fix(#709): anchor the resume/reaper test on the store's own clock

[Weegy]

Closes #709. Follow-up to the CI-flake work in #703 / #708 — different mechanism, same class: a test whose pass/fail depended on how long the test process itself took.

The race

a RESUMED task survives the reaper: provideInput resets the liveness clock aged a parked task by sleep(120), resumed it, then swept with staleAfterMs: 60 and no now — so the reaper used real time.

That left a 60ms budget between provideInput returning and the sweep running. Any stall longer than that aged the freshly reset heartbeat past its own window, the reaper failed the task, and the test went red while the code under test was correct. The suite runs against both stores, so it exposed the normal npm test run and the schema (migrations on pgvector) job.

Reproduction — deterministic

An 80ms stall before the sweep, standing in for a GC pause or a loaded runner:

	in-memory store	real Postgres
before, +80ms stall	✖ the resumed task is NOT abandoned	✖ same
after, +80ms stall	✔	✔
after, +3000ms stall	✔	—

The 3s case is the point: the race is eliminated, not widened.

The fix

const resumeAt = Date.parse(afterResume.updatedAt);
const result = await store.reapOrphans({
  now: new Date(resumeAt + 30),
  staleAfterMs: 60,
  purgeTerminalAfterMs: HOUR_MS,
});

provideInput writes updatedAt and lastHeartbeatAt in a single statement in both stores (last_heartbeat_at = now(), updated_at = now() in SQL; lastHeartbeatAt: ts, updatedAt: ts in memory). So updatedAt is stamped by the same clock as the heartbeat it is compared against — the database's for the durable store, the process's for the in-memory one. That matters: reapOrphans computes its cutoff in JS while the durable store's rows are stamped by Postgres, so an anchor invented in the test process would not be in the row's clock domain. This one is.

Why not the obvious alternatives

• Lengthen the sleep / widen the window — only makes the race rarer and slows the suite. The budget still exists.
• Pass an arbitrary now — breaks the cross-clock property above.
• Drop the reaper assertion — loses the regression coverage this test exists for.

The anchor is not a tautology

updatedAt was chosen precisely because it stays correct when the guarded behaviour regresses: remove the lastHeartbeatAt reset and updatedAt still advances, so the frozen pre-park heartbeat lands outside the window and the test goes red.

Mutation-checked both ways:

Store	mutation	result
in-memory	drop lastHeartbeatAt: ts from provideInput	✖ test red
Postgres	drop last_heartbeat_at = now() from the resume UPDATE	✖ the resumed task is NOT abandoned

Both restored afterwards; only the test file and the changelog are in this diff.

Verification

Gate	Result
npm run test	green ×2, 6459 tests, 0 fail, 0 cancelled
pg conformance (real pgvector, throwaway DB)	15/15
npm run typecheck:test	baseline 406, no regressions
npm run lint	clean
core decoupling ratchet (#470)	held at 3294
PG_TEST_FLOOR	unchanged at 269 — no pg tests added or removed

The Postgres run used a scratch flakeprobe database created and dropped inside the existing dev container, so no developer data was touched.

No lengthened sleeps, no retries, no timeout bumps, no skips.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}