diff --git a/internal/ai/findings_update_safety_emit_test.go b/internal/ai/findings_update_safety_emit_test.go
new file mode 100644
index 000000000..c2d25c127
--- /dev/null
+++ b/internal/ai/findings_update_safety_emit_test.go
@@ -0,0 +1,225 @@
+package ai
+
+import (
+	"testing"
+	"time"
+
+	"github.com/rcourtman/pulse-go-rewrite/internal/models"
+)
+
+// TestUpdateSafetyEmit_FullCycle wires a real FindingsStore to an
+// UpdateSafetyWatcher and drives three observe cycles, mirroring the
+// storm-throttler proof in findings_storm_emit_test.go.
+//
+// Cycle 1: baseline -- no findings emitted.
+// Cycle 2: digest changed -- watcher emits one finding; store accepts it.
+// Cycle 3: same digest, window elapsed, no restarts -- watcher returns a
+//          resolve sentinel; ResolveWithReason marks the finding AutoResolved.
+func TestUpdateSafetyEmit_FullCycle(t *testing.T) {
+	store := NewFindingsStore()
+	w := newUpdateSafetyWatcher()
+	t0 := time.Now()
+
+	hostA := "host-a"
+	ctrA := "ctr-1"
+	key := UpdateSafetyFindingPrefix + ":" + hostA + "/" + ctrA
+
+	// Cycle 1 -- baseline.
+	emit1, res1 := w.Observe([]models.DockerHost{makeHost(hostA, ctrA, "sha256:aaa", 0)}, t0)
+	if len(emit1) != 0 || len(res1) != 0 {
+		t.Fatalf("cycle 1: want silent, got emit=%d resolve=%d", len(emit1), len(res1))
+	}
+
+	// Cycle 2 -- digest changed.
+	emit2, res2 := w.Observe([]models.DockerHost{makeHost(hostA, ctrA, "sha256:bbb", 0)}, t0.Add(5*time.Second))
+	if len(emit2) != 1 {
+		t.Fatalf("cycle 2: want 1 finding, got %d", len(emit2))
+	}
+	if len(res2) != 0 {
+		t.Fatalf("cycle 2: want 0 resolves, got %d", len(res2))
+	}
+
+	// Feed finding into the real store.
+	isNew := store.Add(emit2[0])
+	if !isNew {
+		t.Fatal("cycle 2: store.Add should report new finding")
+	}
+	stored := store.Get(key)
+	if stored == nil {
+		t.Fatalf("cycle 2: finding %q missing from store after Add", key)
+	}
+	if !stored.IsActive() {
+		t.Errorf("cycle 2: finding should be active, got %+v", stored)
+	}
+	if stored.Category != FindingCategoryReliability {
+		t.Errorf("cycle 2 category: want %q, got %q", FindingCategoryReliability, stored.Category)
+	}
+	if stored.ResourceType != "app-container" {
+		t.Errorf("cycle 2 resource_type: want %q, got %q", "app-container", stored.ResourceType)
+	}
+
+	// Cycle 3 -- same digest, window elapsed, no restarts.
+	afterWindow := t0.Add(5*time.Second + updateSafetyVerifyWindow + time.Second)
+	emit3, res3 := w.Observe([]models.DockerHost{makeHost(hostA, ctrA, "sha256:bbb", 0)}, afterWindow)
+	if len(emit3) != 0 {
+		t.Errorf("cycle 3: want 0 findings, got %d", len(emit3))
+	}
+	if len(res3) != 1 {
+		t.Fatalf("cycle 3: want 1 resolve sentinel, got %d", len(res3))
+	}
+	sentinel := res3[0]
+	if sentinel.DedupKey != key {
+		t.Errorf("sentinel key: want %q, got %q", key, sentinel.DedupKey)
+	}
+	if sentinel.Reason != updateSafetyResolveReason {
+		t.Errorf("sentinel reason: want %q, got %q", updateSafetyResolveReason, sentinel.Reason)
+	}
+
+	// Route sentinel through ResolveWithReason.
+	if !store.ResolveWithReason(sentinel.DedupKey, sentinel.Reason) {
+		t.Fatal("ResolveWithReason returned false -- finding missing or already resolved")
+	}
+	resolved := store.Get(key)
+	if resolved == nil {
+		t.Fatal("resolved finding missing from store")
+	}
+	if resolved.IsActive() {
+		t.Errorf("resolved finding still active: %+v", resolved)
+	}
+	if !resolved.AutoResolved {
+		t.Errorf("AutoResolved: want true, got false")
+	}
+	if resolved.ResolveReason != updateSafetyResolveReason {
+		t.Errorf("ResolveReason: want %q, got %q", updateSafetyResolveReason, resolved.ResolveReason)
+	}
+}
+
+// TestUpdateSafetyEmit_TwoContainersDontCollideDedupKeys proves that digest
+// changes on two separate containers produce non-colliding finding IDs in the
+// store, and that resolving one does not affect the other.
+func TestUpdateSafetyEmit_TwoContainersDontCollideDedupKeys(t *testing.T) {
+	store := NewFindingsStore()
+	w := newUpdateSafetyWatcher()
+	t0 := time.Now()
+
+	hosts1 := []models.DockerHost{
+		{
+			ID:       "h1",
+			Hostname: "h1.host",
+			Containers: []models.DockerContainer{
+				{ID: "cX", Name: "alpha", ImageDigest: "sha256:aaa", RestartCount: 0},
+				{ID: "cY", Name: "beta", ImageDigest: "sha256:zzz", RestartCount: 0},
+			},
+		},
+	}
+	w.Observe(hosts1, t0)
+
+	// Change both digests.
+	hosts2 := []models.DockerHost{
+		{
+			ID:       "h1",
+			Hostname: "h1.host",
+			Containers: []models.DockerContainer{
+				{ID: "cX", Name: "alpha", ImageDigest: "sha256:bbb", RestartCount: 0},
+				{ID: "cY", Name: "beta", ImageDigest: "sha256:yyy", RestartCount: 0},
+			},
+		},
+	}
+	emit2, _ := w.Observe(hosts2, t0.Add(5*time.Second))
+	if len(emit2) != 2 {
+		t.Fatalf("want 2 findings for 2 changed digests, got %d", len(emit2))
+	}
+	for _, f := range emit2 {
+		if !store.Add(f) {
+			t.Errorf("store.Add returned false (duplicate?) for finding %q", f.ID)
+		}
+	}
+
+	keyX := UpdateSafetyFindingPrefix + ":h1/cX"
+	keyY := UpdateSafetyFindingPrefix + ":h1/cY"
+	fX := store.Get(keyX)
+	fY := store.Get(keyY)
+	if fX == nil || fY == nil {
+		t.Fatalf("one or both findings missing: fX=%v fY=%v", fX, fY)
+	}
+
+	// Resolve cX via sentinel.
+	afterWindow := t0.Add(5*time.Second + updateSafetyVerifyWindow + time.Second)
+	// Only feed cX into the next observe trip.
+	hostsX := []models.DockerHost{
+		{ID: "h1", Hostname: "h1.host", Containers: []models.DockerContainer{
+			{ID: "cX", Name: "alpha", ImageDigest: "sha256:bbb", RestartCount: 0},
+		}},
+	}
+	_, res3 := w.Observe(hostsX, afterWindow)
+	var sentinelX *resolveSentinel
+	for i := range res3 {
+		if res3[i].DedupKey == keyX {
+			sentinelX = &res3[i]
+			break
+		}
+	}
+	if sentinelX == nil {
+		t.Fatal("expected resolve sentinel for cX, got none")
+	}
+	store.ResolveWithReason(sentinelX.DedupKey, sentinelX.Reason)
+
+	// cX should be resolved; cY should still be active.
+	if store.Get(keyX).IsActive() {
+		t.Errorf("cX finding should be resolved")
+	}
+	if !store.Get(keyY).IsActive() {
+		t.Errorf("cY finding should still be active")
+	}
+}
+
+// TestUpdateSafetyEmit_EscalationUpdatesExistingFinding proves that when a
+// restart occurs after the initial Info emission, re-adding the finding with
+// Warning severity updates the existing record (TimesRaised advances, severity
+// escalates) rather than spawning a duplicate.
+func TestUpdateSafetyEmit_EscalationUpdatesExistingFinding(t *testing.T) {
+	store := NewFindingsStore()
+	w := newUpdateSafetyWatcher()
+	t0 := time.Now()
+
+	// Trip 1: baseline.
+	w.Observe([]models.DockerHost{makeHost("h2", "c1", "sha256:aaa", 0)}, t0)
+
+	// Trip 2: digest changed, no restarts -- Info.
+	emit2, _ := w.Observe([]models.DockerHost{makeHost("h2", "c1", "sha256:bbb", 0)}, t0.Add(5*time.Second))
+	if len(emit2) != 1 {
+		t.Fatalf("trip 2: want 1 finding, got %d", len(emit2))
+	}
+	store.Add(emit2[0])
+	key := UpdateSafetyFindingPrefix + ":h2/c1"
+	afterFirst := store.Get(key)
+	if afterFirst == nil {
+		t.Fatal("finding missing after first add")
+	}
+	if afterFirst.Severity != FindingSeverityInfo {
+		t.Errorf("first emit severity: want Info, got %q", afterFirst.Severity)
+	}
+	raisedAfterFirst := afterFirst.TimesRaised
+
+	// Trip 3: restart count increased -- Warning escalation.
+	emit3, _ := w.Observe([]models.DockerHost{makeHost("h2", "c1", "sha256:bbb", 3)}, t0.Add(20*time.Second))
+	if len(emit3) != 1 {
+		t.Fatalf("trip 3: want 1 escalated finding, got %d", len(emit3))
+	}
+	if emit3[0].Severity != FindingSeverityWarning {
+		t.Errorf("escalated severity: want Warning, got %q", emit3[0].Severity)
+	}
+	store.Add(emit3[0])
+
+	// Same finding ID -- should be updated, not duplicated.
+	afterEscalation := store.Get(key)
+	if afterEscalation == nil {
+		t.Fatal("finding missing after escalation add")
+	}
+	if afterEscalation.Severity != FindingSeverityWarning {
+		t.Errorf("post-escalation severity: want Warning, got %q", afterEscalation.Severity)
+	}
+	if afterEscalation.TimesRaised <= raisedAfterFirst {
+		t.Errorf("TimesRaised should advance: was %d, now %d", raisedAfterFirst, afterEscalation.TimesRaised)
+	}
+}