diff --git a/CHANGELOG.md b/CHANGELOG.md
index c5c38d8..244e434 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,3 +1,22 @@
+## v0.263.1 — the undo asks the OLD probe even when the new one has marked the app unhealthy (2026-09-23, R-637)
+
+**MinAgent: 0.131.0** (unchanged). No new strings.
+
+**Caught by the live proof on 9202, not by a test.** v0.263.0's undo put docmost's data and old
+definition back, and then judged the old version "did not start" after the full 90 s — while it was
+serving. The periodic health probe judges an app with its CURRENT `.felhom.yml`; the new version had
+brought a probe the old one does not answer (the drill's wrong port, and a real case whenever a probe
+moves with a version), so it flipped the app to `unhealthy` every 10 s — and the update's health wait
+only probes an app whose state reads `running`. The old probe was never asked (`last: state unhealthy`).
+
+**Fix:** with the undo's override probe, an `unhealthy` app IS probed, and the old check decides; it
+is never settled on container state (the settle path still requires `running`). The normal update
+path is unchanged. **Why the unit tests missed it:** `TestUndo_UsesTheOldProbe` injects the undo's
+whole health wait, so it never reached the state gate. The new test drives the real wait loop with only
+the network probe faked (new seam `Manager.probeRunFn`) and reproduces the live message verbatim when
+the fix is switched off. The first live attempt ended in an honest HOLD (data put back, old version
+not judged healthy) — kept as evidence.
+
 ## v0.263.0 — a failed update puts the app back by itself (2026-09-23, R-637)
 
 **MinAgent: 0.131.0** (unchanged). New strings, all born as bundle keys (hu + en): `app_info.update_undone`,
diff --git a/controller/internal/stacks/manager.go b/controller/internal/stacks/manager.go
index c957f68..61f9360 100644
--- a/controller/internal/stacks/manager.go
+++ b/controller/internal/stacks/manager.go
@@ -242,6 +242,7 @@ type Manager struct {
 	// which receives the probe it must use (nil ⇒ waitUpdateHealthyMeta).
 	undoCopier         volumeCopier
 	updateUndoHealthFn func(ctx context.Context, name string, timeout time.Duration, meta *Metadata) (bool, string)
+	probeRunFn         func(t probeTarget) *HealthProbeResult // the health wait's network probe; nil ⇒ runChecks
 	updateMemoryFn     func(newReqMB, newLimitMB, releasedReqMB, releasedLimitMB int) (refusal error, warning string)
 	updateDiskFreeFn   func() (freeGiB float64, ok bool)
 	updateNowFn        func() time.Time
diff --git a/controller/internal/stacks/undo_test.go b/controller/internal/stacks/undo_test.go
index b516794..460a8ff 100644
--- a/controller/internal/stacks/undo_test.go
+++ b/controller/internal/stacks/undo_test.go
@@ -415,3 +415,66 @@ func TestUndo_RemovalDeletesKeptCopies(t *testing.T) {
 		t.Errorf("removed %d, left %d", n, fc.nCopies())
 	}
 }
+
+// TestUndo_OldProbeRunsOnAnAppTheCurrentProbeMarkedUnhealthy drives the REAL wait loop
+// (waitUpdateHealthyMeta → RefreshStatus → `docker ps` → the health-probe override → the state gate →
+// findProbeContainerMeta) with only the network probe faked. The app's CURRENT .felhom.yml probe has
+// failed, so the refresh reads it Unhealthy; the undo's OLD probe answers.
+//
+// FOUND LIVE, NOT BY A TEST: v0.263.0's first build gated on StateRunning, so the old probe was never
+// asked and docmost's undo on 9202 was reported "did not start" after the full 90 s while the old
+// version served. TestUndo_UsesTheOldProbe could not see it — it injects updateUndoHealthFn and so
+// never reaches this loop (the seam-boundary class: a test proves only what is behind its seam).
+//
+// COMPANION RED-PROOF (REPORT.md): drop `|| undoProbe` from the state case — the wait times out with
+// `last: state unhealthy` and this test fails.
+func TestUndo_OldProbeRunsOnAnAppTheCurrentProbeMarkedUnhealthy(t *testing.T) {
+	m, dir, _, _, _ := newUndoManager(t)
+	mustWrite(t, filepath.Join(dir, ".felhom.yml"), undoMetaNew) // the catalog's probe (3999) is current
+	m.mu.Lock()
+	m.stacks["nextcloud"].Meta = LoadMetadata(dir)
+	m.stacks["nextcloud"].HealthProbe = &HealthProbeResult{Healthy: false} // the periodic probe failed on 3999
+	m.mu.Unlock()
+	m.execFn = func(name string, args ...string) (string, error) {
+		if name == "docker" && len(args) > 0 && args[0] == "ps" {
+			return "nextcloud\tnextcloud:31.0.14-apache\trunning\tUp 30 seconds\tnextcloud\n", nil
+		}
+		return "", nil
+	}
+	m.inspectRestartPolicyFn = func(string) (string, error) { return "unless-stopped", nil }
+	m.updateNowFn = time.Now // the shared setup freezes the clock; this loop's deadline needs it moving
+	var probed []int
+	m.probeRunFn = func(pt probeTarget) *HealthProbeResult {
+		probed = append(probed, pt.checks[0].Port)
+		return &HealthProbeResult{Healthy: pt.checks[0].Port == 3000}
+	}
+	if err := m.RefreshStatus(); err != nil {
+		t.Fatal(err)
+	}
+	if st, _ := m.GetStack("nextcloud"); st.State != StateUnhealthy {
+		t.Fatalf("setup: the refresh must read the app Unhealthy (the current probe failed), got %q", st.State)
+	}
+	old, err := savePreUpdateMeta(dir) // what the job kept — but it holds the NEW file here, so write the OLD one
+	if err != nil {
+		t.Fatal(err)
+	}
+	mustWrite(t, filepath.Join(old, ".felhom.yml"), undoMetaOld)
+	oldMeta := LoadMetadata(old)
+
+	ok, detail := m.waitUpdateHealthyMeta(context.Background(), "nextcloud", time.Second, &oldMeta)
+	if !ok {
+		t.Fatalf("the old version answers its own probe and must be judged healthy; got %q, probed ports %v", detail, probed)
+	}
+	if len(probed) == 0 || probed[0] != 3000 {
+		t.Errorf("the undo must probe the OLD port (3000), probed %v", probed)
+	}
+	// And without an override the same Unhealthy app is NOT probed or settled — the normal update path
+	// is unchanged.
+	probed = nil
+	m.mu.Lock()
+	m.stacks["nextcloud"].HealthProbe = &HealthProbeResult{Healthy: false} // the first call stored its healthy result
+	m.mu.Unlock()
+	if ok, _ := m.waitUpdateHealthyMeta(context.Background(), "nextcloud", time.Second, nil); ok || len(probed) != 0 {
+		t.Errorf("without an override an Unhealthy app must stay unhealthy and unprobed; ok=%v probed=%v", ok, probed)
+	}
+}
diff --git a/controller/internal/stacks/update.go b/controller/internal/stacks/update.go
index d95a5d4..a113c59 100644
--- a/controller/internal/stacks/update.go
+++ b/controller/internal/stacks/update.go
@@ -918,11 +918,22 @@ func (m *Manager) waitUpdateHealthyMeta(ctx context.Context, name string, timeou
 	for {
 		_ = m.RefreshStatus()
 		st, ok := m.GetStack(name)
+		// v0.263.0 — THE UNDO'S PROBE MUST BE ALLOWED TO RUN. The periodic health probe judges the app
+		// with the CURRENT .felhom.yml and flips a running app to StateUnhealthy when that check fails —
+		// which is exactly the undo's situation when the new version brought a probe the old one does
+		// not answer. Gating on StateRunning alone meant the old probe was never asked and the undo was
+		// reported "did not start" after the full timeout. MEASURED LIVE on 9202 2026-09-23 (docmost,
+		// `last: state unhealthy` for 90 s while the old version served). So with an override that
+		// declares checks, an Unhealthy app is PROBED — and the override's own check decides. It never
+		// settles an Unhealthy app on container state (below: the settle path requires StateRunning).
+		// Pinned by TestUndo_OldProbeRunsOnAnAppTheCurrentProbeMarkedUnhealthy.
+		undoProbe := ok && override != nil && st.State == StateUnhealthy &&
+			override.HealthCheck != nil && len(override.HealthCheck.Checks) > 0
 		switch {
 		case !ok:
 			last = "stack vanished"
 			runningSince = time.Time{}
-		case st.State == StateRunning:
+		case st.State == StateRunning || undoProbe:
 			// A declared health check is only usable if it resolves to a container. When it does
 			// not, the app is judged the same way an app with NO declared check is judged —
 			// settling on container state — and the log says so.
@@ -942,7 +953,7 @@ func (m *Manager) waitUpdateHealthyMeta(ctx context.Context, name string, timeou
 				c, candidates := findProbeContainerMeta(name, &meta, st.Containers)
 				if c != "" {
 					usable = true
-					res := m.runChecks(probeTarget{stackName: name, containerName: c, checks: hc.Checks})
+					res := m.probeRun(probeTarget{stackName: name, containerName: c, checks: hc.Checks})
 					m.mu.Lock()
 					if s, ok := m.stacks[name]; ok {
 						s.HealthProbe = res
@@ -958,7 +969,9 @@ func (m *Manager) waitUpdateHealthyMeta(ctx context.Context, name string, timeou
 						name, name, candidates)
 				}
 			}
-			if !usable {
+			if !usable && st.State != StateRunning {
+				last = "unhealthy, and the check resolves to no container"
+			} else if !usable {
 				if runningSince.IsZero() {
 					runningSince = m.now()
 				}
@@ -1213,3 +1226,13 @@ func (m *Manager) ResumeInterruptedUpdates(ctx context.Context) int {
 	}
 	return len(names)
 }
+
+// probeRun is the network half of the update's health wait — its own seam, so a test can drive the
+// REAL wait loop (the state gate, the container resolution, the settle rule) with only the HTTP/TCP
+// probe faked. nil ⇒ runChecks.
+func (m *Manager) probeRun(t probeTarget) *HealthProbeResult {
+	if m.probeRunFn != nil {
+		return m.probeRunFn(t)
+	}
+	return m.runChecks(t)
+}