diff --git a/CHANGELOG.md b/CHANGELOG.md
index 70c8c7b..83c66c5 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,3 +1,29 @@
+## v0.265.0 — the cause of "runs but not installed", a held app that says so, a louder OOM storm (2026-09-23, R-634, R-625, R-636, R-647)
+
+**MinAgent: 0.131.0** (unchanged). **Needs hub v0.121.0** (deployed first). New strings: yes (hu + en).
+
+- **R-634 — the mechanism, diagnosed and fixed.** Reproduced on 9202: deploy `outline`, press the
+  whole-box backup 20 s later. The backup's list read the in-memory `Deployed` flag, which is true from
+  the moment a deploy is ACCEPTED, so the volume leg stopped the deploying app (`compose down`), dumped its
+  half-made volumes and ran a SECOND `compose up -d` beside the deploy's own; both failed and the deploy
+  recorded „not deployed" (on 2026-09-22 the backup's `up` won and the containers ran under it). Fix:
+  `ListDeployedStacks` leaves out a deploying app; the volume leg asks again right before the stop (SKIP,
+  never a failure); `StopStack` / `StartStack` refuse a deploying stack for every caller
+  (`ErrStackDeploying`). `sparkyfitness` did not reproduce alone (deployed in 47 s).
+- **R-625 — a held app shows the truth.** The update badge of a held app reads „Megállítva —
+  visszaállítás szükséges" / "Stopped — restore needed" (`tag-error`, title = the hold's first sentence
+  in the reader's language). No Update button (the list already hid it; now pinned). Parity fixtures
+  `app_info_deployed` and `stacks_full` regenerated — measured diff: exactly the new badge, one line each.
+- **R-636 — a repeating OOM problem gets louder.** The scan reads the kernel's `oom_kill` counter,
+  `memory.max` and `memory.peak` for flagged containers; 20+ kills in 30 min of the same container run →
+  ONE `app_oom_storm` (error, operator-only). `app_oom` unchanged. RomM's measured rate was ≈375 kills /
+  30 min; one hiccup is 1–3.
+- **R-647 — the three leftovers.** A held update's error is the key `update.error.held`, rendered per
+  reader (page funcs now apply to Hungarian too; same bytes on a Hungarian box); the `app_update_held`
+  details carry the `copy_holds` KEY, not the Hungarian phrase; the undone event no longer logs
+  `hold recorded`, and the disabled-notifier line names the real severity of a health change.
+- Red-proofs: eight in this repo, each seen failing (REPORT).
+
 ## v0.264.0 — the undo reaches the fleet, and the household is TOLD, in its own language (2026-09-23, `09` §6.4 parts 2–3)
 
 **MinAgent: 0.131.0** (unchanged). **Needs hub v0.120.0** (deployed first). New strings: yes (hu + en).
diff --git a/controller/README.md b/controller/README.md
index 227e59d..7a3b43c 100644
--- a/controller/README.md
+++ b/controller/README.md
@@ -650,6 +650,18 @@ the old version back; a power cut while undoing resumes the undo. Reasoning and
 `felhom.eu/documentation/architecture/09-update-architecture.md` §6.1a,
 `felhom.eu/documentation/audits/undo-bakeoff-2026-09-23/`.
 
+**Held apps say so (v0.265.0, R-625).** While a hold stands, the update badge reads „Megállítva —
+visszaállítás szükséges" / "Stopped — restore needed" (`tag-error`, title = the hold sentence's first
+sentence, in the reader's language) and no Update button is rendered; the API still answers 409 `held`. A
+held update's error is stored as the key `update.error.held` and rendered per reader, so a reader in the
+other language than the box reads the hold in theirs (R-647).
+
+**Deploys are not interrupted (v0.265.0, R-634).** A deploy still running is not in any backup leg's app
+list, the volume leg asks again right before it stops an app, and `StopStack` / `StartStack` refuse a
+deploying stack (`ErrStackDeploying`) for every caller. Measured cause: the whole-box backup's `compose down`
+and second `compose up -d` in the middle of the deploy's own `up` — the deploy then recorded „not deployed"
+over running containers.
+
 **The household is told (v0.264.0).** An undone update sends `app_update_undone` (warning) ONCE; an
 update that ends held sends `app_update_held` (error) ONCE — also when the hold itself could not be
 saved. Details `{app, stack_name, from, to, at, copy_tier, copy_date, copy_holds}`. Both are in
@@ -1132,6 +1144,7 @@ The nightly backup has two phases that run sequentially. All paths are **per-dri
 > **Absent-storage tier skip (v0.243.0, R-518).** A tier whose storage the agent (≥ 0.131.0) reports `absent` is dropped from the manual and the scheduled run before anything is stopped (`quiesce.skipAbsentTiers`), logged, and reported once as `backup_tier_skipped`. `unknown` or a legacy agent is never skipped.
 
 > **OOM visibility (v0.243.0, R-514).** The 30 s dead-app check also reads `State.OOMKilled` for running app containers (`Manager.ScanOOMKilled`); the dashboard shows „Memória elfogyott" and the hub gets `app_oom` once per container run. The controller does not restart the app.
+> **OOM storm (v0.265.0, R-636).** For a flagged container the scan also reads the kernel's own kill counter (`memory.events` `oom_kill`, plus `memory.max` / `memory.peak`) with one `docker exec … cat` — `OOMKilled` is sticky, so it cannot count. When the SAME container run's counter rises by **20 or more within 30 minutes**, the notifier sends **`app_oom_storm`** (severity `error`, operator-only, details `{app, container, kills, window_min, mem_limit, peak}`) — ONCE per container run. `app_oom` itself is unchanged. An unreadable counter (an image without `cat`) never escalates.
 
 > **Multi-tier whole-guest backup (v0.174.0, R-82 Slice B).** The agent can serve SEVERAL whole-guest
 > backup tiers with independent cadences — "local daily + PBS weekly" (agent >= v0.97.0,
@@ -2432,6 +2445,7 @@ The controller pushes structured events to the Hub's `/api/v1/event` endpoint. T
 | `app_start_failed` | **warning** | A DEPLOYED app is not running (fix-3) — fired ONCE per running→down transition. **Customer-switchable („Alkalmazás nem fut"), OFF by default; the OPERATOR is e-mailed regardless.** Was `warn` until v0.223.0 — see the severity note below |
 | `app_update_undone` | warning | v0.264.0 — a guarded update failed and the box put the previous version and its data back. Once per app per failed step. Household ON by default |
 | `app_update_held` | **error** | v0.264.0 — a guarded update (or its undo) failed and the app is held until a restore. The mail's line is the hold sentence in the household's language. Household ON by default |
+| `app_oom_storm` | **error** | v0.265.0 (R-636) — the same container run OOM-killed 20+ times in 30 min (the kernel's `oom_kill` counter). Once per container run. Operator-only |
 | `disaster_recovery_started` | warning | DR restore begins |
 | `disaster_recovery_completed` | info/error | DR restore finishes (success/partial) |
 
diff --git a/controller/cmd/controller/main.go b/controller/cmd/controller/main.go
index 176dad2..f44c4af 100644
--- a/controller/cmd/controller/main.go
+++ b/controller/cmd/controller/main.go
@@ -620,7 +620,7 @@ func main() {
 				d.CopyDate = ev.CopyDate.UTC().Format(time.RFC3339)
 			}
 			if backupMgr != nil && ev.CopyTier > 0 {
-				d.CopyHolds = backupMgr.UpdateCopyHolds(ev.App, ev.CopyTier)
+				d.CopyHolds = backupMgr.UpdateCopyHoldsKey(ev.App, ev.CopyTier) // R-647: the key, never the Hungarian phrase
 			}
 			notifier.NotifyAppUpdateHeld(d, func(lang string) string {
 				if ev.HoldRecorded && backupMgr != nil {
@@ -851,8 +851,8 @@ func main() {
 			logger.Printf("[WARN] [deadapp] OOM scan failed: %v", oerr)
 		} else {
 			for _, o := range ooms {
-				logger.Printf("[WARN] [deadapp] %s: container %s was OOM-killed (started %s)", o.Stack, o.Container, o.StartedAt)
-				notifier.NotifyAppOOM(o.Stack, o.Container, o.StartedAt)
+				logger.Printf("[WARN] [deadapp] %s: container %s was OOM-killed (started %s; kills %d, limit %s, peak %s)", o.Stack, o.Container, o.StartedAt, o.Kills, o.MemLimit, o.Peak)
+				notifier.NotifyAppOOM(o.Stack, o.Container, o.StartedAt, o.Kills, o.MemLimit, o.Peak)
 			}
 		}
 		deadAppScans++
@@ -2556,7 +2556,10 @@ func (a *stackAdapter) GetStackClassifiedBinds(name string) ([]backup.Classified
 func (a *stackAdapter) ListDeployedStacks() []backup.StackSummary {
 	var result []backup.StackSummary
 	for _, s := range a.mgr.GetStacks() {
-		if !s.Deployed {
+		// R-634 (v0.265.0): `Deployed` is set true in memory the moment a deploy is ACCEPTED (the
+		// no-stale-button UX), so without the second test a deploy still pulling was in every backup
+		// leg's list — and the volume leg stopped and restarted it in the middle of its own `up`.
+		if !s.Deployed || s.Deploying {
 			continue
 		}
 		result = append(result, backup.StackSummary{
@@ -2574,6 +2577,13 @@ func (a *stackAdapter) StopStack(name string) error {
 	return a.mgr.StopStack(name)
 }
 
+// IsDeploying answers backup's optional deployingReporter (R-634): the volume leg asks it again
+// immediately before it stops an app, because its list is taken once at the start of the run.
+func (a *stackAdapter) IsDeploying(name string) bool {
+	s, ok := a.mgr.GetStack(name)
+	return ok && s.Deploying
+}
+
 func (a *stackAdapter) StartStack(name string) error {
 	return a.mgr.StartStack(name)
 }
@@ -3471,6 +3481,14 @@ func (a *updateGuardsAdapter) HoldFor(name string) (bool, string) {
 	return a.b.RestoreHoldFor(name)
 }
 
+// HoldForLang renders the hold sentence for one reader (R-647, v0.265.0).
+func (a *updateGuardsAdapter) HoldForLang(name, lang string) (bool, string) {
+	if a.b == nil {
+		return false, ""
+	}
+	return a.b.RestoreHoldForLang(name, lang)
+}
+
 func (a *updateGuardsAdapter) Busy(name string) (bool, string) {
 	if a.q.SuppressedStacks()[name] {
 		return true, "a whole-guest backup (quiesce) is holding it"
diff --git a/controller/cmd/controller/update_events_wiring_test.go b/controller/cmd/controller/update_events_wiring_test.go
index 51a3cd5..784bdc5 100644
--- a/controller/cmd/controller/update_events_wiring_test.go
+++ b/controller/cmd/controller/update_events_wiring_test.go
@@ -1,6 +1,10 @@
 package main
 
-import "testing"
+import (
+	"os"
+	"strings"
+	"testing"
+)
 
 // v0.264.0. The update-event sink and the R-646 backfill are both CALLED from main.go (an AST walk —
 // a comment naming them would not count). A sink built and never wired would fail silently: no event,
@@ -21,3 +25,16 @@ func TestUpdateEventSinkIsWiredAtStartup(t *testing.T) {
 		t.Error("BackfillAppliedMeta (R-646) must be called, after AdoptPins")
 	}
 }
+
+// R-647 (2) — the held event carries the copy_holds KEY, never the Hungarian phrase: the hub prints the
+// raw details as the household mail's `Note:` line. COMPANION RED-PROOF (REPORT.md): wire
+// UpdateCopyHolds back — this fails.
+func TestR647_HeldEventCarriesTheCopyHoldsKey(t *testing.T) {
+	src, err := os.ReadFile("main.go")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(src), "d.CopyHolds = backupMgr.UpdateCopyHoldsKey(") {
+		t.Fatal("main.go must fill AppUpdateDetails.CopyHolds with UpdateCopyHoldsKey (the key), not the Hungarian phrase")
+	}
+}
diff --git a/controller/internal/api/router.go b/controller/internal/api/router.go
index 4414c77..e5a6593 100644
--- a/controller/internal/api/router.go
+++ b/controller/internal/api/router.go
@@ -19,7 +19,6 @@ import (
 	"gitea.dooplex.hu/admin/felhom-controller/internal/backup"
 	cf "gitea.dooplex.hu/admin/felhom-controller/internal/cloudflare"
 	"gitea.dooplex.hu/admin/felhom-controller/internal/config"
-	"gitea.dooplex.hu/admin/felhom-controller/internal/i18n"
 	"gitea.dooplex.hu/admin/felhom-controller/internal/integrations"
 	"gitea.dooplex.hu/admin/felhom-controller/internal/metrics"
 	"gitea.dooplex.hu/admin/felhom-controller/internal/notify"
@@ -383,18 +382,12 @@ func (r *Router) getStack(w http.ResponseWriter, req *http.Request, name string)
 	// v0.264.0 (R-606): the page polls this for the Update button's label and the outcome. Rendered in
 	// the reader's language; a Hungarian reader gets exactly the bytes it always got. GetStack returns a
 	// COPY, so nothing here touches the manager's own state.
-	if lang := r.langFor(req); lang != i18n.Default {
-		stack.UpdatePhaseLabel = stacks.UpdatePhaseLabelIn(lang, stack.UpdatePhase)
-		stack.UpdateError = stack.UpdateErrorIn(lang)
-		if stack.HoldReason != "" && r.backupMgr != nil {
-			if held, why := r.backupMgr.RestoreHoldForLang(name, lang); held && why != "" {
-				stack.HoldReason = why
-				if stack.UpdateErrorKey == "" && stack.UpdatePhase == stacks.UpdatePhaseFailed {
-					stack.UpdateError = why // a held update's UpdateError IS the hold sentence
-				}
-			}
-		}
-	}
+	// v0.265.0 (R-647): EVERY reader, Hungarian included — a Hungarian reader of an English box read
+	// the hold in English before. On a Hungarian box a Hungarian reader still gets the same bytes.
+	lang := r.langFor(req)
+	stack.UpdatePhaseLabel = stacks.UpdatePhaseLabelIn(lang, stack.UpdatePhase)
+	stack.UpdateError = r.stackMgr.UpdateErrorFor(*stack, lang)
+	stack.HoldReason = r.stackMgr.HoldReasonFor(*stack, lang)
 	writeJSON(w, http.StatusOK, apiResponse{OK: true, Data: stack})
 }
 
diff --git a/controller/internal/backup/backup.go b/controller/internal/backup/backup.go
index 4b9fd7d..e51e4f3 100644
--- a/controller/internal/backup/backup.go
+++ b/controller/internal/backup/backup.go
@@ -724,6 +724,16 @@ func (m *Manager) runVolumeDumps() (summary []string, dumped int, allOK bool) {
 			continue
 		}
 
+		// R-634 (v0.265.0): an app whose deploy is still running is NOT an installation to back up,
+		// and stopping it is what broke the deploy. The list is taken once at the start of the run, so
+		// the question is asked again here, immediately before the stop. A SKIP, not a failure: there
+		// is nothing of the household's in it yet, and a FAIL would page the operator for a deploy.
+		if m.stackIsDeploying(stack.Name) {
+			m.logger.Printf("[INFO] [backup] Skipping volume dump for %s — the app is still being deployed (R-634)", stack.Name)
+			summary = append(summary, fmt.Sprintf("SKIP %s volumes (deploying)", stack.Name))
+			continue
+		}
+
 		// R-181: the reserve, ahead of DumpAppVolumesSafe so a refused app is NOT stopped. For an app
 		// that already has a DB this is a memo lookup taken before its DB dump; for a volume-only app
 		// this is where its verdict is taken, still before its first byte.
@@ -1398,3 +1408,17 @@ func (m *Manager) note(key string, args ...interface{}) string {
 // noteErr renders an error into a saved note in the box's language: its bundle message when it
 // carries one (release B), its own text otherwise.
 func (m *Manager) noteErr(err error) string { return util.ErrText(m.boxLang(), err) }
+
+// deployingReporter is the OPTIONAL half of the stack provider that can say a deploy is in flight
+// (R-634). Optional on purpose: the production adapter implements it, and the many test fakes that
+// predate it need not — a provider that cannot answer is read as "not deploying", today's behaviour.
+type deployingReporter interface {
+	IsDeploying(name string) bool
+}
+
+func (m *Manager) stackIsDeploying(name string) bool {
+	if d, ok := m.stackProvider.(deployingReporter); ok {
+		return d.IsDeploying(name)
+	}
+	return false
+}
diff --git a/controller/internal/backup/r634_deploying_test.go b/controller/internal/backup/r634_deploying_test.go
new file mode 100644
index 0000000..e359b57
--- /dev/null
+++ b/controller/internal/backup/r634_deploying_test.go
@@ -0,0 +1,61 @@
+package backup
+
+import (
+	"io"
+	"log"
+	"path/filepath"
+	"testing"
+
+	"gitea.dooplex.hu/admin/felhom-controller/internal/config"
+)
+
+// R-634 (v0.265.0). Measured 2026-09-23 on 9202: a whole-box backup pressed 20 s into outline's
+// deploy stopped it (`compose down`), dumped its half-made volumes and ran a SECOND `compose up -d`
+// beside the deploy's own — both failed, and the deploy recorded „not deployed".
+
+type deployingFake struct {
+	volDumpFakeProvider
+	deploying map[string]bool
+}
+
+func (f *deployingFake) IsDeploying(name string) bool { return f.deploying[name] }
+
+// Through the dumpVolumesSafe SEAM, never the real dump: the real one runs `docker run … tar` against a
+// named volume, which on a developer machine CREATES that volume (it did, once, on DooPlex while this
+// test was written). The seam is where the stop happens, so a call to it IS the leak.
+//
+// COMPANION RED-PROOF (REPORT.md): delete the stackIsDeploying block in runVolumeDumps — the dump
+// (and with it the stop) is called for outline and this fails with dumped=[outline].
+func TestR634_VolumeLegNeverStopsADeployingApp(t *testing.T) {
+	cfg := &config.Config{}
+	cfg.Paths.SystemDataPath = filepath.Join(t.TempDir(), "sys")
+	fake := &deployingFake{
+		volDumpFakeProvider: volDumpFakeProvider{
+			stacks:  []StackSummary{{Name: "outline"}},
+			volumes: map[string][]string{"outline": {"outline_outline_data"}},
+		},
+		deploying: map[string]bool{"outline": true},
+	}
+	m := &Manager{cfg: cfg, logger: log.New(io.Discard, "", 0), systemDataPath: cfg.Paths.SystemDataPath,
+		stackProvider: fake}
+	var calls []string
+	m.dumpVolumesSafe = func(name string) error { calls = append(calls, name); return nil }
+
+	summary, dumped, ok := m.runVolumeDumps()
+	if len(calls) != 0 {
+		t.Fatalf("a DEPLOYING app was stopped and dumped by the backup: dumped=%v (R-634's race)", calls)
+	}
+	if !ok || dumped != 0 {
+		t.Errorf("a deploy in flight is a SKIP, never a failure that pages the operator: ok=%v dumped=%d %v", ok, dumped, summary)
+	}
+	if !containsSummary(summary, "SKIP outline volumes (deploying)") {
+		t.Errorf("the skip must be visible in the run summary, got %v", summary)
+	}
+
+	// Control: the same app once its deploy has finished IS backed up (the skip is conditional).
+	fake.deploying["outline"] = false
+	m.runVolumeDumps()
+	if len(calls) != 1 || calls[0] != "outline" {
+		t.Errorf("control: a finished app must be dumped as before, dumped=%v", calls)
+	}
+}
diff --git a/controller/internal/backup/undo_hold_test.go b/controller/internal/backup/undo_hold_test.go
index cc20830..48f027a 100644
--- a/controller/internal/backup/undo_hold_test.go
+++ b/controller/internal/backup/undo_hold_test.go
@@ -7,6 +7,10 @@ import (
 	"strings"
 	"testing"
 	"time"
+
+	"gitea.dooplex.hu/admin/felhom-controller/internal/config"
+	"gitea.dooplex.hu/admin/felhom-controller/internal/i18n"
+	"gitea.dooplex.hu/admin/felhom-controller/internal/util"
 )
 
 // v0.263.0 — a hold after a FAILED UNDO opens with what was tried and what state the data is in, in the
@@ -65,3 +69,15 @@ func TestUndo_HoldSentenceSaysTheUndoWasTriedAndTheDataState(t *testing.T) {
 		t.Errorf("no Hungarian may remain on an English box, got %q", why)
 	}
 }
+
+// R-647 (2) — the key form of the copy verdict is a bundle KEY, the same verdict as the phrase.
+func TestR647_UpdateCopyHoldsKeyIsTheKeyOfThePhrase(t *testing.T) {
+	m := NewManager(&config.Config{}, nil, log.New(io.Discard, "", 0))
+	k := m.UpdateCopyHoldsKey("x", UpdateTierLocal)
+	if !strings.HasPrefix(k, "hold.copy_holds.") {
+		t.Fatalf("want a hold.copy_holds.* key, got %q", k)
+	}
+	if util.Text(i18n.Default, k) != m.UpdateCopyHolds("x", UpdateTierLocal) {
+		t.Fatal("the key must render to exactly the phrase UpdateCopyHolds gives")
+	}
+}
diff --git a/controller/internal/backup/update_guard.go b/controller/internal/backup/update_guard.go
index 82d3b85..5e965ac 100644
--- a/controller/internal/backup/update_guard.go
+++ b/controller/internal/backup/update_guard.go
@@ -152,6 +152,13 @@ func (m *Manager) UpdateCopyHolds(stackName string, tier int) string {
 	return util.Text(i18n.Default, updateCopyHoldsKey(m.DataOutsideUnit(stackName), tier))
 }
 
+// UpdateCopyHoldsKey is the same verdict as UpdateCopyHolds, as its bundle KEY (R-647, v0.265.0). The
+// app_update_held event carries the key, not the Hungarian phrase: the hub prints the raw details as
+// the mail's `Note:` line, and an English household read „a beállításokat, …" there.
+func (m *Manager) UpdateCopyHoldsKey(stackName string, tier int) string {
+	return updateCopyHoldsKey(m.DataOutsideUnit(stackName), tier)
+}
+
 func updateCopyHoldsKey(outside bool, tier int) string {
 	switch tier {
 	case UpdateTierLocal:
diff --git a/controller/internal/i18n/locales/en.json b/controller/internal/i18n/locales/en.json
index 3dc2b40..461f8a4 100644
--- a/controller/internal/i18n/locales/en.json
+++ b/controller/internal/i18n/locales/en.json
@@ -571,6 +571,7 @@
   "badge.update.behind.today": " — today",
   "badge.update.current": "Up to date",
   "badge.update.current.title": "This app is running the newest version available.",
+  "badge.update.held": "Stopped — restore needed",
   "catchall.alkalmazas_kezelese": "Manage app",
   "catchall.alkalmazasok": "Apps",
   "catchall.vezerlopult": "Dashboard",
diff --git a/controller/internal/i18n/locales/hu.json b/controller/internal/i18n/locales/hu.json
index f50549f..8e3cd62 100644
--- a/controller/internal/i18n/locales/hu.json
+++ b/controller/internal/i18n/locales/hu.json
@@ -566,6 +566,7 @@
   "badge.update.behind.today": " — ma",
   "badge.update.current": "Naprakész",
   "badge.update.current.title": "Ez az alkalmazás a legfrissebb elérhető változatot futtatja.",
+  "badge.update.held": "Megállítva — visszaállítás szükséges",
   "catchall.alkalmazas_kezelese": "Alkalmazás kezelése",
   "catchall.alkalmazasok": "Alkalmazások",
   "catchall.vezerlopult": "Vezérlőpult",
diff --git a/controller/internal/notify/notifier.go b/controller/internal/notify/notifier.go
index a1ee98c..68034a2 100644
--- a/controller/internal/notify/notifier.go
+++ b/controller/internal/notify/notifier.go
@@ -46,6 +46,11 @@ type Notifier struct {
 
 	// oomSeen (R-514) remembers container runs already reported as OOM-killed.
 	oomSeen map[string]bool
+	// oomTrack (R-636) remembers, per container run, the kill counter over the last 30 minutes and
+	// whether the ONE storm alarm of that run has gone.
+	oomTrack map[string]*oomRun
+	// nowFn is a test clock (nil → time.Now).
+	nowFn func() time.Time
 	// appDown tracks which deployed apps are currently in the DOWN state so app_start_failed fires
 	// ONCE per running→down transition, not every health cycle (fix-3 anti-spam). In-memory: a
 	// controller restart re-notifies once (acceptable — better than missing). The hub owns the real
@@ -359,7 +364,7 @@ func (n *Notifier) pushEventBoth(eventType, severity, message, messageCustomer s
 // Detects both degradation (ok→warn, ok→fail, warn→fail) and recovery (fail→ok, warn→ok, fail→warn).
 func (n *Notifier) NotifyHealthChange(status string, issues, warnings []string) {
 	if !n.enabled {
-		n.dropped("health_change", status)
+		n.dropped("health_change", healthSeverity(status)) // R-647: the severity, not the health status
 		return
 	}
 
@@ -714,21 +719,108 @@ func (n *Notifier) NotifyAppStartFailures(apps []AppRunState) {
 // limit (Docker State.OOMKilled). Fires ONCE per container run (keyed by StartedAt), so a container
 // that stays OOM-marked does not repeat. "warning" — the hub vocabulary (R-329). Operator-only
 // hub-side (hub >= v0.114.0 registers it): the household sees the dashboard tag.
-func (n *Notifier) NotifyAppOOM(stack, container, startedAt string) {
+func (n *Notifier) NotifyAppOOM(stack, container, startedAt string, kills int64, memLimit, peak string) {
 	key := container + "|" + startedAt
 	n.mu.Lock()
 	if n.oomSeen == nil {
 		n.oomSeen = map[string]bool{}
 	}
-	if n.oomSeen[key] {
-		n.mu.Unlock()
-		return
-	}
+	first := !n.oomSeen[key]
 	n.oomSeen[key] = true
+	storm, inWindow := n.oomStormLocked(key, startedAt, kills)
 	n.mu.Unlock()
-	n.emit("app_oom", "warning",
-		fmt.Sprintf("Alkalmazás memóriája elfogyott: %s (%s) — egy folyamatát a memóriakorlát leállította", stack, container),
-		AppDetails{StackName: stack, DisplayName: container})
+	if first {
+		n.emit("app_oom", "warning",
+			fmt.Sprintf("Alkalmazás memóriája elfogyott: %s (%s) — egy folyamatát a memóriakorlát leállította", stack, container),
+			AppDetails{StackName: stack, DisplayName: container})
+	}
+	if storm {
+		n.logger.Printf("[ERROR] [notify] %s: container %s OOM STORM — %d kills in %d min (limit %s, peak %s)",
+			stack, container, inWindow, oomStormWindowMin, memLimit, peak)
+		n.emit("app_oom_storm", "error",
+			fmt.Sprintf("Alkalmazás memóriája ismételten elfogy: %s (%s) — %d leállítás %d percen belül (korlát %s, csúcs %s)",
+				stack, container, inWindow, oomStormWindowMin, memLimit, peak),
+			OOMStormDetails{App: stack, StackName: stack, Container: container, Kills: inWindow,
+				WindowMin: oomStormWindowMin, MemLimit: memLimit, Peak: peak})
+	}
+}
+
+// R-636 (v0.265.0) — ONE louder alarm per container run when OOM kills keep coming.
+//
+// The once-per-run app_oom stays exactly as it was (it is what stops a crash loop from mailing 4,530
+// times, R-629). Beside it: when the kernel's own kill counter for the SAME run rises by
+// oomStormKills or more within oomStormWindowMin minutes, app_oom_storm goes once, at `error`,
+// operator-only. RomM's real rate on 2026-09-22 was 4,530 kills in 6 h ≈ 375 per 30 min; one hiccup
+// is 1–3. Twenty in thirty minutes sits far from both. Pinned by TestR636_*.
+const (
+	oomStormKills     = 20
+	oomStormWindowMin = 30
+)
+
+type oomSample struct {
+	at    time.Time
+	kills int64
+}
+
+type oomRun struct {
+	samples   []oomSample
+	stormSent bool
+}
+
+// OOMStormDetails is the app_oom_storm payload. stack_name lets the hub's per-app cooldown tell two
+// storming apps apart.
+type OOMStormDetails struct {
+	App       string `json:"app"`
+	StackName string `json:"stack_name"`
+	Container string `json:"container"`
+	Kills     int64  `json:"kills"`
+	WindowMin int    `json:"window_min"`
+	MemLimit  string `json:"mem_limit,omitempty"`
+	Peak      string `json:"peak,omitempty"`
+}
+
+func (n *Notifier) now() time.Time {
+	if n.nowFn != nil {
+		return n.nowFn()
+	}
+	return time.Now()
+}
+
+// oomStormLocked records one reading and says whether the storm alarm fires now. Caller holds n.mu.
+// A run seen for the first time within the window of its own start gets a zero reading AT its start:
+// the counter began at zero there, so a run that is already storming when the controller first looks
+// is not made to wait another thirty minutes.
+func (n *Notifier) oomStormLocked(key, startedAt string, kills int64) (bool, int64) {
+	if kills < 0 {
+		return false, 0 // unreadable: never escalate on a guess
+	}
+	if n.oomTrack == nil {
+		n.oomTrack = map[string]*oomRun{}
+	}
+	now := n.now()
+	window := time.Duration(oomStormWindowMin) * time.Minute
+	r := n.oomTrack[key]
+	if r == nil {
+		r = &oomRun{}
+		if t, err := time.Parse(time.RFC3339Nano, startedAt); err == nil && now.Sub(t) <= window {
+			r.samples = append(r.samples, oomSample{at: t, kills: 0})
+		}
+		n.oomTrack[key] = r
+	}
+	r.samples = append(r.samples, oomSample{at: now, kills: kills})
+	keep := r.samples[:0]
+	for _, s := range r.samples {
+		if now.Sub(s.at) <= window {
+			keep = append(keep, s)
+		}
+	}
+	r.samples = keep
+	inWindow := kills - r.samples[0].kills
+	if r.stormSent || inWindow < oomStormKills {
+		return false, inWindow
+	}
+	r.stormSent = true
+	return true, inWindow
 }
 
 // DiskHealthDetails is the event-detail payload for disk_health_degraded.
@@ -1211,3 +1303,15 @@ func (n *Notifier) NotifyAppUpdateHeld(d AppUpdateDetails, sentence func(lang st
 	}
 	n.pushEventBoth("app_update_held", "error", hu, household, d)
 }
+
+// healthSeverity is the event severity a health status would be sent at (R-647): the disabled path
+// names what it drops, and "warn" is a health STATUS, not a severity the hub knows.
+func healthSeverity(status string) string {
+	switch status {
+	case "fail":
+		return "error"
+	case "warn":
+		return "warning"
+	}
+	return "info"
+}
diff --git a/controller/internal/notify/r636_oom_storm_test.go b/controller/internal/notify/r636_oom_storm_test.go
new file mode 100644
index 0000000..bd83277
--- /dev/null
+++ b/controller/internal/notify/r636_oom_storm_test.go
@@ -0,0 +1,95 @@
+package notify
+
+import (
+	"io"
+	"log"
+	"testing"
+	"time"
+)
+
+// R-636 — a repeating OOM problem gets ONE louder alarm; one hiccup stays one quiet warning.
+func stormRecorder(t *testing.T) (*Notifier, *time.Time, map[string]int) {
+	t.Helper()
+	n := New("", "", "c1", nil, log.New(io.Discard, "", 0), false)
+	got := map[string]int{}
+	n.pushFn = func(eventType, severity, msg, msgCustomer string, details interface{}) {
+		got[eventType+"/"+severity]++
+	}
+	clock := time.Date(2026, 9, 22, 9, 0, 0, 0, time.UTC)
+	n.nowFn = func() time.Time { return clock }
+	return n, &clock, got
+}
+
+// COMPANION RED-PROOF (REPORT.md): make oomStormLocked always return false — the 20-kill case sends no
+// storm and this fails.
+func TestR636_TwentyKillsInThirtyMinutesIsOneStorm(t *testing.T) {
+	for _, c := range []struct {
+		kills int64
+		storm int
+	}{{19, 0}, {20, 1}, {200, 1}} {
+		n, clock, got := stormRecorder(t)
+		start := clock.Add(-time.Hour).Format(time.RFC3339Nano) // started long before: no zero baseline
+		for i := int64(0); i <= 40; i++ { // 20 minutes of 30 s scans, counter 5 → 5+kills
+			n.NotifyAppOOM("romm", "romm", start, 5+c.kills*i/40, "1280M", "1279M")
+			*clock = clock.Add(30 * time.Second)
+		}
+		if got["app_oom_storm/error"] != c.storm {
+			t.Errorf("kills=%d in 20 min: storms=%d, want %d", c.kills, got["app_oom_storm/error"], c.storm)
+		}
+		if c.kills == 200 { // RomM's shape: it keeps going for six hours — still ONE storm for the run
+			for i := int64(1); i <= 720; i++ {
+				n.NotifyAppOOM("romm", "romm", start, 205+i*12, "1280M", "1279M")
+				*clock = clock.Add(30 * time.Second)
+			}
+			if got["app_oom_storm/error"] != 1 {
+				t.Errorf("six hours of storm must stay ONE storm per container run, got %d", got["app_oom_storm/error"])
+			}
+		}
+		if got["app_oom/warning"] != 1 {
+			t.Errorf("kills=%d: the quiet app_oom must still go exactly once per run, got %d", c.kills, got["app_oom/warning"])
+		}
+	}
+}
+
+// One kill, and then the flag just stays set for hours (the sticky OOMKilled): never a storm.
+func TestR636_OneHiccupNeverStorms(t *testing.T) {
+	n, clock, got := stormRecorder(t)
+	start := clock.Format(time.RFC3339Nano)
+	for i := 0; i < 6*120; i++ { // six hours of 30 s scans, counter stuck at 1
+		n.NotifyAppOOM("paperless-ngx", "paperless-webserver", start, 1, "768M", "700M")
+		*clock = clock.Add(30 * time.Second)
+	}
+	if got["app_oom_storm/error"] != 0 || got["app_oom/warning"] != 1 {
+		t.Fatalf("one hiccup: want 1 warning and 0 storms, got %v", got)
+	}
+}
+
+// Slow and steady — 20 kills spread over three hours is not a storm.
+func TestR636_SlowKillsAreNotAStorm(t *testing.T) {
+	n, clock, got := stormRecorder(t)
+	start := clock.Add(-time.Hour).Format(time.RFC3339Nano)
+	for i := int64(0); i <= 360; i++ { // 3 h, counter +20 in total
+		n.NotifyAppOOM("x", "x", start, i*20/360, "", "")
+		*clock = clock.Add(30 * time.Second)
+	}
+	if got["app_oom_storm/error"] != 0 {
+		t.Fatalf("20 kills over 3 h must not storm, got %v", got)
+	}
+}
+
+// A run first seen already storming (the controller restarted mid-storm) fires at once: its counter
+// started at zero at its own start, inside the window. An unreadable counter never escalates.
+func TestR636_AlreadyStormingAtFirstSightAndUnreadable(t *testing.T) {
+	n, clock, got := stormRecorder(t)
+	n.NotifyAppOOM("romm", "romm", clock.Add(-10*time.Minute).Format(time.RFC3339Nano), 150, "1280M", "1279M")
+	if got["app_oom_storm/error"] != 1 {
+		t.Fatalf("150 kills 10 min after the start must storm at first sight, got %v", got)
+	}
+	n2, _, got2 := stormRecorder(t)
+	for i := 0; i < 100; i++ {
+		n2.NotifyAppOOM("romm", "romm", "", -1, "", "")
+	}
+	if got2["app_oom_storm/error"] != 0 {
+		t.Fatalf("an unreadable counter must never escalate, got %v", got2)
+	}
+}
diff --git a/controller/internal/notify/update_events_test.go b/controller/internal/notify/update_events_test.go
index 4f6a095..06c9d19 100644
--- a/controller/internal/notify/update_events_test.go
+++ b/controller/internal/notify/update_events_test.go
@@ -102,3 +102,15 @@ func TestR620_DisabledNotifierSaysWhatItDrops(t *testing.T) {
 		}
 	}
 }
+
+// R-647 (3) — the disabled path names the SEVERITY a health change would be sent at, never the health
+// status. COMPANION RED-PROOF (REPORT.md): pass `status` again — the line reads "severity warn".
+func TestR647_DisabledHealthChangeNamesTheSeverity(t *testing.T) {
+	var buf bytes.Buffer
+	n := New("", "", "c1", nil, log.New(&buf, "", 0), false)
+	buf.Reset()
+	n.NotifyHealthChange("warn", nil, nil)
+	if !strings.Contains(buf.String(), "DROPPED event health_change (severity warning)") {
+		t.Fatalf("want the severity 'warning' named, got:\n%s", buf.String())
+	}
+}
diff --git a/controller/internal/stacks/deploy_errors.go b/controller/internal/stacks/deploy_errors.go
index 1ead34f..fe7a8a1 100644
--- a/controller/internal/stacks/deploy_errors.go
+++ b/controller/internal/stacks/deploy_errors.go
@@ -22,6 +22,11 @@ var (
 	ErrPathMissing = errors.New("path field does not exist")
 	// ErrNotEnoughMemory — the memory verdict refused the deploy (API: 400).
 	ErrNotEnoughMemory = errors.New("not enough memory")
+	// ErrStackDeploying — R-634 (v0.265.0): a stop or start was asked of a stack whose deploy is still
+	// running. Refused, because a `compose down` or a second `compose up -d` in the middle of the
+	// deploy's own `up` makes BOTH fail, and the deploy then records „not deployed" over whatever
+	// containers the race left behind. Pinned by TestR634_StopAndStartRefuseADeployingStack.
+	ErrStackDeploying = errors.New("stack is being deployed")
 )
 
 // msgHU renders a bundle message in Hungarian.
diff --git a/controller/internal/stacks/manager.go b/controller/internal/stacks/manager.go
index 2fd53f7..c94c905 100644
--- a/controller/internal/stacks/manager.go
+++ b/controller/internal/stacks/manager.go
@@ -1147,6 +1147,12 @@ func (m *Manager) StartStack(name string) error {
 	if !ok {
 		return fmt.Errorf("stack %q not found", name)
 	}
+	if stack.Deploying {
+		// R-634: a second `compose up -d` beside the deploy's own is the race that left apps running
+		// under „not deployed". The deploy brings the app up itself; nothing else may.
+		m.logger.Printf("[WARN] [stacks] StartStack %s refused: the app's deploy is still running (R-634)", name)
+		return fmt.Errorf("starting stack %s: %w", name, ErrStackDeploying)
+	}
 
 	if m.isDebug() {
 		m.logger.Printf("[DEBUG] [stacks] StartStack %s: current state=%s deployed=%v", name, stack.State, stack.Deployed)
@@ -1229,6 +1235,13 @@ func (m *Manager) StopStack(name string) error {
 	if !ok {
 		return fmt.Errorf("stack %q not found", name)
 	}
+	if stack.Deploying {
+		// R-634, the backstop for EVERY caller (backup, quiesce, restore, export, storage, the Stop
+		// button): `compose down` in the middle of a deploy was measured to make the deploy fail and
+		// record „not deployed" while the caller's own restart brought the containers back.
+		m.logger.Printf("[WARN] [stacks] StopStack %s refused: the app's deploy is still running (R-634)", name)
+		return fmt.Errorf("stopping stack %s: %w", name, ErrStackDeploying)
+	}
 
 	if m.isDebug() {
 		m.logger.Printf("[DEBUG] [stacks] StopStack %s: current state=%s deployed=%v containers=%d", name, stack.State, stack.Deployed, len(stack.Containers))
diff --git a/controller/internal/stacks/oom.go b/controller/internal/stacks/oom.go
index 59910be..d56a6c5 100644
--- a/controller/internal/stacks/oom.go
+++ b/controller/internal/stacks/oom.go
@@ -1,7 +1,9 @@
 package stacks
 
 import (
+	"fmt"
 	"sort"
+	"strconv"
 	"strings"
 )
 
@@ -18,6 +20,13 @@ type OOMContainer struct {
 	Stack     string
 	Container string
 	StartedAt string // identifies the container run — one event per run
+	// Kills is the kernel's own count of OOM kills in this container's cgroup (`memory.events`
+	// oom_kill) — R-636, v0.265.0. OOMKilled is a STICKY flag: it stays true for the container's whole
+	// life after ONE kill, so "the key re-fired" means nothing; this counter is what separates one
+	// hiccup from RomM's 4,530 kills in six hours. -1 when it could not be read.
+	Kills int64
+	// MemLimit / Peak are the cgroup's memory.max and memory.peak, as "<n>M" (or "max"); "" unread.
+	MemLimit, Peak string
 }
 
 // ScanOOMKilled inspects the containers of every deployed, running-ish stack in ONE docker call and
@@ -56,13 +65,53 @@ func (m *Manager) ScanOOMKilled() ([]OOMContainer, error) {
 		}
 		cname := strings.TrimPrefix(f[0], "/")
 		if st, ok := owner[cname]; ok {
-			found = append(found, OOMContainer{Stack: st, Container: cname, StartedAt: f[2]})
+			found = append(found, OOMContainer{Stack: st, Container: cname, StartedAt: f[2], Kills: -1})
 		}
 	}
+	// R-636: only for the containers already flagged — one `docker exec` each, and the flagged set is
+	// empty on a healthy box. Read from INSIDE the container (its cgroup namespace makes
+	// /sys/fs/cgroup its own cgroup); the controller's own namespace cannot see the others.
+	for i := range found {
+		out, _ := m.execCommand("docker", "exec", found[i].Container, "cat",
+			"/sys/fs/cgroup/memory.events", "/sys/fs/cgroup/memory.max", "/sys/fs/cgroup/memory.peak")
+		found[i].Kills, found[i].MemLimit, found[i].Peak = parseCgroupMemory(out)
+	}
 	m.setOOMCache(found)
 	return found, nil
 }
 
+// parseCgroupMemory reads `cat memory.events memory.max memory.peak`: the key/value lines give
+// oom_kill; the first bare line is memory.max, the second memory.peak (absent on older kernels — cat
+// then fails on that file and still prints the others). Unreadable → Kills -1.
+func parseCgroupMemory(out string) (kills int64, limit, peak string) {
+	kills = -1
+	var bare []string
+	for _, line := range strings.Split(strings.TrimSpace(out), "\n") {
+		f := strings.Fields(line)
+		switch {
+		case len(f) == 2 && f[0] == "oom_kill":
+			if n, err := strconv.ParseInt(f[1], 10, 64); err == nil {
+				kills = n
+			}
+		case len(f) == 1:
+			bare = append(bare, f[0])
+		}
+	}
+	mb := func(v string) string {
+		if n, err := strconv.ParseInt(v, 10, 64); err == nil {
+			return fmt.Sprintf("%dM", n/(1024*1024))
+		}
+		return v // "max"
+	}
+	if len(bare) > 0 {
+		limit = mb(bare[0])
+	}
+	if len(bare) > 1 {
+		peak = mb(bare[1])
+	}
+	return kills, limit, peak
+}
+
 func (m *Manager) setOOMCache(found []OOMContainer) {
 	cache := map[string][]string{}
 	for _, o := range found {
diff --git a/controller/internal/stacks/oom_test.go b/controller/internal/stacks/oom_test.go
index d6e2770..adab645 100644
--- a/controller/internal/stacks/oom_test.go
+++ b/controller/internal/stacks/oom_test.go
@@ -53,3 +53,46 @@ func TestScanOOMKilled_SeesKilledWorkerInRunningContainer(t *testing.T) {
 		t.Fatalf("cache not cleared after a clean scan: %v", got)
 	}
 }
+
+// R-636 — the kernel's kill counter, the limit and the peak are read from `cat memory.events
+// memory.max memory.peak` inside the container; an unreadable one is -1, never 0.
+func TestR636_ParseCgroupMemory(t *testing.T) {
+	out := "low 0\nhigh 0\nmax 4521\noom 4530\noom_kill 4530\noom_group_kill 0\n1342177280\n1341128704\n"
+	k, lim, peak := parseCgroupMemory(out)
+	if k != 4530 || lim != "1280M" || peak != "1279M" {
+		t.Fatalf("got kills=%d limit=%q peak=%q", k, lim, peak)
+	}
+	if k, lim, _ := parseCgroupMemory("oom_kill 3\nmax\n"); k != 3 || lim != "max" {
+		t.Fatalf("no memory.peak (older kernel), unlimited: got %d %q", k, lim)
+	}
+	if k, _, _ := parseCgroupMemory("OCI runtime exec failed: exec: \"cat\": executable file not found"); k != -1 {
+		t.Fatalf("an image without cat must read -1 (unknown), got %d", k)
+	}
+}
+
+// R-636 — the scan asks the kill counter of the FLAGGED containers only (one exec each; none on a
+// healthy box). COMPANION RED-PROOF (REPORT.md): drop the exec loop — Kills stays -1 and this fails.
+func TestR636_ScanReadsTheCounterOfFlaggedContainersOnly(t *testing.T) {
+	var execd []string
+	m := &Manager{logger: log.New(io.Discard, "", 0), stacks: map[string]*Stack{
+		"romm":      {Name: "romm", Deployed: true, Containers: []ContainerInfo{{Name: "romm", State: StateRunning}}},
+		"bookstack": {Name: "bookstack", Deployed: true, Containers: []ContainerInfo{{Name: "bookstack", State: StateRunning}}},
+	}}
+	m.execFn = func(name string, args ...string) (string, error) {
+		if args[0] == "exec" {
+			execd = append(execd, args[1])
+			return "oom_kill 377\n1342177280\n1341128704\n", nil
+		}
+		return "/bookstack|false|t0\n/romm|true|2026-09-22T09:08:00Z\n", nil
+	}
+	ooms, err := m.ScanOOMKilled()
+	if err != nil || len(ooms) != 1 {
+		t.Fatalf("ooms=%+v err=%v", ooms, err)
+	}
+	if ooms[0].Kills != 377 || ooms[0].MemLimit != "1280M" || ooms[0].Peak != "1279M" {
+		t.Errorf("the counter must be read for the flagged container: %+v", ooms[0])
+	}
+	if len(execd) != 1 || execd[0] != "romm" {
+		t.Errorf("exec must reach the flagged container only, got %v", execd)
+	}
+}
diff --git a/controller/internal/stacks/r634_deploying_test.go b/controller/internal/stacks/r634_deploying_test.go
new file mode 100644
index 0000000..85b5c38
--- /dev/null
+++ b/controller/internal/stacks/r634_deploying_test.go
@@ -0,0 +1,36 @@
+package stacks
+
+import (
+	"errors"
+	"testing"
+)
+
+// R-634 (v0.265.0) — the backstop for every caller: a stop or a start while the deploy's own
+// `compose up` is running is refused BEFORE any compose command.
+//
+// COMPANION RED-PROOF (REPORT.md): delete the Deploying check in StopStack — the stop falls through to
+// `compose down` and the error is no longer ErrStackDeploying.
+func TestR634_StopAndStartRefuseADeployingStack(t *testing.T) {
+	m, _, _, _ := newSlice4Manager(t)
+	// NEVER a real compose: the control below falls through to `compose down`, and the stack dir is
+	// named like a real app. An empty PATH + the v1 binary name guarantees the exec cannot find
+	// anything to run, on any machine this test runs on.
+	m.composeCmd = "docker-compose"
+	t.Setenv("PATH", t.TempDir())
+	m.mu.Lock()
+	m.stacks["nextcloud"].Deploying = true
+	m.mu.Unlock()
+	if err := m.StopStack("nextcloud"); !errors.Is(err, ErrStackDeploying) {
+		t.Fatalf("StopStack on a deploying stack = %v, want ErrStackDeploying", err)
+	}
+	if err := m.StartStack("nextcloud"); !errors.Is(err, ErrStackDeploying) {
+		t.Fatalf("StartStack on a deploying stack = %v, want ErrStackDeploying", err)
+	}
+	// Control: once the deploy is over, the same calls are not refused for THIS reason.
+	m.mu.Lock()
+	m.stacks["nextcloud"].Deploying = false
+	m.mu.Unlock()
+	if err := m.StopStack("nextcloud"); errors.Is(err, ErrStackDeploying) {
+		t.Fatal("control: a stack that is not deploying must not be refused as deploying")
+	}
+}
diff --git a/controller/internal/stacks/undo.go b/controller/internal/stacks/undo.go
index 35be22e..b74a67c 100644
--- a/controller/internal/stacks/undo.go
+++ b/controller/internal/stacks/undo.go
@@ -471,7 +471,11 @@ func (m *Manager) emitUpdateEvent(kind, name string, entry *updateJournalEntry,
 	if entry != nil {
 		ev.From, ev.To = entry.PrevPin, entry.NewPin
 	}
-	m.logger.Printf("[INFO] [stacks] update %s: event %s (hold recorded: %v)", name, kind, holdRecorded)
+	if kind == UpdateEventHeld {
+		m.logger.Printf("[INFO] [stacks] update %s: event %s (hold recorded: %v)", name, kind, holdRecorded)
+	} else { // R-647: "hold recorded" means nothing for an undone update
+		m.logger.Printf("[INFO] [stacks] update %s: event %s", name, kind)
+	}
 	sink(ev)
 }
 
diff --git a/controller/internal/stacks/update.go b/controller/internal/stacks/update.go
index 9514cbd..8e45992 100644
--- a/controller/internal/stacks/update.go
+++ b/controller/internal/stacks/update.go
@@ -564,7 +564,7 @@ func (m *Manager) finishUpdate(name, phase, msg string) {
 // beside it for the page. key "" = `plain` is a finished sentence and is stored as it is.
 func (m *Manager) finishUpdateKey(name, phase, key, plain string, args ...interface{}) {
 	msg := plain
-	if key != "" {
+	if key != "" && key != UpdateErrorKeyHeld { // the held key names no bundle entry: plain IS the sentence
 		msg = util.Text(i18n.Default, key, args...)
 	}
 	m.mu.Lock()
@@ -590,12 +590,57 @@ func UpdatePhaseLabelIn(lang, phase string) string {
 
 // UpdateErrorIn is the stack's update sentence in lang (v0.264.0, R-606).
 func (s Stack) UpdateErrorIn(lang string) string {
-	if s.UpdateErrorKey == "" || lang == i18n.Default {
+	if s.UpdateErrorKey == "" || s.UpdateErrorKey == UpdateErrorKeyHeld || lang == i18n.Default {
 		return s.UpdateError
 	}
 	return util.Text(lang, s.UpdateErrorKey, s.UpdateErrorArgs...)
 }
 
+// UpdateErrorKeyHeld marks an UpdateError that IS the hold sentence (R-647, v0.265.0). It names no
+// bundle entry: the sentence is composed by the backup side, so UpdateErrorFor asks it for the
+// reader's language, and UpdateErrorIn (no manager to ask) returns the stored Hungarian.
+const UpdateErrorKeyHeld = "update.error.held"
+
+// holdLanguage is the OPTIONAL half of UpdateGuards that renders the hold sentence in a given
+// language (the production adapter implements it; the test fakes need not).
+type holdLanguage interface {
+	HoldForLang(name, lang string) (bool, string)
+}
+
+func holdForLang(g UpdateGuards, name, lang string) (bool, string) {
+	if g == nil {
+		return false, ""
+	}
+	if hl, ok := g.(holdLanguage); ok {
+		return hl.HoldForLang(name, lang)
+	}
+	return g.HoldFor(name)
+}
+
+// HoldReasonFor is a stack's hold sentence in the READER's language (R-647): "" when nothing holds it.
+// A Hungarian reader on a Hungarian box gets exactly Stack.HoldReason.
+func (m *Manager) HoldReasonFor(st Stack, lang string) string {
+	if st.HoldReason == "" {
+		return ""
+	}
+	if held, why := holdForLang(m.guards(), st.Name, lang); held && why != "" {
+		return why
+	}
+	return st.HoldReason
+}
+
+// UpdateErrorFor is a stack's update error in the READER's language (R-606 + R-647). A held update's
+// error is the hold sentence, rendered by the backup side for this reader.
+func (m *Manager) UpdateErrorFor(st Stack, lang string) string {
+	if st.UpdateErrorKey == UpdateErrorKeyHeld {
+		if held, why := holdForLang(m.guards(), st.Name, lang); held && why != "" {
+			return why
+		}
+		return st.UpdateError
+	}
+	return st.UpdateErrorIn(lang)
+}
+
 func (m *Manager) updateCompose(dir string, env []string, args ...string) (string, error) {
 	if m.updateComposeFn != nil {
 		return m.updateComposeFn(dir, env, args...)
@@ -880,7 +925,7 @@ func (m *Manager) failAndHold(ctx context.Context, name, dir string, env []strin
 		m.logger.Printf("[ERROR] [stacks] update %s: no UpdateGuards — the hold CANNOT be recorded", name)
 	} else if err := g.HoldAfterFailedUpdate(name, m.now(), rp, undoState); err != nil {
 		m.logger.Printf("[ERROR] [stacks] update %s: %v", name, err)
-	} else if _, w := g.HoldFor(name); w != "" {
+	} else if _, w := holdForLang(g, name, i18n.Default); w != "" {
 		holdWhy = w
 		m.markUpdateHeld(name)
 	}
@@ -890,8 +935,10 @@ func (m *Manager) failAndHold(ctx context.Context, name, dir string, env []strin
 	if holdWhy == "" {
 		m.finishUpdateKey(name, UpdatePhaseFailed, "update.error.hold_unsaved", "")
 	} else {
-		// The hold's own sentence (the page renders it per reader through RestoreHoldForLang).
-		m.finishUpdate(name, UpdatePhaseFailed, holdWhy)
+		// R-647 (v0.265.0): stored as the HELD key + the Hungarian sentence, so every reader — the
+		// `?lang=` switch, an operator reading a box in the other language — gets the hold sentence in
+		// THEIR language (UpdateErrorFor), and the stored text is Hungarian as for every other key.
+		m.finishUpdateKey(name, UpdatePhaseFailed, UpdateErrorKeyHeld, holdWhy)
 	}
 	m.emitUpdateEvent(UpdateEventHeld, name, entry, rp, holdWhy != "")
 }
diff --git a/controller/internal/web/i18n_web.go b/controller/internal/web/i18n_web.go
index 59f0d41..d523d41 100644
--- a/controller/internal/web/i18n_web.go
+++ b/controller/internal/web/i18n_web.go
@@ -322,17 +322,6 @@ func (s *Server) localeFuncs(lang string) template.FuncMap {
 			}
 			return stk.UpdatePhaseLabel // no translation for this phase: the stored label, never an empty line
 		},
-		"updateErrorText": func(st interface{}) string { return stackOf(st).UpdateErrorIn(lang) },
-		"holdText": func(st interface{}) string {
-			stk := stackOf(st)
-			if stk.HoldReason == "" || s.backupMgr == nil {
-				return stk.HoldReason
-			}
-			if held, why := s.backupMgr.RestoreHoldForLang(stk.Name, lang); held && why != "" {
-				return why
-			}
-			return stk.HoldReason
-		},
 		"timeAgo": func(t time.Time) string {
 			if t.IsZero() {
 				return "–"
@@ -387,6 +376,13 @@ func (s *Server) localeFuncs(lang string) template.FuncMap {
 		// Found live on 2026-09-20: on an English app page the only Felhom-authored Hungarian left
 		// was this badge, „Naprakész", with a Hungarian tooltip under it.
 		"updateBadge": func(st stacks.Stack) *MetaBadge {
+			if st.HoldReason != "" { // R-625's twin — see updatebadge.go
+				return &MetaBadge{
+					Label: b.Msg(lang, "badge.update.held"),
+					Class: "tag-error",
+					Title: firstSentence(st.HoldReason),
+				}
+			}
 			switch compareInstalledToTemplate(st) {
 			case updateCurrent:
 				return &MetaBadge{
@@ -554,3 +550,43 @@ func (s *Server) note(key string, args ...interface{}) string {
 // noteErr renders an error into a saved note in the box's language: its bundle message when it
 // carries one (release B), its own text otherwise.
 func (s *Server) noteErr(err error) string { return util.ErrText(s.boxLang(), err) }
+
+// readerFuncs are the funcs that need the SERVER to answer in the reader's language, and — unlike
+// localeFuncs — they apply to EVERY language, Hungarian included (R-647, v0.265.0). A hold sentence is
+// composed by the backup side in one language at a time; before this a Hungarian reader of an English
+// box (the `?lang=` switch, an operator) read it in English. On a Hungarian box a Hungarian reader gets
+// exactly the bytes templateFuncMap gave (the parity fixtures are rendered through here).
+// `prev` is the func set they wrap (templateFuncMap + localeFuncs for this language).
+func (s *Server) readerFuncs(lang string, prev template.FuncMap) template.FuncMap {
+	holdFor := func(st stacks.Stack) string {
+		if st.HoldReason == "" || s.backupMgr == nil {
+			return st.HoldReason
+		}
+		if held, why := s.backupMgr.RestoreHoldForLang(st.Name, lang); held && why != "" {
+			return why
+		}
+		return st.HoldReason
+	}
+	out := template.FuncMap{
+		"holdText": func(st interface{}) string { return holdFor(stackOf(st)) },
+		"updateErrorText": func(st interface{}) string {
+			stk := stackOf(st)
+			if stk.UpdateErrorKey == stacks.UpdateErrorKeyHeld && s.backupMgr != nil {
+				if held, why := s.backupMgr.RestoreHoldForLang(stk.Name, lang); held && why != "" {
+					return why
+				}
+			}
+			return stk.UpdateErrorIn(lang)
+		},
+	}
+	if base, ok := prev["updateBadge"].(func(stacks.Stack) *MetaBadge); ok {
+		out["updateBadge"] = func(st stacks.Stack) *MetaBadge {
+			bd := base(st)
+			if bd != nil && st.HoldReason != "" { // R-625: the title is the hold's opening, for THIS reader
+				bd.Title = firstSentence(holdFor(st))
+			}
+			return bd
+		}
+	}
+	return out
+}
diff --git a/controller/internal/web/r625_r647_held_test.go b/controller/internal/web/r625_r647_held_test.go
new file mode 100644
index 0000000..56e1b7f
--- /dev/null
+++ b/controller/internal/web/r625_r647_held_test.go
@@ -0,0 +1,136 @@
+package web
+
+import (
+	"html"
+	"io"
+	"log"
+	"strings"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-controller/internal/backup"
+	"gitea.dooplex.hu/admin/felhom-controller/internal/i18n"
+	"gitea.dooplex.hu/admin/felhom-controller/internal/stacks"
+)
+
+// heldBehindStack is R-625's measured shape: HELD by a failed update, while the catalog has already
+// published a newer version — the moment the page used to say „Frissítés elérhető".
+func heldBehindStack(t *testing.T, s *Server, boxLang string) (stacks.Stack, *backup.Manager) {
+	t.Helper()
+	b := backup.NewManager(s.cfg, s.settings, log.New(io.Discard, "", 0))
+	s.backupMgr = b
+	if boxLang != "" {
+		if err := s.settings.SetLanguage(boxLang); err != nil {
+			t.Fatal(err)
+		}
+	}
+	at := time.Date(2026, 9, 23, 8, 0, 0, 0, time.UTC)
+	if err := b.HoldAfterFailedUpdateHolding("kimai", at, at.Add(-6*time.Hour), backup.UpdateTierLocal, b.UpdateCopyHolds("kimai", backup.UpdateTierLocal), ""); err != nil {
+		t.Fatal(err)
+	}
+	_, why := b.RestoreHoldFor("kimai") // in the BOX's language, as fillHoldReason stores it
+	st := i18nStack("kimai", "Kimai Time", stacks.StateStopped, true)
+	st.AppConfig = &stacks.AppConfig{Deployed: true, InstalledImages: map[string]stacks.InstalledImage{"web": {Ref: "kimai/kimai2:2.30.0"}}}
+	st.CatalogImages = map[string]string{"web": "kimai/kimai2:2.31.0"}
+	if compareInstalledToTemplate(st) != updateBehind {
+		t.Fatal("fixture: the held app must also be BEHIND the catalog (R-625's shape)")
+	}
+	st.HoldReason = why
+	st.UpdatePhase, st.UpdateErrorKey, st.UpdateError = stacks.UpdatePhaseFailed, stacks.UpdateErrorKeyHeld, why
+	return st, b
+}
+
+func renderHeld(t *testing.T, s *Server, st stacks.Stack, lang, tmpl string) string {
+	t.Helper()
+	c := i18nCase{"r625_" + tmpl, tmpl, func() map[string]interface{} {
+		if tmpl == "stacks" {
+			d := i18nLayoutData("stacks", "Alkalmazások")
+			d["Stacks"] = []stacks.Stack{st}
+			for _, k := range []string{"Subdomains", "MissingStorage", "NetworkStubs", "NetworkWarnings", "StorageLabels"} {
+				d[k] = map[string]string{}
+			}
+			return d
+		}
+		d := i18nLayoutData("stacks", "Kimai")
+		cp := st
+		d["Stack"], d["Meta"], d["AppInfo"] = &cp, cp.Meta, cp.Meta.AppInfo
+		return d
+	}}
+	return html.UnescapeString(renderI18nCase(t, s, lang, c))
+}
+
+// R-625 — a held app says it is stopped and needs a restore, in both languages, on the app page; and
+// NEITHER page offers the Update button.
+//
+// COMPANION RED-PROOF (REPORT.md): remove the HoldReason branch from updateBadgeAt (v0.264.0's shape)
+// — the Hungarian page says „Frissítés elérhető" and this fails.
+func TestR625_AHeldAppSaysStoppedRestoreNeeded(t *testing.T) {
+	s := i18nTestServer(t)
+	st, _ := heldBehindStack(t, s, "")
+	for _, c := range []struct{ lang, badge, invite, title string }{
+		{"hu", "Megállítva — visszaállítás szükséges", "Frissítés elérhető", "A(z) kimai frissítése"},
+		{"en", "Stopped — restore needed", "Update available", "The update of kimai"},
+	} {
+		page := renderHeld(t, s, st, c.lang, "app_info")
+		if !strings.Contains(page, `<span class="tag tag-error" title="`) || !strings.Contains(page, c.badge) {
+			t.Errorf("%s app page: the held badge %q (tag-error) is missing", c.lang, c.badge)
+		}
+		if strings.Contains(page, c.invite) {
+			t.Errorf("%s app page still invites an update (%q) on a HELD app — R-625", c.lang, c.invite)
+		}
+		if i := strings.Index(page, c.badge); i > 0 && !strings.Contains(page[max(0, i-400):i], c.title) {
+			t.Errorf("%s: the badge title must open with the hold sentence (%q)", c.lang, c.title)
+		}
+		for _, tmpl := range []string{"app_info", "stacks"} {
+			if strings.Contains(renderHeld(t, s, st, c.lang, tmpl), `'kimai', 'update')`) {
+				t.Errorf("%s %s: the Update button is rendered for a HELD app", c.lang, tmpl)
+			}
+		}
+	}
+	// CONTROL: the same app, not held and running, DOES get the button on the list — so the check above
+	// can see a button when there is one.
+	free := st
+	free.HoldReason, free.UpdateError, free.UpdateErrorKey, free.State = "", "", "", stacks.StateRunning
+	if !strings.Contains(renderHeld(t, s, free, "hu", "stacks"), `'kimai', 'update')`) {
+		t.Fatal("control: a running, not-held app must show the Update button on the list")
+	}
+}
+
+// R-647 (1) — on an ENGLISH box, a Hungarian reader (`?lang=hu`, an operator) reads the hold and the
+// held update's error in HUNGARIAN; the English reader reads English. Both ways, both pages.
+//
+// COMPANION RED-PROOF (REPORT.md): make readerFuncs return nil (v0.264.0: no Hungarian override) — the
+// Hungarian page carries the English hold and this fails.
+func TestR647_AReaderInTheOtherLanguageReadsTheHoldInTheirs(t *testing.T) {
+	s := i18nTestServer(t)
+	st, _ := heldBehindStack(t, s, "en")
+	if !strings.Contains(st.HoldReason, "The update of kimai") {
+		t.Fatalf("fixture: an English box must store the English hold, got %q", st.HoldReason)
+	}
+	const huWant, enWant = "A(z) kimai frissítése 2026-09-23 10:00-kor nem sikerült", "The update of kimai at 2026-09-23 10:00 did not succeed"
+	for _, tmpl := range []string{"app_info", "stacks"} {
+		hu := renderHeld(t, s, st, "hu", tmpl)
+		if !strings.Contains(hu, huWant) || strings.Contains(hu, enWant) {
+			t.Errorf("%s: the Hungarian reader of an English box must read the hold in Hungarian only", tmpl)
+		}
+		en := renderHeld(t, s, st, "en", tmpl)
+		if !strings.Contains(en, enWant) || strings.Contains(en, huWant) {
+			t.Errorf("%s: the English reader must read the hold in English only", tmpl)
+		}
+	}
+}
+
+// The held badge's Hungarian literal and the bundle say the same thing (the parity rule for funcmap).
+func TestR625_HeldBadgeHungarianMatchesTheBundle(t *testing.T) {
+	b, err := i18n.Shared()
+	if err != nil {
+		t.Fatal(err)
+	}
+	st := stacks.Stack{Name: "x", HoldReason: "Egy. Kettő."}
+	if got, want := b.Msg("hu", "badge.update.held"), updateBadgeAt(st, time.Now()).Label; got != want {
+		t.Fatalf("badge.update.held: hu.json %q, funcmap %q", got, want)
+	}
+	if got := updateBadgeAt(st, time.Now()).Title; got != "Egy." {
+		t.Fatalf("the title is the first sentence, got %q", got)
+	}
+}
diff --git a/controller/internal/web/server.go b/controller/internal/web/server.go
index b78c9bf..512f6e8 100644
--- a/controller/internal/web/server.go
+++ b/controller/internal/web/server.go
@@ -310,6 +310,9 @@ func (s *Server) loadTemplates() {
 		for name, fn := range s.localeFuncs(lang) {
 			funcs[name] = fn
 		}
+		for name, fn := range s.readerFuncs(lang, funcs) {
+			funcs[name] = fn
+		}
 		start := time.Now()
 		t, st, err := parseTemplateSet(b, lang, funcs)
 		took := time.Since(start)
diff --git a/controller/internal/web/testdata/i18n_parity/app_info_deployed.html b/controller/internal/web/testdata/i18n_parity/app_info_deployed.html
index 31d4514..4e24bac 100644
--- a/controller/internal/web/testdata/i18n_parity/app_info_deployed.html
+++ b/controller/internal/web/testdata/i18n_parity/app_info_deployed.html
@@ -177,7 +177,7 @@
             <span class="stack-state-badge state-run">Fut</span>
             
             
-            
+            <span class="tag tag-error" title="A frissítés megállt: nincs friss mentés.">Megállítva — visszaállítás szükséges</span>
             <a href="https://paste.example.hu" target="_blank" class="btn btn-sm btn-outline">Megnyitás ↗</a>
             <a href="/stacks/privatebin/logs" class="btn btn-sm btn-outline">Napló</a>
             
diff --git a/controller/internal/web/testdata/i18n_parity/stacks_full.html b/controller/internal/web/testdata/i18n_parity/stacks_full.html
index b48c20a..648a50d 100644
--- a/controller/internal/web/testdata/i18n_parity/stacks_full.html
+++ b/controller/internal/web/testdata/i18n_parity/stacks_full.html
@@ -252,7 +252,7 @@
             <span class="tag tag-neutral"><span class="dot"></span>Leállítva</span>
             
             
-            
+            <span class="tag tag-error" title="A frissítés megállt, a mentés hiányzik.">Megállítva — visszaállítás szükséges</span>
             
             
             
diff --git a/controller/internal/web/updatebadge.go b/controller/internal/web/updatebadge.go
index 2e0d357..96dda11 100644
--- a/controller/internal/web/updatebadge.go
+++ b/controller/internal/web/updatebadge.go
@@ -2,6 +2,7 @@ package web
 
 import (
 	"fmt"
+	"strings"
 	"time"
 
 	"gitea.dooplex.hu/admin/felhom-controller/internal/stacks"
@@ -73,6 +74,18 @@ func compareInstalledToTemplate(s stacks.Stack) updateState {
 // never the decision. A new branch here needs its twin there, and
 // TestLocaleFuncsHungarianBundleMatchesFuncMap fails if hu.json and these literals disagree.
 func updateBadgeAt(s stacks.Stack, now time.Time) *MetaBadge {
+	// R-625 (v0.265.0). A HELD app is stopped, and the only way back is a restore — a newer catalog
+	// version does not lift the hold (`09` §6.1), so „Frissítés elérhető" beside a button that answers
+	// 409 was two verdicts disagreeing. One verdict, read by the badge and by the button (stacks.html
+	// renders no Update button while HoldReason is set), exactly as R-524 did it for Ahead. True for
+	// both hold kinds: a failed restore holds the app stopped too, and the way back is also a restore.
+	if s.HoldReason != "" {
+		return &MetaBadge{
+			Label: "Megállítva — visszaállítás szükséges",
+			Class: "tag-error",
+			Title: firstSentence(s.HoldReason),
+		}
+	}
 	switch compareInstalledToTemplate(s) {
 	case updateCurrent:
 		return &MetaBadge{
@@ -120,3 +133,12 @@ func updateBadgeAt(s stacks.Stack, now time.Time) *MetaBadge {
 //
 // It is INFORMATION ONLY. It is wired to no action, and the Frissítés button is untouched.
 func updateBadge(s stacks.Stack) *MetaBadge { return updateBadgeAt(s, time.Now().UTC()) }
+
+// firstSentence is the hold sentence's opening clause, for a badge title (R-625): up to and including
+// the first full stop that ends a sentence; the whole text when it has none.
+func firstSentence(s string) string {
+	if i := strings.Index(s, ". "); i >= 0 {
+		return s[:i+1]
+	}
+	return s
+}
diff --git a/controller/scripts/i18n_go_keys.json b/controller/scripts/i18n_go_keys.json
index 2340da3..e20cd5d 100644
--- a/controller/scripts/i18n_go_keys.json
+++ b/controller/scripts/i18n_go_keys.json
@@ -61,7 +61,8 @@
     "page.title.storage_network": "slice 1 -- pinned by TestHandlerTitleKeysMatchHungarianTitle",
     "event.app_update_undone": "BORN AS A KEY, v0.264.0 (R-606 / the update events) -- a NEW sentence, never a Go literal.",
     "settings_notifications.app_update_undone": "BORN AS A KEY, v0.264.0 (R-606 / the update events) -- a NEW sentence, never a Go literal.",
-    "settings_notifications.app_update_held": "BORN AS A KEY, v0.264.0 (R-606 / the update events) -- a NEW sentence, never a Go literal."
+    "settings_notifications.app_update_held": "BORN AS A KEY, v0.264.0 (R-606 / the update events) -- a NEW sentence, never a Go literal.",
+    "badge.update.held": "BORN AS A KEY, v0.265.0 (R-625) -- a NEW badge for a held app; its Hungarian twin in updatebadge.go is pinned by TestR625_HeldBadgeHungarianMatchesTheBundle."
   },
   "flash.share.already_on": "A megosztás már be van kapcsolva.",
   "flash.share.enable_failed": "A megosztás bekapcsolása nem sikerült.",