.mu.Lock(); f.calls = append(f.calls, c); f.mu.Unlock() }
+func (f *fakeGuards) callList() []string {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return append([]string(nil), f.calls...)
+}
+func (f *fakeGuards) HoldFor(string) (bool, string) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return f.held, f.holdWhy
+}
+func (f *fakeGuards) Busy(string) (bool, string) { return f.busy, "fake busy" }
+func (f *fakeGuards) RestorePoint(string) (UpdateRestorePoint, error) {
+	f.note("RestorePoint")
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return f.rp, f.rpErr
+}
+func (f *fakeGuards) BackupNow(context.Context, string) error {
+	f.note("BackupNow")
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if f.backupErr == nil && f.rpAfterBackup != nil {
+		f.rp = *f.rpAfterBackup
+	}
+	return f.backupErr
+}
+func (f *fakeGuards) SafetyDump(context.Context, string) ([]string, error) {
+	f.note("SafetyDump")
+	if cfg := LoadAppConfig(f.stackDir); cfg != nil {
+		f.mu.Lock()
+		f.pinAtDump = cfg.PinnedImages["web"]
+		f.mu.Unlock()
+	}
+	return []string{"/fake/pre-restore-x.sql"}, f.dumpErr
+}
+func (f *fakeGuards) HoldAfterFailedUpdate(_ string, _ time.Time, provenAt time.Time) error {
+	f.note("HoldAfterFailedUpdate")
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if f.holdErr != nil {
+		return f.holdErr
+	}
+	f.held, f.holdWhy, f.holdProvenAt = true, "HELD-SENTENCE", provenAt
+	return nil
+}
+
+type composeRec struct {
+	mu    sync.Mutex
+	calls []string
+	fail  map[string]error // first arg → error
+}
+
+func (c *composeRec) fn(_ string, _ []string, args ...string) (string, error) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	c.calls = append(c.calls, strings.Join(args, " "))
+	if err := c.fail[args[0]]; err != nil {
+		return "", err
+	}
+	return "", nil
+}
+func (c *composeRec) list() []string {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return append([]string(nil), c.calls...)
+}
+
+// newSlice4Manager: a pinned nextcloud on the OLD version, the catalog offering the NEW one, fresh
+// proven copy, and every boundary faked. Returns the manager, its stack dir, the guards and compose.
+func newSlice4Manager(t *testing.T) (*Manager, string, *fakeGuards, *composeRec) {
+	t.Helper()
+	m, dir := newPinManager(t, pinTplOld, pinTplNew,
+		"deployed: true\nenv: {}\npinned_images:\n    web: nextcloud:31.0.14-apache\n")
+	mustWrite(t, AppliedComposePath(dir), pinTplOld)
+	g := &fakeGuards{rp: UpdateRestorePoint{Restorable: true, Proven: true, ProvenAt: slice4T0.Add(-1 * time.Hour)}, stackDir: dir}
+	c := &composeRec{fail: map[string]error{}}
+	m.updateGuards = g
+	m.updateComposeFn = c.fn
+	m.updateHealthFn = func(context.Context, string, time.Duration) (bool, string) { return true, "fake healthy" }
+	m.updateMemoryFn = func(int, int, int, int) (string, string) { return "", "" }
+	m.updateDiskFreeFn = func() (float64, bool) { return 50, true }
+	m.updateNowFn = func() time.Time { return slice4T0 } // R-457: the SAME clock the age check reads
+	m.execFn = func(string, ...string) (string, error) { return "", nil }
+	return m, dir, g, c
+}
+
+func waitUpdateDone(t *testing.T, m *Manager, name string) *Stack {
+	t.Helper()
+	deadline := time.Now().Add(5 * time.Second)
+	for time.Now().Before(deadline) {
+		if st, ok := m.GetStack(name); ok && !st.Updating {
+			return st
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+	t.Fatal("the update never finished")
+	return nil
+}
+
+func pinOf(t *testing.T, dir string) string { return readPin(t, dir).PinnedImages["web"] }
+
+func fileBody(t *testing.T, p string) string {
+	t.Helper()
+	b, err := os.ReadFile(p)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return string(b)
+}
+
+func journalExists(m *Manager) bool {
+	_, err := os.Stat(m.updateJournalPath())
+	return err == nil
+}
+
+// ── A: the happy path, and the truth ─────────────────────────────────────────────────────────────
+
+// TestSlice4_A_SuccessIsDeclaredOnlyAfterHealth. The health wait BLOCKS until the test releases it;
+// while it blocks, the app must read Updating=true / phase=verifying / no error — and the safety dump
+// must have run while the pin still named the OLD version.
+//
+// COMPANION RED-PROOF 1 (REPORT.md): delete the updateHealth call from verifyAndConclude so success is
+// declared on the compose exit code. This test then fails at "Updating went false before health was
+// known" — which is R-443 exactly.
+func TestSlice4_A_SuccessIsDeclaredOnlyAfterHealth(t *testing.T) {
+	m, dir, g, c := newSlice4Manager(t)
+	release := make(chan struct{})
+	healthCalled := make(chan struct{}, 1)
+	m.updateHealthFn = func(ctx context.Context, name string, timeout time.Duration) (bool, string) {
+		healthCalled <- struct{}{}
+		<-release
+		return true, "fake healthy"
+	}
+	if err := m.StartGuardedUpdate("nextcloud"); err != nil {
+		t.Fatalf("a fully-qualified update must start: %v", err)
+	}
+	select {
+	case <-healthCalled:
+	case <-time.After(5 * time.Second):
+		st, _ := m.GetStack("nextcloud")
+		t.Fatalf("the health wait was never reached; state: updating=%v phase=%s err=%q", st.Updating, st.UpdatePhase, st.UpdateError)
+	}
+	st, _ := m.GetStack("nextcloud")
+	if !st.Updating {
+		t.Fatal("Updating went false before health was known — success reported on the compose exit code (R-443)")
+	}
+	if st.UpdatePhase != UpdatePhaseVerifying || st.UpdatePhaseLabel != "Működés ellenőrzése…" {
+		t.Errorf("while waiting for health the phase must be verifying, got %q / %q", st.UpdatePhase, st.UpdatePhaseLabel)
+	}
+	if st.UpdateError != "" {
+		t.Errorf("no error may be shown while verifying, got %q", st.UpdateError)
+	}
+	if !journalExists(m) {
+		t.Error("the journal must exist while the update is in flight (Scenario G depends on it)")
+	}
+	if got := pinOf(t, dir); got != "nextcloud:34.0.1-apache" {
+		t.Errorf("by verifying, the pin must have advanced; got %q", got)
+	}
+	close(release)
+	st = waitUpdateDone(t, m, "nextcloud")
+	if st.UpdatePhase != UpdatePhaseDone || st.UpdateError != "" || st.UpdatePhaseLabel != "Frissítve" {
+		t.Fatalf("after health the update is done: phase=%q label=%q err=%q", st.UpdatePhase, st.UpdatePhaseLabel, st.UpdateError)
+	}
+	if journalExists(m) {
+		t.Error("a completed update must clear its journal entry")
+	}
+	if _, err := os.Stat(filepath.Join(dir, preUpdateComposeFile)); err == nil {
+		t.Error("the pre-update copy must be removed after success")
+	}
+	if g.pinAtDump != "nextcloud:31.0.14-apache" {
+		t.Errorf("the safety dump must run BEFORE the pin moves (\"a minute ago\"); the pin at dump time was %q", g.pinAtDump)
+	}
+	if got, want := strings.Join(c.list(), " | "), "pull | up -d --remove-orphans"; got != want {
+		t.Errorf("compose calls = %q, want %q", got, want)
+	}
+	// RestorePoint twice by design: once in the preflight (the refusal), once inside the job (the
+	// precondition must still hold when the job actually starts).
+	if got := strings.Join(g.callList(), ","); got != "RestorePoint,RestorePoint,SafetyDump" {
+		t.Errorf("a fresh copy needs no backup-first; guard calls = %s", got)
+	}
+}
+
+// ── B: the proven copy is too old ─────────────────────────────────────────────────────────────────
+
+func TestSlice4_B_StaleCopyIsRefreshedFirst(t *testing.T) {
+	m, dir, g, _ := newSlice4Manager(t)
+	g.rp.ProvenAt = slice4T0.Add(-30 * time.Hour) // > 24 h default
+	g.rpAfterBackup = &UpdateRestorePoint{Restorable: true, Proven: true, ProvenAt: slice4T0.Add(-1 * time.Minute)}
+	if err := m.StartGuardedUpdate("nextcloud"); err != nil {
+		t.Fatal(err)
+	}
+	st := waitUpdateDone(t, m, "nextcloud")
+	if st.UpdatePhase != UpdatePhaseDone {
+		t.Fatalf("with a successful backup-first the update completes, got phase=%q err=%q", st.UpdatePhase, st.UpdateError)
+	}
+	calls := strings.Join(g.callList(), ",")
+	if !strings.HasPrefix(calls, "RestorePoint,RestorePoint,BackupNow,RestorePoint,SafetyDump") {
+		t.Errorf("a stale copy must be backed up FIRST and the precondition re-read; calls = %s", calls)
+	}
+	if got := pinOf(t, dir); got != "nextcloud:34.0.1-apache" {
+		t.Errorf("pin = %q", got)
+	}
+}
+
+func TestSlice4_B_BackupFailureMovesNothing(t *testing.T) {
+	m, dir, g, c := newSlice4Manager(t)
+	g.rp.ProvenAt = slice4T0.Add(-30 * time.Hour)
+	g.backupErr = errors.New("disk full")
+	if err := m.StartGuardedUpdate("nextcloud"); err != nil {
+		t.Fatal(err)
+	}
+	st := waitUpdateDone(t, m, "nextcloud")
+	if want := fmt.Sprintf(MsgUpdateBackupFailFmt, g.backupErr); st.UpdateError != want {
+		t.Errorf("UpdateError = %q, want the backup's own error in the sentence %q", st.UpdateError, want)
+	}
+	if got := pinOf(t, dir); got != "nextcloud:31.0.14-apache" {
+		t.Errorf("a failed backup must move nothing; pin = %q", got)
+	}
+	if len(c.list()) != 0 {
+		t.Errorf("a failed backup must reach no compose call; got %v", c.list())
+	}
+	if journalExists(m) {
+		t.Error("the journal must be cleared on a refusal")
+	}
+}
+
+func TestSlice4_B_BackupThatYieldsNoFreshUnitRefuses(t *testing.T) {
+	m, dir, g, c := newSlice4Manager(t)
+	g.rp.ProvenAt = slice4T0.Add(-30 * time.Hour) // stays stale: rpAfterBackup nil
+	if err := m.StartGuardedUpdate("nextcloud"); err != nil {
+		t.Fatal(err)
+	}
+	st := waitUpdateDone(t, m, "nextcloud")
+	if st.UpdateError != MsgUpdateBackupNoUnit {
+		t.Errorf("UpdateError = %q", st.UpdateError)
+	}
+	if pinOf(t, dir) != "nextcloud:31.0.14-apache" || len(c.list()) != 0 {
+		t.Error("nothing may move when the backup did not produce a fresh restorable copy")
+	}
+}
+
+// ── C: no backup exists that could restore this app ───────────────────────────────────────────────
+
+// COMPANION RED-PROOF 2 (REPORT.md): make the precondition in UpdatePreflight proceed when
+// !rp.Restorable. This test then fails with the update started.
+func TestSlice4_C_NoRestorableCopyRefusesBeforeAnythingMoves(t *testing.T) {
+	m, dir, g, c := newSlice4Manager(t)
+	// A PROVEN, FRESH copy whose unit cannot be opened — the realistic half-copied mirror. Proven and
+	// fresh on purpose: a fixture that is also unproven would be refused by the proven check alone,
+	// and a red-proof that drops the restorable check would then pass inertly (observed on the first
+	// run of red-proof 2, 2026-09-13).
+	g.rp = UpdateRestorePoint{Restorable: false, Proven: true, ProvenAt: slice4T0.Add(-time.Hour)}
+	err := m.StartGuardedUpdate("nextcloud")
+	var ref *UpdateRefusal
+	if !errors.As(err, &ref) || ref.Reason != "no_backup" {
+		t.Fatalf("an app with no restorable copy must be REFUSED (no_backup), got %v", err)
+	}
+	if want := fmt.Sprintf(MsgUpdateNoBackupFmt, "nextcloud"); ref.Message != want {
+		t.Errorf("message = %q", ref.Message)
+	}
+	time.Sleep(50 * time.Millisecond)
+	if st, _ := m.GetStack("nextcloud"); st.Updating {
+		t.Error("a refused update must not set Updating")
+	}
+	if pinOf(t, dir) != "nextcloud:31.0.14-apache" || len(c.list()) != 0 {
+		t.Error("a refused update must move nothing")
+	}
+	// A copy that exists but was never PROVEN is not a copy (R-101).
+	g.rp = UpdateRestorePoint{Restorable: true, Proven: false}
+	if ref := m.UpdatePreflight("nextcloud"); ref == nil || ref.Reason != "no_backup" {
+		t.Errorf("an unproven copy must refuse too, got %v", ref)
+	}
+}
+
+// ── D: the cheap refusals, each one ─────────────────────────────────────────────────────────────────
+
+func TestSlice4_D_CheapRefusals(t *testing.T) {
+	cases := []struct {
+		name   string
+		setup  func(m *Manager, g *fakeGuards, dir string)
+		reason string
+		msg    string
+	}{
+		{"held", func(m *Manager, g *fakeGuards, _ string) { g.held, g.holdWhy = true, "THE HOLD TEXT" }, "held", "THE HOLD TEXT"},
+		{"busy", func(m *Manager, g *fakeGuards, _ string) { g.busy = true }, "busy", MsgUpdateBusy},
+		{"already updating", func(m *Manager, _ *fakeGuards, _ string) { m.stacks["nextcloud"].Updating = true }, "updating", fmt.Sprintf(MsgUpdateAlreadyFmt, "nextcloud")},
+		{"deploying", func(m *Manager, _ *fakeGuards, _ string) { m.stacks["nextcloud"].Deploying = true }, "deploying", fmt.Sprintf(MsgUpdateDeployingFmt, "nextcloud")},
+		{"memory", func(m *Manager, _ *fakeGuards, _ string) {
+			catDir := filepath.Join(m.cfg.Paths.DataDir, "catalog-cache", "templates", "nextcloud")
+			if err := os.WriteFile(filepath.Join(catDir, ".felhom.yml"), []byte("resources:\n  mem_request: 900M\n"), 0o644); err != nil {
+				panic(err)
+			}
+			m.updateMemoryFn = func(newReq, _, _, _ int) (string, string) {
+				return fmt.Sprintf("Nincs elég memória (%d MB)", newReq), ""
+			}
+		}, "memory", "Nincs elég memória (900 MB)"},
+		{"disk", func(m *Manager, _ *fakeGuards, _ string) {
+			m.updateDiskFreeFn = func() (float64, bool) { return 1.0, true }
+		}, "disk", fmt.Sprintf(MsgUpdateDiskFmt, 1.0, updateDiskFloorGiB)},
+		{"guards unwired", func(m *Manager, _ *fakeGuards, _ string) { m.updateGuards = nil }, "guards_unwired", MsgUpdateNoGuards},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			m, dir, g, c := newSlice4Manager(t)
+			tc.setup(m, g, dir)
+			ref := m.UpdatePreflight("nextcloud")
+			if ref == nil || ref.Reason != tc.reason {
+				t.Fatalf("want refusal %q, got %+v", tc.reason, ref)
+			}
+			if ref.Message != tc.msg {
+				t.Errorf("message = %q, want %q", ref.Message, tc.msg)
+			}
+			if err := m.StartGuardedUpdate("nextcloud"); err == nil {
+				t.Fatal("StartGuardedUpdate must refuse the same")
+			}
+			time.Sleep(20 * time.Millisecond)
+			if len(c.list()) != 0 || pinOf(t, dir) != "nextcloud:31.0.14-apache" {
+				t.Errorf("a cheap refusal reached the act: compose=%v pin=%s", c.list(), pinOf(t, dir))
+			}
+		})
+	}
+}
+
+// ── E: the pull fails ────────────────────────────────────────────────────────────────────────────
+
+func TestSlice4_E_PullFailurePutsThePinBack(t *testing.T) {
+	m, dir, g, c := newSlice4Manager(t)
+	c.fail["pull"] = errors.New("exit code 1\nstderr: manifest unknown")
+	if err := m.StartGuardedUpdate("nextcloud"); err != nil {
+		t.Fatal(err)
+	}
+	st := waitUpdateDone(t, m, "nextcloud")
+	if st.UpdateError != MsgUpdatePullFailed {
+		t.Errorf("UpdateError = %q, want the Hungarian sentence and never raw stderr", st.UpdateError)
+	}
+	if strings.Contains(st.UpdateError, "manifest unknown") {
+		t.Error("raw Docker stderr leaked into the customer sentence")
+	}
+	if got := pinOf(t, dir); got != "nextcloud:31.0.14-apache" {
+		t.Errorf("the pin must be PUT BACK after a failed pull, got %q", got)
+	}
+	if got := fileBody(t, filepath.Join(dir, "docker-compose.yml")); got != pinTplOld {
+		t.Errorf("the live file must be re-rendered to the old version:\n%s", got)
+	}
+	if got := fileBody(t, AppliedComposePath(dir)); got != pinTplOld {
+		t.Errorf("the stored definition must be the old one again:\n%s", got)
+	}
+	if got := strings.Join(c.list(), " | "); got != "pull" {
+		t.Errorf("after a failed pull nothing else runs; compose calls = %q", got)
+	}
+	for _, call := range g.callList() {
+		if call == "HoldAfterFailedUpdate" {
+			t.Error("a failed PULL ran nothing and must not hold the app")
+		}
+	}
+}
+
+// ── F: the new version does not come up ─────────────────────────────────────────────────────────
+
+// COMPANION RED-PROOF 3 (REPORT.md): remove the HoldAfterFailedUpdate call from failAndHold. This test
+// then fails: no hold, and the customer is not told the route back.
+func TestSlice4_F_HealthFailureHoldsTheAppAndKeepsTheNewPin(t *testing.T) {
+	m, dir, g, c := newSlice4Manager(t)
+	m.updateHealthFn = func(context.Context, string, time.Duration) (bool, string) { return false, "crash loop" }
+	if err := m.StartGuardedUpdate("nextcloud"); err != nil {
+		t.Fatal(err)
+	}
+	st := waitUpdateDone(t, m, "nextcloud")
+	if st.UpdatePhase != UpdatePhaseFailed {
+		t.Errorf("phase = %q", st.UpdatePhase)
+	}
+	held, _ := g.HoldFor("nextcloud")
+	if !held {
+		t.Fatal("an app that did not come up must be HELD")
+	}
+	if !g.holdProvenAt.Equal(g.rp.ProvenAt) {
+		t.Errorf("the hold must name the PROVEN copy date %s, got %s", g.rp.ProvenAt, g.holdProvenAt)
+	}
+	if st.UpdateError != "HELD-SENTENCE" {
+		t.Errorf("the page must carry the hold's own sentence, got %q", st.UpdateError)
+	}
+	if st.HoldReason != "HELD-SENTENCE" {
+		t.Errorf("GetStack must carry the hold text, got %q", st.HoldReason)
+	}
+	if got := pinOf(t, dir); got != "nextcloud:34.0.1-apache" {
+		t.Errorf("the pin must STAY on the new version (its migration may have run), got %q", got)
+	}
+	if got := strings.Join(c.list(), " | "); got != "pull | up -d --remove-orphans | down" {
+		t.Errorf("the failed app must be stopped; compose calls = %q", got)
+	}
+	if journalExists(m) {
+		t.Error("the journal is cleared once the hold (the durable record) is written")
+	}
+}
+
+func TestSlice4_F_UnsavedHoldSaysSo(t *testing.T) {
+	m, _, g, _ := newSlice4Manager(t)
+	m.updateHealthFn = func(context.Context, string, time.Duration) (bool, string) { return false, "crash loop" }
+	g.holdErr = errors.New("settings.json read-only")
+	if err := m.StartGuardedUpdate("nextcloud"); err != nil {
+		t.Fatal(err)
+	}
+	if st := waitUpdateDone(t, m, "nextcloud"); st.UpdateError != MsgUpdateHoldUnsaved {
+		t.Errorf("an unrecorded hold must be disclosed, got %q", st.UpdateError)
+	}
+}
+
+// ── G: the controller restarts mid-update ────────────────────────────────────────────────────────
+
+func writeTestJournal(t *testing.T, m *Manager, name string, e updateJournalEntry) {
+	t.Helper()
+	if err := m.writeUpdateJournal(updateJournal{Updates: map[string]updateJournalEntry{name: e}}); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// simulateAdvanced puts the stack in the state a crash AFTER the pin moved would leave: pin, live file
+// and stored definition all new, the pre-update copy on disk.
+func simulateAdvanced(t *testing.T, m *Manager, dir string) updateJournalEntry {
+	t.Helper()
+	mustWrite(t, filepath.Join(dir, preUpdateComposeFile), pinTplOld)
+	mustWrite(t, filepath.Join(dir, preUpdateAppliedFile), pinTplOld)
+	if err := m.advancePinToCatalog("nextcloud", dir); err != nil {
+		t.Fatal(err)
+	}
+	return updateJournalEntry{
+		StartedAt: slice4T0, PrevPin: map[string]string{"web": "nextcloud:31.0.14-apache"},
+		PrevCompose: filepath.Join(dir, preUpdateComposeFile), PrevApplied: filepath.Join(dir, preUpdateAppliedFile),
+		ProvenCopyAt: slice4T0.Add(-time.Hour).Format(time.RFC3339),
+	}
+}
+
+func TestSlice4_G_InterruptedBeforeUpIsPutBack(t *testing.T) {
+	m, dir, _, c := newSlice4Manager(t)
+	e := simulateAdvanced(t, m, dir)
+	e.Phase = UpdatePhasePulling
+	writeTestJournal(t, m, "nextcloud", e)
+
+	if resumed := m.RecoverUpdates(); len(resumed) != 0 {
+		t.Fatalf("an update interrupted before `up` is not resumed, got %v", resumed)
+	}
+	if got := pinOf(t, dir); got != "nextcloud:31.0.14-apache" {
+		t.Errorf("the pin must be put back, got %q", got)
+	}
+	if got := fileBody(t, filepath.Join(dir, "docker-compose.yml")); got != pinTplOld {
+		t.Errorf("the live file must be the old definition again:\n%s", got)
+	}
+	st, _ := m.GetStack("nextcloud")
+	if st.Updating || st.UpdateError != MsgUpdateInterrupted {
+		t.Errorf("updating=%v err=%q", st.Updating, st.UpdateError)
+	}
+	if journalExists(m) || len(c.list()) != 0 {
+		t.Error("recovery of a pre-up interruption runs nothing and clears the journal")
+	}
+}
+
+func TestSlice4_G_InterruptedBeforeThePinIsDroppedUntouched(t *testing.T) {
+	m, dir, _, _ := newSlice4Manager(t)
+	writeTestJournal(t, m, "nextcloud", updateJournalEntry{Phase: UpdatePhaseSafetyDump, StartedAt: slice4T0})
+	m.RecoverUpdates()
+	if pinOf(t, dir) != "nextcloud:31.0.14-apache" || journalExists(m) {
+		t.Error("an update interrupted before the pin moved must leave the pin and clear the journal")
+	}
+}
+
+func TestSlice4_G_InterruptedAfterUpResumesTheHealthWait(t *testing.T) {
+	m, dir, g, c := newSlice4Manager(t)
+	e := simulateAdvanced(t, m, dir)
+	e.Phase = UpdatePhaseVerifying
+	writeTestJournal(t, m, "nextcloud", e)
+	guards := m.updateGuards
+	m.updateGuards = nil // at RecoverUpdates time the backup side is NOT wired yet (main.go order)
+
+	resumed := m.RecoverUpdates()
+	if len(resumed) != 1 || !m.IsUpdating("nextcloud") || !m.UpdatingStacks()["nextcloud"] {
+		t.Fatalf("an update interrupted after `up` must be marked Updating for the boot sweep; resumed=%v", resumed)
+	}
+	if got := pinOf(t, dir); got != "nextcloud:34.0.1-apache" {
+		t.Errorf("after `up` the pin is NOT put back — something may have run; got %q", got)
+	}
+
+	m.updateGuards = guards
+	m.updateHealthFn = func(context.Context, string, time.Duration) (bool, string) { return false, "still broken" }
+	if n := m.ResumeInterruptedUpdates(context.Background()); n != 1 {
+		t.Fatalf("resumed %d, want 1", n)
+	}
+	st := waitUpdateDone(t, m, "nextcloud")
+	if held, _ := g.HoldFor("nextcloud"); !held || st.UpdatePhase != UpdatePhaseFailed {
+		t.Errorf("a resumed update that is still unhealthy must end HELD; held=%v phase=%q", held, st.UpdatePhase)
+	}
+	if got := strings.Join(c.list(), " | "); got != "up -d --remove-orphans | down" {
+		t.Errorf("resumption re-runs `up` then stops the failed app; compose calls = %q", got)
+	}
+	if !g.holdProvenAt.Equal(slice4T0.Add(-time.Hour)) {
+		t.Errorf("the resumed hold must name the journaled proven copy date, got %s", g.holdProvenAt)
+	}
+}
+
+// ── the page reads the hold from the ONE store ──────────────────────────────────────────────────
+
+func TestSlice4_GetStacksCarriesTheHoldText(t *testing.T) {
+	m, _, g, _ := newSlice4Manager(t)
+	g.held, g.holdWhy = true, "HOLD"
+	for _, st := range m.GetStacks() {
+		if st.Name == "nextcloud" && st.HoldReason != "HOLD" {
+			t.Errorf("GetStacks HoldReason = %q", st.HoldReason)
+		}
+	}
+	g.held = false
+	if st, _ := m.GetStack("nextcloud"); st.HoldReason != "" {
+		t.Errorf("a lifted hold must disappear on the next read, got %q", st.HoldReason)
+	}
+}
+
+func TestSlice4_PhaseLabelsAreTheSpecifiedCopy(t *testing.T) {
+	want := map[string]string{
+		UpdatePhaseChecking:   "Ellenőrzés…",
+		UpdatePhaseBackingUp:  "Biztonsági mentés készül a frissítés előtt…",
+		UpdatePhaseSafetyDump: "Adatbázis pillanatkép…",
+		UpdatePhasePulling:    "Új verzió letöltése…",
+		UpdatePhaseStarting:   "Indítás az új verzióval…",
+		UpdatePhaseVerifying:  "Működés ellenőrzése…",
+		UpdatePhaseDone:       "Frissítve",
+	}
+	for p, l := range want {
+		if got := UpdatePhaseLabel(p); got != l {
+			t.Errorf("label(%s) = %q, want %q", p, got, l)
+		}
+	}
+}

controller/internal/web/handlers.go

@@ -1429,12 +1429,15 @@ func (s *Server) buildAppBackupRows(status *backup.FullBackupStatus) []AppBackup
 				// a disconnected destination, a pre-v2 layout — so an offer is never rendered for a
 				// copy the action would refuse. On any refusal the action is simply not offered; the
 				// row keeps rendering everything else it already showed.
-				if cov, covErr := s.backupMgr.Tier2RestoreCoverage(app.StackName); covErr == nil {
-					row.Tier2UnitRestorable = cov.CanRestoreUnit()
+				//
+				// Slice 4: the computation lives in backup.Tier2UnitRestorePoint, because the guarded
+				// update asks the same question and a second copy of a predicate drifts (R-203).
+				if rp, rpErr := s.backupMgr.Tier2UnitRestorePoint(app.StackName); rpErr == nil {
+					row.Tier2UnitRestorable = rp.Restorable
 					// R-403: the UNIT action names the PACKAGE's date, not the run's. After a
 					// preserved leg those are different dates and the run's is the flattering one.
-					pkgDate, stale := cov.UnitRestoreDate()
-					row.Tier2CopyDate, row.Tier2CopyDateProven = pkgDate, cov.CopyLastSuccess != ""
+					pkgDate, stale := rp.CopyDate, rp.PackagePreserved
+					row.Tier2CopyDate, row.Tier2CopyDateProven = pkgDate, rp.CopyDateProven
 					row.Tier2UnitConfirm = tier2UnitConfirmWithStaleness(pkgDate, row.Tier2CopyDateProven, stale)
 					if stale && pkgDate != "" {
 						row.Tier2UnitStaleNotice = fmt.Sprintf(tier2UnitStaleNoticeFmt, fmtRFC3339Local(pkgDate))

controller/internal/web/intermediary.go

@@ -219,6 +219,12 @@ func (s *Server) stopAppsOnPath(storagePath string) []string {
 // restartStacks starts each named stack (the gate-stopped set on drive return). Best-effort per app.
 func (s *Server) restartStacks(names []string) {
 	for _, name := range names {
+		// Slice 4 / R-379: the drive-return gate starts apps UNATTENDED, so it must honour a hold
+		// exactly as the customer's button does. Until v0.237.0 it did not — a held app whose drive
+		// blinked would have been started again. "A hold that only one path honours is not a hold."
+		if s.appHeld(name) {
+			continue
+		}
 		if err := s.stackMgr.StartStack(name); err != nil {
 			s.logger.Printf("[WARN] [gate] restart %s: %v", name, err)
 		}
@@ -454,6 +460,9 @@ func (s *Server) processGuestBootChange() {
 	}
 	recreate := func(bs bootStack) {
 		s.logger.Printf("[INFO] [gate] boot %s: live bind confirmed — recreating drive-backed app %s (state=%s) onto %s", resp.GuestBootID, bs.name, bs.state, bs.hdd)
+		if s.appHeld(bs.name) {
+			return
+		}
 		_ = s.stackMgr.StopStack(bs.name)
 		if serr := s.stackMgr.StartStack(bs.name); serr != nil {
 			s.logger.Printf("[WARN] [gate] boot recreate %s: %v", bs.name, serr)
@@ -694,3 +703,16 @@ func (s *Server) notifyDriveReturned(path string, isTarget map[string]bool) {
 	}
 	s.notifier.NotifyStorageReconnected(label)
 }
+
+// appHeld reports whether an app carries a hold (failed update or failed restore) and logs the skip.
+// Nil-safe: no backup manager means no hold store, so nothing is held.
+func (s *Server) appHeld(name string) bool {
+	if s.backupMgr == nil {
+		return false
+	}
+	held, _ := s.backupMgr.RestoreHoldFor(name)
+	if held {
+		s.logger.Printf("[WARN] [gate] NOT starting %s — the app is HELD (a failed update or restore); a person releases it", name)
+	}
+	return held
+}

controller/internal/web/slice4_update_test.go

@@ -0,0 +1,74 @@
+package web
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-controller/internal/backup"
+	"gitea.dooplex.hu/admin/felhom-controller/internal/settings"
+)
+
+// Slice 4 — the predicate moved from this package into backup.Tier2UnitRestorePoint. The page must
+// render IDENTICALLY: the row's four unit-restore fields are compared against the coverage computed
+// the old way (the inline expression that used to live in buildAppBackupRows).
+func TestSlice4_BackupRowUnitFieldsAreUnchangedByTheExtraction(t *testing.T) {
+	s, sett, m := newOffboxWebServer(t)
+	dest := t.TempDir()
+	if err := sett.AddStoragePath(settings.StoragePath{Path: dest, Label: "flash"}); err != nil {
+		t.Fatal(err)
+	}
+	if err := sett.SetCrossDriveConfig("app", &settings.CrossDriveBackup{
+		Enabled: true, Method: "rsync", DestinationPath: dest,
+		LastRun: "2026-09-13T01:30:00Z", LastStatus: "ok", LastSuccess: "2026-09-13T01:30:00Z", SuccessTracked: true,
+	}); err != nil {
+		t.Fatal(err)
+	}
+	unit := filepath.Join(dest, "backups", "secondary", "app", "recovery-unit")
+	if err := os.MkdirAll(unit, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(unit, "manifest.json"), []byte(`{"schema_version":2,"app_name":"app","created_at":"2026-09-12T02:15:29Z"}`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dest, "backups", "secondary", "app", ".felhom-tier2-layout"), []byte("2"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	cov, err := m.Tier2RestoreCoverage("app")
+	if err != nil {
+		t.Fatal(err)
+	}
+	wantDate, wantStale := cov.UnitRestoreDate()
+	rows := s.buildAppBackupRows(&backup.FullBackupStatus{AppDataInfo: []backup.AppBackupInfo{{StackName: "app", DisplayName: "App"}}})
+	row := findRow(rows, "app")
+	if row == nil {
+		t.Fatal("no row")
+	}
+	if !cov.CanRestoreUnit() {
+		t.Fatal("fixture: the copy must hold an openable unit")
+	}
+	if row.Tier2UnitRestorable != cov.CanRestoreUnit() || row.Tier2CopyDate != wantDate ||
+		row.Tier2CopyDateProven != (cov.CopyLastSuccess != "") ||
+		row.Tier2UnitConfirm != tier2UnitConfirmWithStaleness(wantDate, cov.CopyLastSuccess != "", wantStale) {
+		t.Errorf("the row changed: restorable=%v date=%q proven=%v confirm=%q", row.Tier2UnitRestorable, row.Tier2CopyDate, row.Tier2CopyDateProven, row.Tier2UnitConfirm)
+	}
+}
+
+// The drive-return gate starts apps UNATTENDED. A held app must be skipped.
+//
+// COMPANION RED-PROOF (REPORT.md): delete the appHeld check from restartStacks. The server has NO stack
+// manager here on purpose, so the unguarded StartStack call panics and this test fails.
+func TestSlice4_DriveReturnGateSkipsAHeldApp(t *testing.T) {
+	s, _, m := newOffboxWebServer(t)
+	if err := m.HoldAfterFailedUpdate("held", time.Now(), time.Now()); err != nil {
+		t.Fatal(err)
+	}
+	defer func() {
+		if r := recover(); r != nil {
+			t.Fatalf("the drive-return gate tried to START a held app: %v", r)
+		}
+	}()
+	s.restartStacks([]string{"held"})
+}