test(metadata-fs): bracket case 2's negative assertion instead of waiting 4s (#7408)

[os-zhuang]

Fixes #7408

Real file surface

File	Change
packages/metadata-fs/test/watch-dot-root.test.ts	Case 2's negative assertion rewritten from a wall-clock quiet window to an event-driven bracket; waitForEvent → waitForEvents + new waitForNames; QUIET_WINDOW_MS deleted; CASE_TIMEOUT_MS 120s → 180s

That is the whole diff — one file, test-only (+176 / −45). No source file is touched; packages/metadata-fs/src/repository.ts is byte-identical to main.

The defect

Case 2 proves a negative: the watcher must not report dot entries under the root. It proved it by sleeping QUIET_WINDOW_MS = 4_000 and asserting sink.events was empty.

#7369 measured delivery of a single external edit at 24–36s under event-loop starvation. Past that load level the array is empty because nothing has arrived yet — not because the noise was ignored. The assertion passes more reliably the more loaded the runner is, which is the one correlation a guard must never have.

Re-measured here against the real defect (isIgnoredWatchPath reduced to return false at the source, reverted before commit): the old assertion passed with the defect live in 9 of 13 loaded runs, and 5 of 5 when .cache/x.json was the only noise.

⛔ A bigger number is not the fix — every "sleep, then assert empty" shape has the same defect at some load level, and #7369 already measured what moving one fixed budget buys (the 10s hookTimeout immediately became the next binding one).

The shape

You cannot wait for an absence, so the case brackets the noise between deliveries it can actually observe:

• an opener — an external edit that must arrive, proving the watcher is awake before the noise exists. Nothing checked this before: a late-armed watcher swallows the noise into its own baseline and produces exactly the empty array a correct one does.
• two closers — each a new entry placed so the same directory read that discovers it must also enumerate one of the leaks: a new file in view/ for view/.scratch.json, a new type directory for .cache/x.json.
• the bracket closes on the closers' identity, never on an event count. With a count, a leaked event satisfies the count itself, so the run that should fail loudest is the one that stops listening soonest.

Reverse verification

Break: isIgnoredWatchPath → return false (the "let parseItemPath decide" variant the source docblock describes). Load: in-process event-loop starvation, BLOCKms blocked out of every BLOCK+GAPms. Base SHA 08363a0.

Probe	Shape	Source	Load	Predicted	Measured
P1	new	intact	none	green	✅ green
P2	new	intact	4000/5	green	✅ 5/5 green, case 2 40–44s
P3	new	broken	none	red	✅ red
P4c	new	broken	2000/5	red	✅ 8/8 red
P4d	new	broken	4000/5	red	✅ 9/9 red
P5	old	broken	none	red	✅ red
P6	old	broken	2000/5 + 4000/5	quiet assert green = false pass	⚠️ 9/13 false pass, but case red 13/13 — see below

Missed prediction (P6), with cause. I predicted the old case would go green under load. It did not, 13/13. What false-passed was the assertion this card is about — expect(sink.events).toEqual([]) passed with the defect live in 9 of 13 loaded runs (0/1 unloaded), 5/5 on the isolated .cache leak. The case was rescued each time by its trailing toHaveLength(1), an assertion documented as a liveness control, which caught the leak only because the leak (written ~24–28s earlier under starvation) happened to arrive before the control event. That rescue is an ordering accident, not a design — and it is #7369's deadline-driven half doing the work, not the quiet window.

Guards (green in BOTH directions), labelled as such in the test file:

• The .objectstack/.log/main.jsonl append cannot become a MetadataEvent even with the matcher removed entirely — parseItemPath rejects the .objectstack prefix. Measured: with the matcher broken, the leak set is exactly .cache/x + view/.scratch, never the log. It stays as the traffic the matcher exists to keep out of the poll set.
• Case 1 stays green under the break, confirming the file's premise that a fix which only widened the matcher would pass case 1 and fail case 2.

Residual, reported rather than hidden. view/.scratch.json's bracket is exact (one view/ read enumerates both) — caught 11/11. .cache/x.json's is tight, not exact: the root read that discovers dashboard/ enumerates .cache/ in the same call, but each is then descended into by its own async read and those can finish in either order. Caught 28/29 (9/9 at 2000/5, 19/20 at 4000/5). The single escape was a late delivery, not a lost one — draining 30s past the bracket found the event 6/6. The file records this and states why a sleep must not be used to close it; the direction that would is another event-driven round, traded against one more delivery inside the ceiling.

Unmeasured: whether the opener catches a genuinely late-arming watcher. That needs a different break (#7000/#7282 arming defects) than the matcher break used here; the opener's teeth against it are argued from ignoreInitial, not measured.

CASE_TIMEOUT_MS 120s → 180s

Not a quiet window — this is the deadline #7369 made the single source of every wait's budget, and it was sized for one delivery. Case 2 now makes three sequential ones. Measured at 4000/5 (heavier than any row #7369 ran): case 1 28–32s, case 2 40–44s, up to 72s when the broken matcher adds two events. 120s would have held at 2.6x margin; 180s restores the ~3.3x #7369 chose. A healthy run pays none of it — case 2 is 2.1s unloaded, down from 5.5s.

Changeset

skip-changeset label, not an empty changeset. This is test-only and releases nothing, so it is the textbook case for the label; an empty changeset would linger in the release pipeline. Label applied to this PR.

Gates

Gate	Result
pnpm check:nul-bytes	✅ pass (6839 files)
packages/metadata-fs suite	✅ 31/31, 5 files, 8.9s
pnpm typecheck (metadata-fs)	✅ clean

Per instruction, the whole lint farm was not run locally.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
objectstack	Ignored		Aug 10, 2026 3:03pm

[github-actions]

📓 Docs Drift Check

No hand-written docs reference the 0 changed package(s). ✅