diff --git a/TEST_INTEGRITY.md b/TEST_INTEGRITY.md
index 809aeb65..9e29df29 100644
--- a/TEST_INTEGRITY.md
+++ b/TEST_INTEGRITY.md
@@ -731,3 +731,32 @@ test_integrity_note:
 ```
 
 Grouped mutation coverage: last-owner cancellation mutant kills SCN-last-probe-owner-disposal; unconditional cancel kills SCN-shared-queued-owner; unconditional inflight deletion and cancelled-normalization mutants each kill SCN-cancelled-replacement-inflight; reversed queue priority kills SCN-probe-priority-concurrency; disabled eager path kills SCN-hidden-healthy-reveal; stale result guard and late-entry membership mutants each kill SCN-active-shared-late-entry; non-idempotent disposer kills SCN-repeat-disposal-reattach; owner-dependent active caching kills SCN-ownerless-active-cache; missing cache writes kills SCN-existing-probe-cache-ttl. All11 variants exit1 with1 failed scenario,0 cancelled, and are reverted. Browser checks corroborate the same disposal/shared-owner/hidden-reveal/source-visibility scenarios with native observers; no browser CPU or actual media fallback claim.
+
+Additional client-reliability-3 scenarios (same invariant and controls):
+
+```yaml
+test_integrity_note:
+  change_type: ["new_tests"]
+  scenarios:
+    - id: SCN-shared-badge-listener-lifetime
+      given: "two feed registrations share the app callback"
+      when: "badge updates occur and registrations dispose one at a time"
+      then: "each event notifies once, one surviving owner retains the callback, last disposal stops notifications"
+    - id: SCN-checking-callback-disposal
+      given: "a live card's checking callback synchronously disposes the feed"
+      when: "badge notification returns to probe submission"
+      then: "no probe is submitted after unregister"
+  observable_outcomes:
+    - "callback event count/state sequence and external probe start count"
+  determinism_controls:
+    - "real badge/observer path, synchronous disposal callback and deferred probe IO"
+  anti_cheat_rationale:
+    prevents:
+      - "duplicated shared callbacks or dropping live callback owners"
+      - "post-disposal work submitted after synchronous reentrant teardown"
+  relaxation:
+    did_relax_any_assertion: false
+    if_true_explain_spec_basis: ""
+```
+
+Duplicate wrapper and unconditional listener release variants each fail SCN-shared-badge-listener-lifetime. Removing the post-checking membership guard fails SCN-checking-callback-disposal (actual starts1 vs expected0). These3 additional variants exit1 with1 failed/0 cancelled and are reverted; total14 killed variants for11 lifecycle scenarios. Original source/priority/observer expectations remain unchanged.
diff --git a/decisions/DECISIONS_2026-10-19_22-02-00_client-reliability-3.md b/decisions/DECISIONS_2026-10-19_22-02-00_client-reliability-3.md
new file mode 100644
index 00000000..419b6663
--- /dev/null
+++ b/decisions/DECISIONS_2026-10-19_22-02-00_client-reliability-3.md
@@ -0,0 +1,7 @@
+# Final callback ownership corrections
+
+- Caller review shows app.boundStreamHealthBadgeHandler is shared across feeds; preserve global callback dedup with registration reference counts. Disposing one owner must retain the handler for another owner; final disposal removes it. Wrapping each registration separately would duplicate callbacks.
+- Synchronous checking notifications can dispose/rerender the feed before queue submission. Measured1 start after disposal, expected0; check live membership after notification before creating queued work.
+- Late observer scenario also asserts no disposed-card notifications reach a remaining registration, not just final badge state. This rejects a removed observer guard even when the reentrant guard still prevents queue submission.
+- Final source has11 lifecycle scenarios;14 independent mutants were rerun against this exact product tree, all exit1 with1 failed/0 cancelled, then restored. Source outside gridHealth.js remains untouched.
+- Corrected build and Chromium/Firefox built-dist/native-observer smoke both pass; no external requests or page errors. Actual media playback fallback and CPU/RSS remain unverified.
diff --git a/js/gridHealth.js b/js/gridHealth.js
index bbe57fc4..8d59c06c 100644
--- a/js/gridHealth.js
+++ b/js/gridHealth.js
@@ -21,14 +21,14 @@ import {
   getLivenessProbePrefetchMargin,
 } from "./constants.js";
 
-const badgeUpdateListeners = new Set();
+const badgeUpdateListeners = new Map();
 
 function notifyBadgeUpdate(payload) {
   if (!payload) {
     return;
   }
 
-  badgeUpdateListeners.forEach((listener) => {
+  badgeUpdateListeners.forEach((_owners, listener) => {
     if (typeof listener !== "function") {
       return;
     }
@@ -479,7 +479,7 @@ function setBadge(card, state, details) {
   notifyBadgeUpdate(payload);
 }
 
-function handleCardVisible({ card, pendingByCard, priority = 0 }) {
+function handleCardVisible({ card, pendingByCard, liveCards, priority = 0 }) {
   if (!(card instanceof HTMLElement)) {
     return;
   }
@@ -515,6 +515,9 @@ function handleCardVisible({ card, pendingByCard, priority = 0 }) {
   }
 
   setBadge(card, "checking");
+  // Badge callbacks can synchronously dispose/rerender the feed. Do not create
+  // a new request after its unregister hook has already run.
+  if (!liveCards.has(card)) return;
 
   const encodedUrl = card.dataset.urlHealthUrl || "";
   const webSeed = encodedUrl ? decodeURIComponent(encodedUrl) : "";
@@ -601,6 +604,7 @@ const gridCardObserver = createCardObserver({
       handleCardVisible({
         card,
         pendingByCard: state.pendingByCard,
+        liveCards: state.liveCards,
         priority: PRIORITY_BASELINE,
       });
     }
@@ -624,6 +628,7 @@ const gridCardObserver = createCardObserver({
     handleCardVisible({
       card,
       pendingByCard: state.pendingByCard,
+      liveCards: state.liveCards,
       priority,
     });
   },
@@ -634,16 +639,23 @@ export function attachHealthBadges(container, options = {}) {
     return () => {};
   }
   const listener =
-    typeof options?.onUpdate === "function"
-      ? (payload) => options.onUpdate(payload)
-      : null;
-  if (listener) badgeUpdateListeners.add(listener);
+    typeof options?.onUpdate === "function" ? options.onUpdate : null;
+  if (listener) {
+    badgeUpdateListeners.set(
+      listener,
+      (badgeUpdateListeners.get(listener) || 0) + 1
+    );
+  }
   gridCardObserver.observe(container);
   let disposed = false;
   return () => {
     if (disposed) return;
     disposed = true;
-    if (listener) badgeUpdateListeners.delete(listener);
+    if (listener) {
+      const remaining = badgeUpdateListeners.get(listener) - 1;
+      if (remaining > 0) badgeUpdateListeners.set(listener, remaining);
+      else badgeUpdateListeners.delete(listener);
+    }
     gridCardObserver.disconnect(container);
   };
 }
diff --git a/test_logs/TEST_LOG_2026-10-19_21-59-00_client-reliability-3-review.md b/test_logs/TEST_LOG_2026-10-19_21-59-00_client-reliability-3-review.md
new file mode 100644
index 00000000..29579d14
--- /dev/null
+++ b/test_logs/TEST_LOG_2026-10-19_21-59-00_client-reliability-3-review.md
@@ -0,0 +1,10 @@
+# client-reliability-3 final lifecycle review
+
+- At712b05cf, final lint11/11 and build/artifact verification passed; full426-file run was interrupted130 after133 file exits, before file134 completed. No full green claimed for that partial run.
+- Final caller review found app.boundStreamHealthBadgeHandler is shared across feed registrations. Preserve the existing global dedup contract with per-callback registration counts rather than per-registration callback wrappers. New shared-listener scenario checks one notification, survival after one owner disposes, and no notifications after the last owner disposes. Both duplicate-listener and unconditional-release mutants fail naturally (exit1,1 failed,0 cancelled).
+- New actual callback-disposal reproduction: onUpdate disposes feed at checking, but old code still submits1 probe after unregister. Added post-notification live-card guard; missing-guard mutant fails naturally (exit1,1 failed,0 cancelled). No shared observer/VideoListView code edits needed; all callers read only.
+- Corrected lifecycle suite11/11 passes naturally; total focused count now27 scenarios across4 files. New scoped correction does not change queue/cache/concurrency policy or active probes. Existing11 mutants plus these3 variants total14 rejected/restored variants.
+- Browser/build/lint and full gates rerun below on final correction; earlier built-dist browser evidence belongs to the prior product tree and is not substituted for the corrected build.
+- Final14-variant rerun against corrected listener/reentrant tree: all14 killed (each exit1/1 failed/0 cancelled), source restored. Late-entry case additionally asserts no stale disposed-card badge event reaches a remaining live listener.
+- Corrected build/artifact verification exit0 and lint11/11 exit0. Corrected Chromium/Firefox native-observer smoke exit0: each4 active/0 obsolete starts, shared live healthy/disposed checking/0 stale updates, healthy hidden reveal, both source visibility fallbacks true,0 page errors/0 attempted external requests. Final log /tmp/client-reliability-3-browser-corrected.log; server/browsers closed and browser reservation released.
+- Final direct focused runs after full mutant restoration: lifecycle11/11 (513.51ms), priority6/6 (381.59ms), visibility9/9 (394.34ms), observer1/1 (225.61ms), each exit0 naturally with0 failed/cancelled/skipped. Total27 scenarios across4 files. No source changes follow these focused checks.
diff --git a/tests/grid-health-lifecycle.test.mjs b/tests/grid-health-lifecycle.test.mjs
index 563e1340..f6bd78bb 100644
--- a/tests/grid-health-lifecycle.test.mjs
+++ b/tests/grid-health-lifecycle.test.mjs
@@ -210,6 +210,59 @@ test("disposing one of two queued source owners preserves the live owner", async
   );
 });
 
+test("disposal from a checking callback prevents probe submission", async (t) => {
+  const f = fixture(t);
+  const feed = f.feed([magnet()], {
+    onUpdate: ({ state }) => {
+      if (state === "checking") feed.dispose();
+    }
+  });
+  feed.observer.show();
+  await flush();
+  assert.equal(
+    f.calls.length,
+    0,
+    "checking notification must not submit work after synchronous disposal"
+  );
+});
+
+test("shared badge listener stays deduplicated until its last registration disposes", async (t) => {
+  const f = fixture(t);
+  const updates = [];
+  const listener = (event) => updates.push(event);
+  const first = f.feed([magnet()], { onUpdate: listener });
+  const live = f.feed([magnet()], { onUpdate: listener });
+  first.observer.show();
+  assert.equal(
+    updates.length,
+    1,
+    "same callback must receive each update once across registrations"
+  );
+  first.dispose();
+  updates.length = 0;
+  live.observer.show();
+  assert.equal(
+    updates.length,
+    1,
+    "disposing one registration must retain the shared callback"
+  );
+  f.calls[1].finish();
+  await flush();
+  assert.deepEqual(
+    updates.map((event) => event.state),
+    ["checking", "healthy"]
+  );
+  live.dispose();
+  updates.length = 0;
+  const other = f.feed([magnet()]);
+  other.observer.show();
+  assert.equal(
+    updates.length,
+    0,
+    "last listener owner disposal must stop later notifications"
+  );
+});
+
 test("cancelled source can be re-requested before old settlement without cache poison or duplicate work", async (t) => {
   const f = fixture(t);
   f.saturate();
@@ -317,9 +370,17 @@ test("active probe remains shared while disposed owners and late observer entrie
   const first = f.feed([source]);
   first.observer.show();
   first.dispose();
+  const liveUpdates = [];
+  const live = f.feed([source], {
+    onUpdate: (event) => liveUpdates.push(event)
+  });
   first.observer.show(); // An already-delivered entry may arrive after disconnect.
-  const live = f.feed([source]);
   live.observer.show();
+  assert.equal(
+    liveUpdates.some((event) => event.card === first.cards[0]),
+    false,
+    "late observer entries must not emit disposed-card badge events"
+  );
   assert.equal(
     f.calls.length,
     1,
diff --git a/todo/TODO_2026-10-19_22-02-00_client-reliability-3.md b/todo/TODO_2026-10-19_22-02-00_client-reliability-3.md
new file mode 100644
index 00000000..e604cb54
--- /dev/null
+++ b/todo/TODO_2026-10-19_22-02-00_client-reliability-3.md
@@ -0,0 +1,6 @@
+- [x] Actual disposal reproduction1→0 obsolete starts, queued/shared/active lifecycle preservation.
+- [x] Caller review preserves shared-handler dedup and checks synchronous checking-callback disposal1→0 starts.
+- [x] Focused27/27 across4 files (11 lifecycle+6 priority+9 visibility+1 observer),14 reverted mutants killed.
+- [x] Corrected product lint11/11, build/artifact verification and Chromium/Firefox built-dist/native-observer smoke pass; browser released.
+- [ ] Commit corrected tree, merge unstable, restart full426-file60s gate and final lint; normal queue/durable landing/ancestry, dispatch landing event, close/final report.
+- Not qualified: actual media/torrent fallback, live peers, CPU/RSS, URL cancellation or active abort. Dispatch owns preview refresh. No new item without routing.