Flaky: useTimelineMode live-poll resume fails when a leaked loadSource writes SCAN_PROGRESS

[thalida]

tests/hooks/useTimelineMode.test.ts > live poll suspends in Timeline mode > does not probe the signature endpoint while in Timeline mode, and resumes on exit fails intermittently on main.

AssertionError: expected "fetch" to be called at least once
  tests/hooks/useTimelineMode.test.ts:361:22

Twice in three weeks, both on main, both the same assertion:

• run 31991836275 — 2026-08-17
• run 33795322008 — 2026-09-03 (on bb586d33, which touches no app/ or api/ file)

Not reproducible on demand: 12/12 passes running the file alone, and the full suite with coverage passes 2153/2153 locally. It needs CI-level scheduling pressure.

Cause

TIMELINE_MODE never restarts the poll timer. setupLiveUpdates starts one interval and tick() early-returns while Timeline is on (app/src/hooks/useManifestSource.ts:286):

if (inFlight) return;
if (TIMELINE_MODE.peek()) return;
if (SCAN_PROGRESS.peek() !== null) return;   // <- this one

So the resume assertion needs a tick to land in a 1000ms window and clear all three guards. SCAN_PROGRESS is module-level, and an earlier test in the same file leaks a load that writes it.

exitTimelineMode (line 297) starts a live-HEAD reload and only asserts the stream opened:

exitTimelineMode();          // -> void loadSource(...), never awaited
await flush();
expect(StubEventSource.instances.length).toBeGreaterThan(0);

That loadSource outlives the test. pumpManifestStream writes SCAN_PROGRESS.value = {...} per stream event (useManifestSource.ts:100) and clears it only in finally (line 214). The flaky test's beforeEach nulls SCAN_PROGRESS, but the leaked stream writes it non-null again — and if that lands inside advanceTimersByTimeAsync, the guard fires and no fetch happens.

Load-dependent, which is why it survives isolation and only shows under a loaded CI run.

Fix

Make exitTimelineMode's test finish what it starts — drive the stub stream to completion and await it — so no async work outlives the test. Not a retry wrapper, and not a wider timer window: both just lower the odds without removing the leak.

Worth checking the other void-dispatched loads in the file for the same shape while in there.