diff --git a/CHANGELOG.md b/CHANGELOG.md
index 8ee6fa1..016d265 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,3 +1,32 @@
+## v0.294.0 — the off-site clean-up deletes honest old copies; a first install survives the image clean-up; failed compose logs its reason; the move-aside line names its destination (R-867, R-863, R-864, R-869) (2026-10-05)
+
+**MinAgent: 0.131.0** (unchanged). No new household string.
+
+- **R-867.** The fake-snapshot guard's "young" line is now built from the SAME constants as the retention policy
+  (`keepDaily` 7 / `keepWeekly` 4 / `keepMonthly` 6 → `retentionPolicy`): a removal is refused only when its CALENDAR
+  day is fewer than `keepDaily` days before today (in the snapshot's own zone, as restic buckets it) and no newer
+  same-day snapshot of its group supersedes it. v0.289–0.293 used a fixed 8-day age, which sat inside the keep window:
+  the snapshot a keep-7-dailies policy drops each night is 7 days + seconds old, so every window on a box with more than
+  7 nightly snapshots refused, deleted nothing and mailed `offsite_prune_guard_refused` (measured demo-felhom
+  2026-10-05 02:15 UTC). Every other refusal is unchanged: future-dated, newer than the hub allows, a plan above the
+  week's cap (R-833's grant still raises it), the lab's 13-fake shape. Tests run restic 0.14.0's policy itself
+  (`restic0140Plan`, proven identical to the 0.14.0 binary over 92 snapshots): the measured shape, 60 nights × 3 apps
+  with two same-day manual runs across a month boundary, weekly windows, and a skew-window fake that the line still
+  refuses. Recorded limit: past-dated gap-fills that steer OLDER keeps stay R-822's residual (bounded by the cap and
+  the hub's count check), as before.
+- **R-863.** The image clean-up no longer runs while ANY compose command that can pull is running (`up`, `pull`,
+  `create`, `run` — install, update, restore, undo, the FileBrowser sync). New `dockerexec.BeginImageWork` /
+  `TryImageCleanup`: pulling commands hold a shared lock for their whole run; a clean-up pass takes it exclusively
+  without waiting and, when it cannot, does not run. The one-time clean-up writes its marker ONLY after a pass that ran
+  and is retried every 2 minutes (at most 30 times) instead of being skipped until the next start. Measured cause: on
+  the fresh Tester 1 box the one-time pass deleted `mariadb@sha256:…` (untagged, named by no container yet) one second
+  before `compose up` created BookStack's database container.
+- **R-864.** A failed compose command logs and returns the TAIL of stderr (new `tailStr`, rune-safe) — compose prints
+  pull progress first and the reason last; the head kept only "Image … Pulling".
+- **R-869.** The decision-78 move-aside log line now prints the destination the hub returned.
+- Tests: `TestOffsiteGuard_RealPolicy*`, `TestRetentionPolicy_BuiltFromTheGuardConstants`, `TestR863_*`, `TestR864_*`,
+  `TestR869_*`. Red-proofs (9): `felhom.eu/documentation/audits/night-fixes-2026-10-05/part{A,B,D}/`.
+
 ## v0.293.0 — the controller heals itself after the guest's Docker socket is re-created (R-860, generalises R-858) (2026-10-04)
 
 **MinAgent: 0.131.0** (unchanged). No new household string.
diff --git a/REUSE.md b/REUSE.md
index 807b7f5..592a512 100644
--- a/REUSE.md
+++ b/REUSE.md
@@ -30,7 +30,8 @@
 | `stacks.OpenSignupWindow` / `SignupBlocked` / `SetupGateProbe` · `web.ServeSignupClosed` (v0.281.0, decision 47) | controller/internal/stacks/signup_block.go · controller/internal/web/setup_gate.go | `(name)` | Sign-up closed at the app's own address once the gate opens; the household's 15-minute window; the press asks the probe | **The block goes up BEFORE the gate comes down** (a failed write keeps the gate closed). Never on an app this box did not gate |
 | `family.Store` · `stacks.FamilyGateHost` / `familyGateTick` / `FamilyExceptRegexp` · `web.ServeFamilyGateAuth` / `ServeFamilyStart` / `ServeFamilyLogin` / `ServeFamilyLogout` (v0.287.0, decisions 63/64) | controller/internal/family/family.go · controller/internal/stacks/family_gate.go · controller/internal/web/family_gate.go | store `Add/Reset/Remove/Verify/NewSession/Valid/EndSession` · `(host)` / `()` / `(prefix)` · handlers | The PERMANENT family gate: family members with their own logins in front of a `family_gate: true` app | **A family cookie never opens the dashboard** (RequireAuth reads only `felhom_session`); the app cookie names a STORE session, so the store is asked on every request (reset/remove/logout end access at once); **every exception goes through `FamilyExceptRegexp`** (anchored — never a hand-written PathPrefix); door written before the first start, like the setup gate. `familyStoreOverride` is the test seam. Fifth atomic-write helper (family.json, fsync) — see §6 |
 | `stacks.OpenInstallHold` / `installHoldTick` (v0.284.0, R-741) | controller/internal/stacks/install_hold.go | `(name, by)` / `()` | An `after_install` app held behind the setup gate's door until its known login is replaced | **Written before the first start**, like the gate; opens on `after_install` success or the household's "I changed it"; the door (`SetupGateHost`) reads holds first |
-| `stacks.RetainImagesAfterUpdate` / `RetainImagesAfterRemove` / `RunImageRetentionOnce` · seam `imageDocker` (v0.284.0, decision 53) | controller/internal/stacks/image_retention.go | `(name, previous)` / `(name, repos)` / `()` | Deletes an app's images older than its running + previous one | **The keep set is box-wide and read at delete time** (containers, installed composes, installed/previous records); exact id, never forced or pruned; skipped while any update runs; tests use the `imageDocker` seam, never Docker |
+| `stacks.RetainImagesAfterUpdate` / `RetainImagesAfterRemove` / `RunImageRetentionOnce` (→ `(deleted, done)` since v0.294.0) / `RunImageRetentionOnceUntilDone` · seam `imageDocker` (v0.284.0, decision 53) | controller/internal/stacks/image_retention.go | `(name, previous)` / `(name, repos)` / `()` | Deletes an app's images older than its running + previous one | **The keep set is box-wide and read at delete time** (containers, installed composes, installed/previous records); exact id, never forced or pruned; skipped while any update runs AND while any image-pulling compose command runs (R-863, `dockerexec.TryImageCleanup`); the one-time pass writes its marker only after a pass that RAN; tests use the `imageDocker` seam, never Docker |
+| `dockerexec.BeginImageWork(args)` / `TryImageCleanup()` / `ImagePulling(args)` (v0.294.0, R-863) | controller/internal/dockerexec/imagework.go | `defer dockerexec.BeginImageWork(args)()` | **Every NEW place that runs `compose up/pull/create/run` must hold it** (the stacks compose helpers, appexport's restore and the FileBrowser sync already do) | An image pulled by compose is named by no container until compose creates one; a clean-up pass in that window deleted it (measured, R-863). The pass takes the lock exclusively without waiting; pulling commands wait for a running pass (seconds) |
 | `stacks.RetainControllerImages(ControllerImageRecord)` · `selfupdate.UpdateState.RecordedPrevious` (v0.285.0, decision 56) | controller/internal/stacks/controller_image_retention.go | `({Repo, Running, Previous})` | Deletes controller images older than the running + previous one | The previous comes from the SWAP RECORD (success onto the running version), version order only as the fallback; versions above the running one and non-version tags are kept; skipped while the controller swaps itself; same `imageDocker` seam |
 | `stacks.CloseSignupNow` / `CloseSignupOffered` / `applyNativeLock` (v0.282.0, decisions 47/49) | controller/internal/stacks/after_setup.go | `(name)` | The app's own sign-up switch after the setup; "close sign-up now" for an app installed before the rule | **Check the installed compose reads the variable** (an old install carries the old compose until its next update) — never record a lock that is not there. One run per app at a time (`nativeLockBusy`) |
 | `backup.judgeCopy` / `HollowCopies` / `SetHollowCopyNotify` (Part D, v0.279.0) | controller/internal/backup/hollow_watch.go | `(app, tier, unitDir)` | A RUNNING app whose newest copy holds no data → operator digest once/day + page sentence | Uses `unitCarriesData` (the manifest, never size); a stopped held app is never flagged |
@@ -246,7 +247,7 @@
 | `metrics.FetchContainerLogTail` | controller/internal/metrics/logscanner.go | `(name, tailLines) (string, error)` | Raw per-container `docker logs --tail=N` | 15s timeout; caller caps/redacts (capTailLines) |
 | `ConfigRefresher.Reconcile` | controller/internal/report/config_refresh.go | `(ackVersion int)` | Pull-based config refresh | Re-pulls controller.yaml (re-merging local_api), then graceful self-restart; first-run = baseline, no restart |
 | `offsiteapply.HubRegistrar` / `HubWindowClient` / `PinnedProber` (v0.289.0, decisions 68–69) | controller/internal/offsiteapply/seams.go | `Register(ctx,pub)(fp,err)` · `Confirm` · `MoveAside` · `Open/Close` window · `Probe(ctx,host,user,port,kh,privPEM) bool` | EVERY off-site key install, the hub move-aside, the clean-up window | **The box never handles the sub-account password** — there is no consume path any more. `PinnedProber` is a POSITIVE observable (exit 0 + rclone output); "authenticates" is NOT enough — an unpinned key authenticates and can delete. |
-| `Manager.offsiteWindowRetention` + `offsiteGuard` (v0.289.0, R-822) | controller/internal/backup/offbox_window.go | `(ctx, base, env, why)` · pure `(all, plan, now, newestAllowed, max) (ids, refuse)` | THE retention step for both callers (after a run, over quota) | Pinned tier: no window → nothing deleted; the guard runs BEFORE any forget; a young snapshot superseded the same day is EXCLUDED, any other young removal / future date / plan above `max_remove` REFUSES; forget is by explicit ids, oldest first. A due abandonment goes to `OffsiteAbandonClient` (hub, 7-day wait). NAS tier: the old SP-2 policy, unchanged. |
+| `Manager.offsiteWindowRetention` + `offsiteGuard` (v0.289.0, R-822) | controller/internal/backup/offbox_window.go | `(ctx, base, env, why)` · pure `(all, plan, now, newestAllowed, max) (ids, refuse)` | THE retention step for both callers (after a run, over quota) | Pinned tier: no window → nothing deleted; the guard runs BEFORE any forget; "young" = fewer than `keepDaily` CALENDAR days old (v0.294.0, R-867 — the line is built from the same constants as `retentionPolicy`; never an hour count); a young snapshot superseded the same day is EXCLUDED, any other young removal / future date / plan above `max_remove` REFUSES; tests run restic 0.14.0's policy itself (`restic0140Plan`, proven identical to the binary); forget is by explicit ids, oldest first. A due abandonment goes to `OffsiteAbandonClient` (hub, 7-day wait). NAS tier: the old SP-2 policy, unchanged. |
 | `offsiteapply.SettleProvider` / `SettleFunc` / `Bridge.AwaitSettle` / `ReconcileWhenSettled` (R-71a, v0.162.0) | controller/internal/offsiteapply/offsiteapply.go + seams.go | `SettleState() (version, floor string, updateRunning, floorKnown bool)` | THE settle-gate: defers the offsite one-time-password consume past a managed day-0 floor-update (the F10 race). Wire the `SettleFunc` adapter over `updater.GetFloor()`/`IsUpdateRunning()` — **the updater's knowledge is the ONE floor source; never fetch the floor a second way**. Gate ONLY the bridge goroutine, and only when an updater exists (nil `Settle` = reconcile immediately). Bounds `settlePoll`/`settleFloorSubBound`/`settleOverallBound`; the floor is in-memory (report-ACK-derived, ~5–10 s), NOT persisted → unknown until the first ACK on any restart. Inject `Now`/`Sleep` in tests (no real sleeps). B′: at/above-floor GOes on the first poll, zero wait. Do NOT touch the consume/persist order or the 404 contract — ordering only |
 | `bootstrap.MaybeIngest` / `RefreshConfig` | controller/internal/bootstrap/bootstrap.go | bootstrap.json → controller.yaml | Day-0 + refresh | Overwrites controller.yaml, NEVER settings.json |
 | `api.GracefulSelfRestart` | controller/internal/api/selfrestart.go | `(logger)` | Controller self-restart | Detached exit; bootstrap unit re-runs the image |
@@ -396,7 +397,7 @@ Cross-repo edges:
 | Duplication | Locations |
 |---|---|
 | Atomic write ×4 (+inline in settings.save) | controller/internal/backup/recovery_unit.go `atomicWrite`; controller/internal/bootstrap/bootstrap.go `writeFileAtomic`; controller/internal/setup/handlers.go `atomicWriteFile`; controller/internal/api/router.go `writeConfig0600` |
-| String truncation ×3 | controller/internal/util/strings.go `TruncateStr` (rune-safe, 1 caller); controller/internal/stacks/manager.go `truncateStr` (byte-based, widely used); controller/internal/backup/offbox.go `truncate` |
+| String truncation ×4 | controller/internal/util/strings.go `TruncateStr` (rune-safe, 1 caller); controller/internal/stacks/manager.go `truncateStr` (byte-based HEAD, widely used) and `tailStr` (rune-safe TAIL, v0.294.0 R-864 — use it for a command's stderr: the error is printed last); controller/internal/backup/offbox.go `truncate` |
 | humanizeBytes ×3 | controller/internal/appbackup/appdata.go `HumanizeBytes` (canonical) + private twin; controller/internal/appexport/estimate.go `humanizeBytes`; controller/internal/backup/appbackup_bridge.go wrapper (deliberate bridge) |
 | copyFile ×2 | controller/internal/stacks/migrate.go (returns bytes) vs controller/internal/appexport/export.go |
 | dir-size ×6 | controller/internal/stacks/delete.go `getDirSizeBytes`/`getDirSizeHuman`; controller/internal/backup/tier2.go `dirSizeBytes` (du -sb); controller/internal/appexport/estimate.go `dirSize`+`duBytes`; controller/internal/appexport/export.go `calcDirSize`; controller/internal/web/handlers.go `dirSizeHuman` |
diff --git a/controller/cmd/controller/main.go b/controller/cmd/controller/main.go
index 67ed5cf..280468c 100644
--- a/controller/cmd/controller/main.go
+++ b/controller/cmd/controller/main.go
@@ -1731,7 +1731,9 @@ func main() {
 			return
 		case <-time.After(3 * time.Minute):
 		}
-		stackMgr.RunImageRetentionOnce()
+		// R-863: retried every 2 minutes while an install/update/restore is pulling (a new household's first
+		// installs land exactly here), instead of deleting an image compose has pulled but not yet used.
+		go stackMgr.RunImageRetentionOnceUntilDone(ctx, 2*time.Minute, 30)
 		// decision 56 (R-745): the controller's own images — keep the running one and the one before it. At every
 		// start, 3 minutes in: a swap has ended by then (the agent's verify window is 90 s), so the image the agent
 		// would roll back to is the running one, and the one before it is in the swap record read here.
diff --git a/controller/internal/appexport/restore.go b/controller/internal/appexport/restore.go
index 2047d42..80fd964 100644
--- a/controller/internal/appexport/restore.go
+++ b/controller/internal/appexport/restore.go
@@ -938,6 +938,7 @@ func classifyDBImage(service, image, container string) *dbServiceInfo {
 
 // composeExecEnv runs docker compose in the given stack directory with env vars.
 func composeExecEnv(stackDir string, env map[string]string, args ...string) ([]byte, error) {
+	defer dockerexec.BeginImageWork(args)() // R-863: no image clean-up while a restore may pull
 	cmdArgs := append([]string{"compose"}, args...)
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
 	defer cancel()
diff --git a/controller/internal/backup/offbox.go b/controller/internal/backup/offbox.go
index 14feb4e..46bef1e 100644
--- a/controller/internal/backup/offbox.go
+++ b/controller/internal/backup/offbox.go
@@ -331,8 +331,8 @@ func (m *Manager) resetOrphanedRepo(ctx context.Context, base, env []string, rea
 		if err != nil {
 			return fmt.Errorf("offbox move-aside failed: %w", err)
 		}
+		newPath = np // R-869 (v0.294.0): assigned BEFORE the log line — v0.293.0 logged an empty destination
 		m.logger.Printf("[INFO] [offbox] the hub set the orphaned repo aside: %s -> %s (nothing deleted)", t.RepoPath, newPath)
-		newPath = np
 	} else {
 		port := t.Port
 		if port == 0 {
diff --git a/controller/internal/backup/offbox_window.go b/controller/internal/backup/offbox_window.go
index 4a8be9e..3fdcbb6 100644
--- a/controller/internal/backup/offbox_window.go
+++ b/controller/internal/backup/offbox_window.go
@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"sort"
+	"strconv"
 	"strings"
 	"time"
 )
@@ -22,9 +23,14 @@ import (
 // snapshot for removal. So, refuse when:
 //   - any snapshot is dated in the future (beyond offsiteGuardSkew), or after the hub's newest-allowed
 //     bound (the moment the window opened, plus the same skew);
-//   - the plan would remove a snapshot younger than offsiteGuardMinAge — the honest policy
-//     (--keep-daily 7) never removes the newest snapshot of any of the last 7 days, while a poisoning
-//     shape does exactly that.
+//   - the plan would remove a snapshot whose calendar day is fewer than keepDaily days before today and
+//     that is not superseded the same day — the honest policy (--keep-daily keepDaily) never does that,
+//     while a poisoning shape does exactly that. v0.294.0 (R-867): the line is DERIVED from keepDaily, the
+//     same constant the policy is built from. v0.289–0.293 used a fixed 8-day AGE, which sat inside the
+//     keep window: the snapshot a keep-7-dailies policy drops each night is 7 days + seconds old, so every
+//     window on a box with more than 7 nightly snapshots refused and mailed an error (measured
+//     demo-felhom 2026-10-05). Pinned by TestOffsiteGuard_RealPolicy* (the policy itself, simulated as
+//     restic 0.14.0 applies it, over 15+ nightly snapshots and a month boundary).
 // And a plan larger than MaxRemove (the hub's number for one week) REFUSES (v0.290.0, per the 2026-10-04
 // brief, replacing v0.289's cap). The cost, recorded (R-96 rule 4): after a long gap without windows the
 // honest backlog exceeds a week and the guard refuses until the operator grants a window by hand — R-833.
@@ -33,9 +39,14 @@ import (
 //
 // Pinned by TestOffsiteGuard_* (offbox_window_test.go), including the lab's 13-fake shape.
 
+const offsiteGuardSkew = time.Hour
+
+// The ruled retention policy (SP-2), as constants: BOTH the policy's arguments and the guard's day line
+// are built from these, so the guard can never sit inside the keep window (R-867).
 const (
-	offsiteGuardSkew   = time.Hour
-	offsiteGuardMinAge = 8 * 24 * time.Hour
+	keepDaily   = 7
+	keepWeekly  = 4
+	keepMonthly = 6
 )
 
 // OffsiteWindow is the hub's answer to "may I prune now?".
@@ -67,7 +78,20 @@ type OffsiteWindowClient interface {
 func (m *Manager) SetOffsiteWindowClient(c OffsiteWindowClient) { m.offsiteWindow = c }
 
 // retentionPolicy is the ruled policy, unchanged since SP-2 (`--group-by host,tags`).
-var retentionPolicy = []string{"--group-by", "host,tags", "--keep-daily", "7", "--keep-weekly", "4", "--keep-monthly", "6"}
+var retentionPolicy = []string{"--group-by", "host,tags",
+	"--keep-daily", strconv.Itoa(keepDaily), "--keep-weekly", strconv.Itoa(keepWeekly), "--keep-monthly", strconv.Itoa(keepMonthly)}
+
+// calendarDaysBefore: how many calendar days s lies before now, both read in s's own zone — the zone
+// restic 0.14.0 buckets a snapshot's day in (the offset stored with the snapshot). Edge, recorded: in the
+// hour after local midnight on a DST change, a box whose snapshots carry two different offsets can read one
+// day short; the guard then REFUSES (the safe direction) and the next week's window passes.
+func calendarDaysBefore(s, now time.Time) int {
+	loc := s.Location()
+	a, b := s.In(loc), now.In(loc)
+	da := time.Date(a.Year(), a.Month(), a.Day(), 0, 0, 0, 0, time.UTC)
+	db := time.Date(b.Year(), b.Month(), b.Day(), 0, 0, 0, 0, time.UTC)
+	return int(db.Sub(da).Hours() / 24)
+}
 
 type guardSnap struct {
 	ID       string    `json:"id"`
@@ -98,7 +122,8 @@ func supersededSameDay(s guardSnap, all []guardSnap) bool {
 }
 
 // offsiteGuard is the PURE decision: from all snapshots and the policy's remove-plan, either the ids to
-// remove (oldest first) or a refusal reason. v0.290.0 (R-824): a YOUNG snapshot that a newer same-day
+// remove (oldest first) or a refusal reason. v0.294.0 (R-867): "young" means fewer than keepDaily calendar
+// days before today, not an age in hours. v0.290.0 (R-824): a YOUNG snapshot that a newer same-day
 // snapshot of its group supersedes is EXCLUDED (kept for a later window, when it is old) instead of
 // refusing the run — v0.289.x refused every window after any manual run. A young removal WITHOUT that
 // explanation still refuses: it is the poisoning signature. Future-dated snapshots, snapshots newer than
@@ -114,12 +139,12 @@ func offsiteGuard(all, plan []guardSnap, now, newestAllowed time.Time, maxRemove
 	}
 	var keep []guardSnap
 	for _, s := range plan {
-		if now.Sub(s.Time) < offsiteGuardMinAge {
+		if calendarDaysBefore(s.Time, now) < keepDaily {
 			if supersededSameDay(s, all) {
-				continue // excluded: removed in a later window, once older than offsiteGuardMinAge
+				continue // excluded: removed in a later window, once keepDaily days old
 			}
-			return nil, fmt.Sprintf("the policy would remove snapshot %s from %s — younger than %d days and not superseded the same day, which honest retention never does",
-				s.ShortID, s.Time.UTC().Format(time.RFC3339), int(offsiteGuardMinAge.Hours()/24))
+			return nil, fmt.Sprintf("the policy would remove snapshot %s from %s — within the last %d days kept daily and not superseded the same day, which honest retention never does",
+				s.ShortID, s.Time.UTC().Format(time.RFC3339), keepDaily)
 		}
 		keep = append(keep, s)
 	}
diff --git a/controller/internal/backup/offbox_window_test.go b/controller/internal/backup/offbox_window_test.go
index b880577..d22881a 100644
--- a/controller/internal/backup/offbox_window_test.go
+++ b/controller/internal/backup/offbox_window_test.go
@@ -1,9 +1,12 @@
 package backup
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"fmt"
+	"log"
+	"sort"
 	"strings"
 	"testing"
 	"time"
@@ -160,7 +163,7 @@ func TestOffsiteGuard_RecentRemovalRefused(t *testing.T) {
 	now := time.Now()
 	all := []guardSnap{snap("old", now.Add(-60*24*time.Hour)), snap("recent", now.Add(-3*24*time.Hour))}
 	_, why := offsiteGuard(all, []guardSnap{snap("recent", now.Add(-3*24*time.Hour))}, now, now, 50)
-	if !strings.Contains(why, "younger than 8 days") {
+	if !strings.Contains(why, "within the last 7 days kept daily") {
 		t.Fatalf("why = %q", why)
 	}
 	_, why = offsiteGuard(all, nil, now, now.Add(-10*24*time.Hour), 50)
@@ -251,6 +254,24 @@ func TestResetOrphaned_PinnedAsksTheHub(t *testing.T) {
 	}
 }
 
+// R-869 (v0.294.0): the move-aside line names the destination. MEASURED 2026-10-05 03:08 UTC, Tester 1 box
+// (v0.293.0): `the hub set the orphaned repo aside: /home/felhom-repo ->  (nothing deleted)`.
+// COMPANION RED-PROOF: move `newPath = np` below the log line → "the line names no destination".
+func TestR869_MoveAsideLineNamesTheDestination(t *testing.T) {
+	m, sett := newOffboxManager(t)
+	var buf bytes.Buffer
+	m.logger = log.New(&buf, "", 0)
+	pinTarget(t, sett)
+	m.SetOffsiteMoveAside(func(context.Context) (string, error) { return "/home/felhom-repo.orphaned-20261005", nil })
+	m.SetOffboxRunner(func(context.Context, []string, ...string) ([]byte, error) { return nil, nil })
+	if err := m.resetOrphanedRepo(context.Background(), nil, nil, "test"); err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(buf.String(), "aside: /home/felhom-repo -> /home/felhom-repo.orphaned-20261005 (nothing deleted)") {
+		t.Fatalf("the line names no destination:\n%s", buf.String())
+	}
+}
+
 // The provider's rclone notice must not reach a JSON parser — measured live on demo-felhom (v0.289.0).
 func TestStripRcloneNotice(t *testing.T) {
 	in := "rclone: 2026/10/03 15:05:42 NOTICE: Config file \"/home/.config/rclone/rclone.conf\" not found - using defaults\n[{\"id\":\"s1\"}]\n"
@@ -343,3 +364,200 @@ func TestAbandon_PinnedHandsToHubAndFollows(t *testing.T) {
 		t.Fatalf("del=%v err=%v status=%+v evs=%v", del, err, m.AbandonStatus(), evs)
 	}
 }
+
+// ── R-867 (v0.294.0): the guard against the REAL policy ─────────────────────────────────────────────────
+
+// restic0140Plan is restic 0.14.0's `forget --group-by host,tags --keep-daily D --keep-weekly W
+// --keep-monthly M` (internal/restic/snapshot_policy.go ApplyPolicy): per group, newest first, a snapshot
+// is kept when it opens a new day / ISO week / month bucket while that rule still has count left; every
+// other snapshot is removed. Equivalence with the real binary over the same snapshot set is recorded in
+// felhom.eu/documentation/audits/night-fixes-2026-10-05/partA/ (the lab run).
+func restic0140Plan(all []guardSnap, daily, weekly, monthly int) (remove []guardSnap) {
+	groups := map[string][]guardSnap{}
+	var order []string
+	for _, s := range all {
+		if _, ok := groups[s.group()]; !ok {
+			order = append(order, s.group())
+		}
+		groups[s.group()] = append(groups[s.group()], s)
+	}
+	for _, g := range order {
+		list := append([]guardSnap{}, groups[g]...)
+		sort.SliceStable(list, func(i, j int) bool { return list[i].Time.After(list[j].Time) })
+		type bucket struct {
+			count int
+			f     func(time.Time) int
+			last  int
+		}
+		b := []bucket{
+			{daily, func(d time.Time) int { return d.Year()*10000 + int(d.Month())*100 + d.Day() }, -1},
+			{weekly, func(d time.Time) int { y, w := d.ISOWeek(); return y*100 + w }, -1},
+			{monthly, func(d time.Time) int { return d.Year()*100 + int(d.Month()) }, -1},
+		}
+		for _, cur := range list {
+			keep := false
+			for i := range b {
+				if b[i].count > 0 {
+					if v := b[i].f(cur.Time); v != b[i].last {
+						keep = true
+						b[i].last = v
+						b[i].count--
+					}
+				}
+			}
+			if !keep {
+				remove = append(remove, cur)
+			}
+		}
+	}
+	return remove
+}
+
+func without(all []guardSnap, ids []string) []guardSnap {
+	gone := map[string]bool{}
+	for _, id := range ids {
+		gone[id] = true
+	}
+	var out []guardSnap
+	for _, s := range all {
+		if !gone[s.ID] {
+			out = append(out, s)
+		}
+	}
+	return out
+}
+
+// THE MEASURED SHAPE (demo-felhom window 3, 2026-10-05 02:15 UTC): the night's snapshot is taken, then the
+// window runs; the policy drops the snapshot of 7 days + 5 s ago. v0.293.0's 8-day line refused it.
+func TestOffsiteGuard_RealPolicyMeasuredShapeAllowed(t *testing.T) {
+	var all []guardSnap
+	for d := 0; d < 8; d++ {
+		at := time.Date(2026, 9, 28+d, 2, 15, 5, 0, time.UTC)
+		all = append(all, guardSnap{ID: fmt.Sprintf("n%d-full", d), ShortID: fmt.Sprintf("n%d", d), Time: at, Hostname: "demo-felhom", Tags: []string{"opengist"}})
+	}
+	now := time.Date(2026, 10, 5, 2, 15, 10, 0, time.UTC)
+	plan := restic0140Plan(all, keepDaily, keepWeekly, keepMonthly)
+	if len(plan) != 1 || plan[0].ShortID != "n0" {
+		t.Fatalf("the policy's plan = %v, want the 2026-09-28 snapshot only", plan)
+	}
+	ids, why := offsiteGuard(all, plan, now, now, 5)
+	if why != "" || len(ids) != 1 || ids[0] != "n0-full" {
+		t.Fatalf("the honest drop of a 7 d + 5 s snapshot must pass: ids=%v why=%q", ids, why)
+	}
+}
+
+// Sixty nights of a household with three apps, a window after every night's run (the worst case: the hub
+// opens weekly), a same-day manual run on two days, and a month boundary: the real policy's plan is never
+// refused, deletes happen, and the store settles at the policy's size. Weekly and monthly keeps survive.
+func TestOffsiteGuard_RealPolicySixtyNightsNeverRefuses(t *testing.T) {
+	apps := []string{"opengist", "bookstack", "immich"}
+	var all []guardSnap
+	removedTotal, windowsWithRemoval := 0, 0
+	start := time.Date(2026, 8, 20, 2, 15, 0, 0, time.UTC)
+	for night := 0; night < 60; night++ {
+		at := start.AddDate(0, 0, night)
+		for i, a := range apps {
+			all = append(all, guardSnap{ID: fmt.Sprintf("%s-%d-full", a, night), ShortID: fmt.Sprintf("%s%d", a, night),
+				Time: at.Add(time.Duration(i) * time.Second), Hostname: "box", Tags: []string{a}})
+		}
+		if night == 20 || night == 41 { // the household pressed "back up now" in the afternoon
+			for _, a := range apps {
+				all = append(all, guardSnap{ID: fmt.Sprintf("%s-%d-manual-full", a, night), ShortID: fmt.Sprintf("%s%dm", a, night),
+					Time: at.Add(13 * time.Hour), Hostname: "box", Tags: []string{a}})
+			}
+		}
+		now := at.Add(2 * time.Minute)
+		if night == 20 || night == 41 {
+			now = at.Add(13*time.Hour + 2*time.Minute)
+		}
+		plan := restic0140Plan(all, keepDaily, keepWeekly, keepMonthly)
+		ids, why := offsiteGuard(all, plan, now, now, 500)
+		if why != "" {
+			t.Fatalf("night %d (%s): the guard refused the honest policy: %s", night, now.Format(time.RFC3339), why)
+		}
+		if len(ids) > 0 {
+			windowsWithRemoval++
+			removedTotal += len(ids)
+		}
+		all = without(all, ids)
+	}
+	if windowsWithRemoval < 15 || removedTotal < 45 {
+		t.Fatalf("too little was ever removed (%d windows, %d snapshots): the test does not exercise deletion", windowsWithRemoval, removedTotal)
+	}
+	// The store holds the policy's shape per app, never only the last 7 days: the end-of-month keeps remain.
+	per := map[string]int{}
+	monthEnds := 0
+	for _, s := range all {
+		per[s.Tags[0]]++
+		if s.Tags[0] == "opengist" && (s.Time.Format("01-02") == "08-31" || s.Time.Format("01-02") == "09-30") {
+			monthEnds++
+		}
+	}
+	for _, a := range apps {
+		if per[a] < keepDaily || per[a] > keepDaily+keepWeekly+keepMonthly {
+			t.Fatalf("%s holds %d snapshots after 60 nights; policy bounds %d..%d", a, per[a], keepDaily, keepDaily+keepWeekly+keepMonthly)
+		}
+	}
+	if monthEnds != 2 {
+		t.Fatalf("the monthly keeps of 31 Aug and 30 Sep must survive; found %d", monthEnds)
+	}
+}
+
+// A weekly window (the hub's real cadence) over the same nights: the backlog of one week is removed in one
+// go and is never refused by the day line.
+func TestOffsiteGuard_RealPolicyWeeklyWindows(t *testing.T) {
+	var all []guardSnap
+	start := time.Date(2026, 9, 1, 2, 15, 0, 0, time.UTC)
+	removed := 0
+	for night := 0; night < 35; night++ {
+		at := start.AddDate(0, 0, night)
+		all = append(all, guardSnap{ID: fmt.Sprintf("s%d-full", night), ShortID: fmt.Sprintf("s%d", night), Time: at, Hostname: "box", Tags: []string{"app"}})
+		if night%7 != 6 {
+			continue
+		}
+		now := at.Add(time.Minute)
+		plan := restic0140Plan(all, keepDaily, keepWeekly, keepMonthly)
+		ids, why := offsiteGuard(all, plan, now, now, 7)
+		if why != "" {
+			t.Fatalf("weekly window on %s refused: %s", now.Format("2006-01-02"), why)
+		}
+		removed += len(ids)
+		all = without(all, ids)
+	}
+	if removed == 0 {
+		t.Fatal("five weekly windows removed nothing")
+	}
+}
+
+// Poisoning that the day line still catches with the REAL policy: just before midnight an add-only
+// attacker plants a snapshot 50 minutes ahead — inside the future-date skew, so not refused as future —
+// dated TOMORROW. The policy then counts tomorrow as a day and drops the real snapshot of six days ago.
+// (Past-dated fakes cannot make the policy drop a snapshot inside the last keepDaily calendar days: those
+// days are at most keepDaily distinct days and keep-daily keeps them all. Past-dated gap-fills that steer
+// OLDER keeps are R-822's residual, bounded by MaxRemove and the hub's count check, not by this line.)
+func TestOffsiteGuard_RealPolicySkewWindowFakeRefused(t *testing.T) {
+	now := time.Date(2026, 10, 12, 23, 30, 0, 0, time.UTC)
+	var all []guardSnap
+	for d := 0; d < 7; d++ {
+		all = append(all, guardSnap{ID: fmt.Sprintf("real%d-full", d), ShortID: fmt.Sprintf("real%d", d),
+			Time: time.Date(2026, 10, 12-d, 2, 15, 0, 0, time.UTC), Hostname: "box", Tags: []string{"app"}})
+	}
+	all = append(all, guardSnap{ID: "fake-full", ShortID: "fake", Time: now.Add(50 * time.Minute), Hostname: "box", Tags: []string{"app"}})
+	plan := restic0140Plan(all, keepDaily, keepWeekly, keepMonthly)
+	if len(plan) != 1 || plan[0].ShortID != "real6" {
+		t.Fatalf("plan = %v (want the real snapshot of 6 days ago)", plan)
+	}
+	_, why := offsiteGuard(all, plan, now, now, 50)
+	if !strings.Contains(why, "within the last 7 days kept daily") {
+		t.Fatalf("why = %q — a real snapshot inside the daily window must not be removed", why)
+	}
+}
+
+// The policy's arguments and the guard's line come from the same constants (R-867).
+func TestRetentionPolicy_BuiltFromTheGuardConstants(t *testing.T) {
+	got := strings.Join(retentionPolicy, " ")
+	want := fmt.Sprintf("--group-by host,tags --keep-daily %d --keep-weekly %d --keep-monthly %d", keepDaily, keepWeekly, keepMonthly)
+	if got != want || keepDaily != 7 || keepWeekly != 4 || keepMonthly != 6 {
+		t.Fatalf("policy = %q (want %q, the ruled 7/4/6)", got, want)
+	}
+}
diff --git a/controller/internal/dockerexec/imagework.go b/controller/internal/dockerexec/imagework.go
new file mode 100644
index 0000000..dbdf4e9
--- /dev/null
+++ b/controller/internal/dockerexec/imagework.go
@@ -0,0 +1,62 @@
+package dockerexec
+
+import "sync"
+
+// ── R-863 (v0.294.0): no image clean-up while an image is being pulled ───────────────────────────────
+//
+// MEASURED 2026-10-04 night drill, a fresh box: the controller's one-time image clean-up ran while the
+// household's first BookStack install was inside `docker compose up -d`. compose had pulled
+// `mariadb@sha256:…` (stored untagged) and not yet created the container, so no container, installed app
+// or undo named it; the clean-up deleted it one second before compose created the container, and the
+// install failed. An app being installed is not "installed" yet, and a restore or undo pulls the same way.
+//
+// THE RULE: every compose command that can pull an image and then create a container from it (up, pull,
+// create, run) holds the image-work lock SHARED for its whole run; an image clean-up pass takes it
+// EXCLUSIVELY, without waiting (TryLock). So a pass never runs while any such command runs, and a command
+// that starts during a pass waits the few seconds the pass takes. A pass that cannot take the lock does
+// not run; its caller tries again later. Pinned by TestR863_* (internal/stacks/image_retention_r863_test.go).
+
+var imageWorkMu sync.RWMutex
+
+// ImagePulling reports whether compose args are an image-pulling verb (up, pull, create, run). Flags
+// before the verb (`-p name`, `--profile x`) are skipped.
+func ImagePulling(args []string) bool {
+	for i := 0; i < len(args); i++ {
+		a := args[i]
+		if a == "compose" {
+			continue
+		}
+		if len(a) > 0 && a[0] == '-' {
+			switch a {
+			case "-p", "--project-name", "-f", "--file", "--profile", "--env-file", "--project-directory":
+				i++ // the flag's value
+			}
+			continue
+		}
+		switch a {
+		case "up", "pull", "create", "run":
+			return true
+		}
+		return false
+	}
+	return false
+}
+
+// BeginImageWork holds the image-work lock shared while compose args pull images; the returned func
+// releases it. For any other verb it returns a no-op.
+func BeginImageWork(args []string) (end func()) {
+	if !ImagePulling(args) {
+		return func() {}
+	}
+	imageWorkMu.RLock()
+	return imageWorkMu.RUnlock
+}
+
+// TryImageCleanup takes the image-work lock exclusively if no image work runs now. ok=false: something is
+// pulling — do not clean up now.
+func TryImageCleanup() (end func(), ok bool) {
+	if !imageWorkMu.TryLock() {
+		return nil, false
+	}
+	return imageWorkMu.Unlock, true
+}
diff --git a/controller/internal/stacks/deploy.go b/controller/internal/stacks/deploy.go
index bcc97b4..a5ca9c5 100644
--- a/controller/internal/stacks/deploy.go
+++ b/controller/internal/stacks/deploy.go
@@ -15,6 +15,7 @@ import (
 
 	"gitea.dooplex.hu/admin/felhom-controller/internal/appbackup"
 	"gitea.dooplex.hu/admin/felhom-controller/internal/crypto"
+	"gitea.dooplex.hu/admin/felhom-controller/internal/dockerexec"
 	"gitea.dooplex.hu/admin/felhom-controller/internal/system"
 	"gitea.dooplex.hu/admin/felhom-controller/internal/util"
 	"gopkg.in/yaml.v3"
@@ -837,7 +838,8 @@ func (m *Manager) PersistUnitRedeployConfig(name string, env map[string]string)
 // so USERDATA_PATH must be injected here too (mirrors stackEnv), else the FIRST deploy resolves
 // ${USERDATA_PATH} to "" and binds a bogus root-owned dir at the container root.
 func (m *Manager) composeExecWithEnv(dir string, env map[string]string, args ...string) (string, error) {
-	if m.composeExecFn != nil { // test seam (v0.280.0): a deploy test never reaches Docker
+	defer dockerexec.BeginImageWork(args)() // R-863 (held around the seam too, so a test sees the real rule)
+	if m.composeExecFn != nil {             // test seam (v0.280.0): a deploy test never reaches Docker
 		return m.composeExecFn(dir, env, args...)
 	}
 	cmdEnv := os.Environ()
diff --git a/controller/internal/stacks/image_retention.go b/controller/internal/stacks/image_retention.go
index b38f866..0b06a33 100644
--- a/controller/internal/stacks/image_retention.go
+++ b/controller/internal/stacks/image_retention.go
@@ -1,12 +1,15 @@
 package stacks
 
 import (
+	"context"
+	"errors"
 	"fmt"
 	"os"
 	"path/filepath"
 	"sort"
 	"strings"
 	"sync"
+	"time"
 
 	"gitea.dooplex.hu/admin/felhom-controller/internal/dockerexec"
 )
@@ -35,6 +38,10 @@ var imageDocker = func(args ...string) (string, error) {
 
 var imageRetentionMu sync.Mutex
 
+// errImageRetentionBusy: a pass did not run because an image may be in use by work in flight (an update, or
+// any compose command that pulls — R-863). The caller tries again later; nothing was judged.
+var errImageRetentionBusy = errors.New("image work in flight")
+
 type localImage struct {
 	ID, Repo, Tag, Digest, Size string
 }
@@ -194,8 +201,17 @@ func (m *Manager) deleteUnkeptImages(why string, repos map[string]bool, except s
 	m.mu.RUnlock()
 	if busy != "" {
 		m.logger.Printf("[INFO] [stacks] image retention (%s): skipped — %s is updating (its undo may need an image nothing else names)", why, busy)
-		return nil, nil
+		return nil, errImageRetentionBusy
 	}
+	// R-863: an install, a restore or an undo inside `compose up` may have pulled an image (by digest, so
+	// untagged) that no container names YET. No pass while any image-pulling compose command runs; one that
+	// starts now waits for this pass (seconds).
+	endCleanup, ok := dockerexec.TryImageCleanup()
+	if !ok {
+		m.logger.Printf("[INFO] [stacks] image retention (%s): skipped — an app install, update, restore or undo is pulling images now (R-863); tried again later", why)
+		return nil, errImageRetentionBusy
+	}
+	defer endCleanup()
 	imgs, err := listLocalImages()
 	if err != nil {
 		return nil, err
@@ -293,7 +309,7 @@ func (m *Manager) RetainImagesAfterUpdate(name string, previous map[string]Insta
 		m.logger.Printf("[INFO] [stacks] image retention after the update of %s: the app is gone — nothing to do here", name)
 		return
 	}
-	if _, err := m.deleteUnkeptImages("update of "+name, appImageRepos(dir, st.AppConfig), ""); err != nil {
+	if _, err := m.deleteUnkeptImages("update of "+name, appImageRepos(dir, st.AppConfig), ""); err != nil && !errors.Is(err, errImageRetentionBusy) {
 		m.logger.Printf("[WARN] [stacks] image retention after the update of %s: %v", name, err)
 	}
 }
@@ -317,7 +333,7 @@ func (m *Manager) RetainImagesAfterRemove(name string, repos map[string]bool) {
 	if len(repos) == 0 {
 		return
 	}
-	if _, err := m.deleteUnkeptImages("remove of "+name, repos, name); err != nil {
+	if _, err := m.deleteUnkeptImages("remove of "+name, repos, name); err != nil && !errors.Is(err, errImageRetentionBusy) {
 		m.logger.Printf("[WARN] [stacks] image retention after the remove of %s: %v", name, err)
 	}
 }
@@ -349,28 +365,49 @@ func (m *Manager) imageRetentionMarker() string {
 
 // RunImageRetentionOnce is the one-time clean-up at the first start of this release: the same rule, applied to every
 // app image the catalog names (so the images of apps removed before this release go too). Logged; a marker file
-// keeps it to once. Returns what it deleted.
-func (m *Manager) RunImageRetentionOnce() []string {
+// keeps it to once. Returns what it deleted, and done=false when it must be tried again (no catalog yet, or image
+// work in flight — R-863: the marker is written ONLY after a pass that ran).
+func (m *Manager) RunImageRetentionOnce() (deleted []string, done bool) {
 	if _, err := os.Stat(m.imageRetentionMarker()); err == nil {
-		return nil
+		return nil, true
 	}
 	repos := m.catalogImageRepos()
 	if len(repos) == 0 {
-		m.logger.Printf("[WARN] [stacks] image retention (one-time): no catalog read — skipped, tried again at the next start")
-		return nil
+		m.logger.Printf("[WARN] [stacks] image retention (one-time): no catalog read — skipped, tried again later")
+		return nil, false
 	}
 	before, _ := imageDocker("system", "df", "--format", "{{.Type}} {{.Size}} {{.Reclaimable}}")
 	deleted, err := m.deleteUnkeptImages("one-time clean-up", repos, "")
+	if errors.Is(err, errImageRetentionBusy) {
+		return nil, false // logged by the pass; tried again later
+	}
 	if err != nil {
-		m.logger.Printf("[WARN] [stacks] image retention (one-time): %v — tried again at the next start", err)
-		return nil
+		m.logger.Printf("[WARN] [stacks] image retention (one-time): %v — tried again later", err)
+		return nil, false
 	}
 	after, _ := imageDocker("system", "df", "--format", "{{.Type}} {{.Size}} {{.Reclaimable}}")
 	m.logger.Printf("[INFO] [stacks] image retention (one-time): deleted %d image(s). docker disk before: %s | after: %s",
 		len(deleted), strings.Join(strings.Fields(firstLine(before)), " "), strings.Join(strings.Fields(firstLine(after)), " "))
 	_ = os.MkdirAll(filepath.Dir(m.imageRetentionMarker()), 0o755)
 	_ = os.WriteFile(m.imageRetentionMarker(), []byte(fmt.Sprintf("deleted %d\n%s\n", len(deleted), strings.Join(deleted, "\n"))), 0o644)
-	return deleted
+	return deleted, true
+}
+
+// RunImageRetentionOnceUntilDone runs the one-time clean-up, and again every `every` until it has run (R-863:
+// a pass that met image work in flight, or found no catalog yet, is retried — no longer only at the next start),
+// at most `tries` times.
+func (m *Manager) RunImageRetentionOnceUntilDone(ctx context.Context, every time.Duration, tries int) {
+	for i := 0; i < tries; i++ {
+		if _, done := m.RunImageRetentionOnce(); done {
+			return
+		}
+		select {
+		case <-ctx.Done():
+			return
+		case <-time.After(every):
+		}
+	}
+	m.logger.Printf("[WARN] [stacks] image retention (one-time): not run after %d tries — tried again at the next start", tries)
 }
 
 func sortedKeys(m map[string]bool) []string {
diff --git a/controller/internal/stacks/image_retention_r863_test.go b/controller/internal/stacks/image_retention_r863_test.go
new file mode 100644
index 0000000..e972eb9
--- /dev/null
+++ b/controller/internal/stacks/image_retention_r863_test.go
@@ -0,0 +1,109 @@
+package stacks
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-controller/internal/dockerexec"
+)
+
+// R-863 (v0.294.0) — THE NIGHT'S EXACT SHAPE (2026-10-04, a fresh box): the household's first install is inside
+// `compose up`; compose has pulled `mariadb@sha256:…` (stored untagged, `mariadb:<none>`) and not yet created the
+// container; the controller's one-time image clean-up fires (3 minutes after its first start). v0.293.0 deleted
+// the image — no container, installed app or undo named it — and `compose up` failed one second later.
+// COMPANION RED-PROOF: drop the dockerexec.TryImageCleanup check in deleteUnkeptImages → the clean-up deletes
+// sha256:MDB and the install fails ("the install failed").
+func TestR863_OneTimeCleanupDuringFirstInstallKeepsThePulledImage(t *testing.T) {
+	m := gateManager(t, "display_name: Book\ndeploy_fields:\n  - env_var: DOMAIN\n    type: domain\n  - env_var: SUBDOMAIN\n    type: subdomain\n    default: gapp\n")
+	m.cfg.Paths.DataDir = filepath.Join(t.TempDir(), "data")
+	cat := filepath.Join(m.cfg.Paths.DataDir, "catalog-cache", "templates", "bookstack")
+	must(t, os.MkdirAll(cat, 0o755))
+	must(t, os.WriteFile(filepath.Join(cat, "docker-compose.yml"), []byte("services:\n  db:\n    image: mariadb:11.4@sha256:mdb\n"), 0o644))
+
+	f := &fakeImages{containers: map[string]string{}}
+	var fmu sync.Mutex
+	prev := imageDocker
+	imageDocker = func(args ...string) (string, error) { fmu.Lock(); defer fmu.Unlock(); return f.run(args...) }
+	t.Cleanup(func() { imageDocker = prev })
+
+	var cleanupDone bool
+	m.composeExecFn = func(_ string, _ map[string]string, args ...string) (string, error) {
+		if len(args) == 0 || args[0] != "up" {
+			return "", nil
+		}
+		// compose pulls by digest: the image exists, untagged, named by no container yet
+		fmu.Lock()
+		f.imgs = append(f.imgs, localImage{ID: "sha256:MDB", Repo: "mariadb", Tag: "<none>", Digest: "sha256:mdb", Size: "334MB"})
+		fmu.Unlock()
+		// the one-time clean-up fires NOW, from its own goroutine, as at 19:45:04
+		res := make(chan bool, 1)
+		go func() { _, done := m.RunImageRetentionOnce(); res <- done }()
+		select {
+		case cleanupDone = <-res:
+		case <-time.After(10 * time.Second):
+			return "", fmt.Errorf("the clean-up blocked")
+		}
+		// compose creates the container from the pulled image — if it is still there
+		fmu.Lock()
+		defer fmu.Unlock()
+		for _, im := range f.imgs {
+			if im.ID == "sha256:MDB" {
+				f.containers["c-db"] = "sha256:MDB"
+				return "", nil
+			}
+		}
+		return "", fmt.Errorf("exit code 1\nstderr: Error response from daemon: No such image: mariadb@sha256:mdb")
+	}
+	done := make(chan bool, 1)
+	m.SetDeployDoneHook(func(_ string, ok bool, _ string) { done <- ok })
+	if _, err := m.DeployStack(DeployRequest{StackName: "gapp"}); err != nil {
+		t.Fatal(err)
+	}
+	select {
+	case ok := <-done:
+		if !ok {
+			t.Fatalf("the install failed: the clean-up deleted the image compose had just pulled (rmi %v)", f.rmi)
+		}
+	case <-time.After(20 * time.Second):
+		t.Fatal("the deploy never ended")
+	}
+	if len(f.rmi) != 0 {
+		t.Fatalf("the clean-up deleted during the install: %v", f.rmi)
+	}
+	if cleanupDone {
+		t.Fatal("the clean-up reported done although it did not run — its marker would end it for good")
+	}
+	if _, err := os.Stat(m.imageRetentionMarker()); err == nil {
+		t.Fatal("marker written for a pass that did not run")
+	}
+	// After the install the retried clean-up runs, and the app's image is kept (a container names it).
+	if _, done := m.RunImageRetentionOnce(); !done {
+		t.Fatal("the retried clean-up did not run once nothing was pulling")
+	}
+	if strings.Contains(strings.Join(f.rmi, ","), "sha256:MDB") {
+		t.Fatal("the installed app's database image was deleted")
+	}
+}
+
+// Image-pulling verbs hold the lock; others do not (a `ps` or `down` must never hold off a clean-up).
+func TestR863_OnlyPullingVerbsHoldTheLock(t *testing.T) {
+	for _, c := range []struct {
+		args []string
+		want bool
+	}{
+		{[]string{"up", "-d"}, true}, {[]string{"compose", "up", "-d", "--remove-orphans"}, true},
+		{[]string{"pull"}, true}, {[]string{"-p", "x", "create"}, true}, {[]string{"run", "--rm", "x"}, true},
+		{[]string{"ps"}, false}, {[]string{"down"}, false}, {[]string{"stop"}, false}, {[]string{"-p", "up", "down"}, false},
+	} {
+		if got := imagePullingForTest(c.args); got != c.want {
+			t.Errorf("%v: pulling=%v, want %v", c.args, got, c.want)
+		}
+	}
+}
+
+func imagePullingForTest(args []string) bool { return dockerexec.ImagePulling(args) }
diff --git a/controller/internal/stacks/image_retention_test.go b/controller/internal/stacks/image_retention_test.go
index ddf3d6a..d59bfc3 100644
--- a/controller/internal/stacks/image_retention_test.go
+++ b/controller/internal/stacks/image_retention_test.go
@@ -1,6 +1,7 @@
 package stacks
 
 import (
+	"errors"
 	"fmt"
 	"os"
 	"path/filepath"
@@ -214,7 +215,7 @@ func TestImageRetention_OneTimeSweepOnlyCatalogRepos(t *testing.T) {
 	cat := filepath.Join(m.cfg.Paths.DataDir, "catalog-cache", "templates", "gone")
 	must(t, os.MkdirAll(cat, 0o755))
 	must(t, os.WriteFile(filepath.Join(cat, "docker-compose.yml"), []byte("services:\n  gone:\n    image: acme/gone:6\n"), 0o644))
-	deleted := m.RunImageRetentionOnce()
+	deleted, _ := m.RunImageRetentionOnce()
 	got := strings.Join(f.rmi, ",")
 	if !strings.Contains(got, "sha256:OLD") {
 		t.Fatalf("an earlier-removed app's image was not swept: %v", deleted)
@@ -242,8 +243,8 @@ func TestImageRetention_NoPassWhileAnUpdateRuns(t *testing.T) {
 	m.stacks["docs"].Updating = true
 	m.mu.Unlock()
 	st, _ := m.GetStack("web")
-	if _, err := m.deleteUnkeptImages("test", appImageRepos(filepath.Dir(st.ComposePath), st.AppConfig), ""); err != nil {
-		t.Fatal(err)
+	if _, err := m.deleteUnkeptImages("test", appImageRepos(filepath.Dir(st.ComposePath), st.AppConfig), ""); !errors.Is(err, errImageRetentionBusy) {
+		t.Fatalf("err = %v, want errImageRetentionBusy (the caller tries again later)", err)
 	}
 	if len(f.rmi) != 0 {
 		t.Fatalf("a pass ran while docs was updating: %v", f.rmi)
diff --git a/controller/internal/stacks/manager.go b/controller/internal/stacks/manager.go
index f7462f2..8380587 100644
--- a/controller/internal/stacks/manager.go
+++ b/controller/internal/stacks/manager.go
@@ -14,6 +14,7 @@ import (
 	"strings"
 	"sync"
 	"time"
+	"unicode/utf8"
 
 	"gitea.dooplex.hu/admin/felhom-controller/internal/appbackup"
 	"gitea.dooplex.hu/admin/felhom-controller/internal/config"
@@ -1446,6 +1447,7 @@ func (m *Manager) composeExec(dir string, args ...string) (string, error) {
 }
 
 func (m *Manager) composeExecCustomEnv(dir string, env []string, args ...string) (string, error) {
+	defer dockerexec.BeginImageWork(args)() // R-863: no image clean-up while this may pull
 	var cmd *exec.Cmd
 
 	if m.composeCmd == "docker compose" {
@@ -1511,10 +1513,12 @@ func (m *Manager) composeExecCustomEnv(dir string, env []string, args ...string)
 		if stdoutStr := truncateStr(stdout.String(), 500); stdoutStr != "" {
 			m.logger.Printf("[ERROR] [stacks] stdout: %s", stdoutStr)
 		}
-		if stderrStr := truncateStr(stderr.String(), 500); stderrStr != "" {
-			m.logger.Printf("[ERROR] [stacks] stderr: %s", stderrStr)
+		// R-864 (v0.294.0): the TAIL of stderr — compose prints its pull progress first and the reason last, so
+		// the head (v0.293.0 and earlier) logged "Image … Pulling" and cut the error itself.
+		if stderrStr := tailStr(stderr.String(), 500); stderrStr != "" {
+			m.logger.Printf("[ERROR] [stacks] stderr (last part): %s", stderrStr)
 		}
-		return stdout.String(), fmt.Errorf("exit code %d\nstderr: %s", exitCode, truncateStr(stderr.String(), 500))
+		return stdout.String(), fmt.Errorf("exit code %d\nstderr: %s", exitCode, tailStr(stderr.String(), 500))
 	}
 
 	m.logger.Printf("[DEBUG] Command completed: %s %s (took %.1fs)", m.composeCmd, strings.Join(args, " "), time.Since(start).Seconds())
@@ -1549,6 +1553,20 @@ func (m *Manager) isDebug() bool {
 }
 
 // truncateStr truncates a string to maxLen characters, appending "..." if truncated.
+// tailStr keeps the LAST maxLen bytes of s (an error is printed last), marked with a leading "...". The cut
+// moves forward to a UTF-8 boundary so a Hungarian letter is never split. Pinned by TestR864_*.
+func tailStr(s string, maxLen int) string {
+	s = strings.TrimSpace(s)
+	if len(s) <= maxLen {
+		return s
+	}
+	i := len(s) - maxLen
+	for i < len(s) && !utf8.RuneStart(s[i]) {
+		i++
+	}
+	return "..." + s[i:]
+}
+
 func truncateStr(s string, maxLen int) string {
 	s = strings.TrimSpace(s)
 	if len(s) <= maxLen {
diff --git a/controller/internal/stacks/r864_stderr_tail_test.go b/controller/internal/stacks/r864_stderr_tail_test.go
new file mode 100644
index 0000000..057d3bf
--- /dev/null
+++ b/controller/internal/stacks/r864_stderr_tail_test.go
@@ -0,0 +1,61 @@
+package stacks
+
+import (
+	"bytes"
+	"log"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// R-864 (v0.294.0): a failed compose command logs and returns the TAIL of stderr. THE NIGHT'S SHAPE (2026-10-04,
+// BookStack on the fresh Tester 1 box, audits/night-2026-10-04/tester1/t8-*): the stderr opened with the pull
+// progress below (the first line is verbatim from that night) and the reason came last; v0.293.0 kept the first
+// 500 bytes and logged only "Image … Pulling" — the reason was lost. The night's last line itself was lost by
+// exactly this defect; the one here is Docker's message for an image deleted under compose (R-863).
+// COMPANION RED-PROOF: use truncateStr instead of tailStr for stderr in composeExecCustomEnv → "the reason was cut".
+func TestR864_FailedComposeLogsTheReasonNotThePullProgress(t *testing.T) {
+	var lines []string
+	lines = append(lines, "Image lscr.io/linuxserver/bookstack:26.09.1@sha256:99cd1f5707c1911afad213adec5c9739763b76f843d1477142231834ecdcb6f7 Pulling")
+	lines = append(lines, "Image mariadb:11.4@sha256:dfff46ef3f9d Pulling")
+	for i := 0; i < 16; i++ {
+		lines = append(lines, "  a1b2c3d4e5f6 Pull complete")
+	}
+	lines = append(lines, "Image mariadb:11.4@sha256:dfff46ef3f9d Pulled")
+	reason := "Error response from daemon: No such image: mariadb@sha256:dfff46ef3f9d — árvíztűrő"
+	lines = append(lines, reason)
+	stderr := strings.Join(lines, "\n")
+	if len(stderr) < 700 {
+		t.Fatalf("the fixture must exceed the 500-byte cut (%d)", len(stderr))
+	}
+	m := gateManager(t, "display_name: G\n")
+	dir := t.TempDir() // a stub under os.TempDir(): R-650's sanctioned seam, never this host's Docker
+	errf := filepath.Join(dir, "stderr.txt")
+	must(t, os.WriteFile(errf, []byte(stderr), 0o644))
+	must(t, os.WriteFile(filepath.Join(dir, "docker"), []byte("#!/bin/sh\nwhile IFS= read -r l || [ -n \"$l\" ]; do printf '%s\\n' \"$l\" >&2; done < "+errf+"\nexit 1\n"), 0o755))
+	t.Setenv("PATH", dir)
+	var buf bytes.Buffer
+	m.logger = log.New(&buf, "", 0)
+	_, err := m.composeExecCustomEnv(dir, []string{"PATH=" + dir}, "up", "-d")
+	if err == nil {
+		t.Fatal("no error from a failed compose")
+	}
+	if !strings.Contains(err.Error(), reason) {
+		t.Fatalf("the reason was cut from the returned error: %q", err.Error())
+	}
+	if !strings.Contains(buf.String(), reason) {
+		t.Fatalf("the reason was cut from the log:\n%s", buf.String())
+	}
+}
+
+func TestR864_TailStrKeepsTheEndOnARuneBoundary(t *testing.T) {
+	s := strings.Repeat("x", 10) + "őőő"
+	got := tailStr(s, 5) // the cut falls inside "ő" (2 bytes each)
+	if got != "...őő" {
+		t.Fatalf("tailStr = %q", got)
+	}
+	if tailStr("short", 500) != "short" {
+		t.Fatal("a short string must pass unchanged")
+	}
+}
diff --git a/controller/internal/stacks/update.go b/controller/internal/stacks/update.go
index e505feb..b8a5b12 100644
--- a/controller/internal/stacks/update.go
+++ b/controller/internal/stacks/update.go
@@ -9,6 +9,7 @@ import (
 	"strings"
 	"time"
 
+	"gitea.dooplex.hu/admin/felhom-controller/internal/dockerexec"
 	"gitea.dooplex.hu/admin/felhom-controller/internal/i18n"
 	"gitea.dooplex.hu/admin/felhom-controller/internal/system"
 	"gitea.dooplex.hu/admin/felhom-controller/internal/util"
@@ -690,6 +691,7 @@ func (m *Manager) UpdateErrorFor(st Stack, lang string) string {
 }
 
 func (m *Manager) updateCompose(dir string, env []string, args ...string) (string, error) {
+	defer dockerexec.BeginImageWork(args)() // R-863
 	if m.updateComposeFn != nil {
 		return m.updateComposeFn(dir, env, args...)
 	}
diff --git a/controller/internal/web/handlers.go b/controller/internal/web/handlers.go
index f3c05b2..ad2c439 100644
--- a/controller/internal/web/handlers.go
+++ b/controller/internal/web/handlers.go
@@ -3661,7 +3661,10 @@ func (s *Server) syncFileBrowserMounts(resetDBOnChange bool) {
 	}
 	cmd := dockerexec.CommandContext(ctx, "docker", args...)
 	cmd.Dir = stackDir
-	if out, err := cmd.CombinedOutput(); err != nil {
+	endImageWork := dockerexec.BeginImageWork(args) // R-863
+	out, err := cmd.CombinedOutput()
+	endImageWork()
+	if err != nil {
 		s.logger.Printf("[ERROR] [web] Failed to bring up FileBrowser: %s — %v", string(out), err)
 	} else if changed {
 		s.logger.Printf("[INFO] [web] FileBrowser mounts synced (recreated) — %d storage path(s), config updated", len(paths))