[finding] lint-startup-registry-verdict.corpus.test.ts sweeps 1872 files / 28.1 MB twice inside vitest's default 5s per-test timeout — ejected #10733 from the merge queue at 9.1s under queue load

[os-elon]

What happened

Merge-queue CI run 32481624229 (2026-08-21 ~12:23Z, entry for PR #10733) failed on Test Core (4/6):

FAIL src/lint-startup-registry-verdict.corpus.test.ts > startup open-vocabulary verdicts across packages/ (#4776) > no package records a verdict the boot can still contradict
Error: Test timed out in 5000ms.

Actual duration 9144ms (file total 9731ms); 1 failed / 2250 passed on the shard. PR #10733 was ejected from the queue; its diff is describe/message prose in packages/spec — it does not touch packages/lint. The PR passed on requeue, consistent with a near-threshold timeout, not a regression.

Why this is structural, not incidental

packages/lint/src/lint-startup-registry-verdict.corpus.test.ts (#4776, the vocabulary gate):

• The corpus is every non-test .ts under packages/ — on the current main that is 1,872 files, 28.1 MB.
• The sweep (synchronous readFileSync of the whole corpus + findStartupRegistryVerdicts over each file) runs twice: once in no package records a verdict the boot can still contradict and again, from scratch, in no ledger entry is stale. Nothing is shared between the two cases.
• Each case sits under vitest's default 5000ms per-test timeout — no testTimeout override anywhere in the file or the package config.

So the time budget is fixed while the work grows with the repo and the wall-clock varies with shard load. Queue builds run the full suite on loaded runners (this shard reported import 106.30s); 9.1s for a 28 MB synchronous read-and-scan there is unsurprising. The margin only shrinks from here.

Context: this is currently the only unfixed repeat-risk queue signature

A sweep of all 262 completed merge-group CI runs in the 24h window ending 2026-08-21T12:45Z: 256 success, 6 failures (2.3%), every one PR-independent (all six PRs passed on requeue). Of the six:

• 2× scripts/gen-sdui-manifest-collision.test.ts port-collision harness — the only repeated signature, already fixed by test(spec): let the sdui collision harness own the port it calls busy #10456 (landed ~00:11Z, no recurrence since).
• 1× this timeout.
• 1× @objectstack/metadata-fs test/watch-dot-root.test.ts watcher-count race (expected […] to have a length of 3 but got 4) — single occurrence.
• 1× corepack ECONNRESET downloading pnpm (infra, pre-build).
• 1× @objectstack/example-showcase#test, undetermined (grouped turbo log truncated) — the same run PR fix(ci): teach the completeness guard turbo's failing-task header, so a red shard's own package stops being the one it cannot grade #10466's completeness-guard work analyzes.

Each ejection also rebuilds every queue entry behind it (e.g. pr-10358 ran CI 5×), so a single flaky signature multiplies full-suite load. Precedent for the class: #6855 (ordering assertion failing under merge-queue load, ejecting unrelated PRs).

Possible shapes for a fix (for whoever picks this up — not prescribed)

1. Hoist the corpus read + sweep into a shared beforeAll (or module-level) computation so the 28 MB scan runs once per file, and both cases assert over the cached findings. Halves the work and moves it out of the per-test budget (beforeAll gets its own hook timeout, which can be set explicitly).
2. And/or declare an explicit per-file testTimeout sized to the corpus (with a comment tying the number to the corpus size so it reads as a decision, not a magic constant).
3. Keep the two false-green refusals the file is built around (non-empty corpus, the sweep can still fire) intact — nothing here argues for weakening the gate, only for paying its cost once and budgeting it honestly.

Filed unassigned per findings discipline. The merge-queue triage bot will auto-open a cross-PR aggregate if a second distinct PR hits this signature; this card exists because the structural cause is already diagnosed and doesn't need a second ejection to be actionable.

[huangyiirene]

Triage first-touch: promoted to queue, type Bug — this is the only unfixed repeat-risk merge-queue signature in the 24h sweep, and the merge queue is a shared serial resource: one ejection rebuilds every entry behind it, so this taxes every lane. Scope per the card's shapes 1–3: hoist the corpus read + sweep into a shared beforeAll/module-level computation (own hook timeout, set explicitly), and/or an explicit testTimeout whose number is tied to the corpus size in a comment; ⛔ keep both false-green refusals (non-empty corpus, "the sweep can still fire") intact — nothing here weakens the gate, it pays its cost once and budgets it honestly. Bug priority in selection order applies. Size/model suggestion: S, sonnet.

---

Generated by Claude Code

[claude]

PM claim — dispatching now.

• branch: claude/issue-10838-corpus-sweep-budget

Dispatched as a bug with a measured cost, not a tidy-up: this card is the only unfixed repeat-risk queue signature in its own 24h sweep, and an ejection from the merge queue costs every PR behind it a full rebuild. That price is running today — a separate measurement this seat took at 16:14Z found 26 of 26 merge-queue failures in a 2.5h window, so queue capacity is already the scarce resource.

⚠️ The agent is told explicitly that raising the timeout is not automatically the fix: the card's own diagnosis is that the budget is fixed while the work grows with the repo, so widening the budget buys time and hides the trend. Removing work (the sweep runs twice, sharing nothing) is preferred, and any timeout change has to be justified with a measurement rather than picked.

No file conflict: packages/lint/** is untouched by the four other cards in flight.

• session: c970724d-303c-5614-9d20-a3f92205cfad
• seat: domain:devx

If an earlier claim comment with a different session ID appears above this one, that claim wins and this one is void.

---

Generated by Claude Code

[claude]

os-dev-report

Dev report — issue #10838 (draft PR #10911)

{
 "issue": 10838,
 "status": "done",
 "branch": "claude/issue-10838-corpus-sweep-budget",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/10911",
 "premise_still_valid": true,
 "summary": "REMOVED one of the two full corpus sweeps: the read plus TS-parse of every non-test .ts under packages/ (now 1,880 files / 28.15 MB) ran once inside each corpus case, sharing nothing, and is now paid once in a beforeAll both cases read. WIDENED nothing per-test: the sweep's budget moved off vitest's implicit default 5000ms per-test timeout onto an explicit, commented 60s hook timeout, and the largest thing still measured against the per-test budget fell from 972ms to 2ms while the file's `tests` total fell from 1441ms to 816ms (43% of the work gone). The 60s is measured, not picked: local cold sweep 986ms and warm 465ms/440ms, the ejecting shard ran that same cold work at 9144ms (~9.2x local), and vitest's DEFAULT hook timeout of 10000ms would have cleared that run by only 856ms, so hoisting alone would still have ejected #10733; 60s is ~6.5x the worst wall-clock ever observed for this sweep and is commented as a liveness backstop, deliberately NOT a performance tripwire, because the corpus grew from 1,872 to 1,880 files in the four hours between the card and this branch. Sharing is pinned rather than asserted: the findings array and every finding in it are frozen, the shared value starts undefined behind a corpusFindings() accessor that throws, and four ablation legs prove each case still reds individually. Both original false-green refusals are untouched; the file now documents a third.",
 "tests": "All runs via scripts/pm/os-verify-lock.sh, verdicts read from each tool's own printed line (never a bare $?). FILTER CONFIRMED: every run echoed `@objectstack/lint@17.1.0`, and `pnpm --filter @objectstack/definitely-not-a-package test` was run once as a negative control -- it printed `No projects matched the filters` and EXITED 0 (#10853), so the readings below are confirmed rather than assumed. BEFORE (pnpm --filter @objectstack/lint test src/lint-startup-registry-verdict.corpus.test.ts --reporter=verbose): 'no package records a verdict the boot can still contradict 972ms', 'no ledger entry is stale 467ms', tests total 1.44s. AFTER, at the committed HEAD b97a70c9a7: those two cases read 1ms and 0ms, all four green, 'Test Files 1 passed (1) / Tests 4 passed (4) / Duration 1.37s (... tests 854ms ...)'. ISOLATED SWEEP COST (tsx, this worktree): 'corpus files=1880 / sweep#1 (cold, includes lazy typescript load): 986ms / sweep#2 (warm): 465ms / sweep#3 (warm): 440ms'. CORPUS RE-DERIVED: files=1880 bytes=29516906 (28.15 MB) vs the card's 1,872 / 28.1 MB at 12:23Z -- it grew 8 files in ~4h, which is the trend the card is about. WHOLE PACKAGE: pnpm --filter @objectstack/lint test -> 'Test Files 80 passed (80) / Tests 2246 passed | 5 skipped (2251)', exit 0. TYPECHECK: dependency closure built first (pnpm --workspace-concurrency=2 --filter '@objectstack/lint^...' build), then pnpm --filter @objectstack/lint typecheck -> exit 0, no 'error TS'. ABLATION -- predictions written to disk BEFORE any leg ran (PREDICTIONS.md, 16:59:56Z). NO REBUILD IS INVOLVED and that is load-bearing to state: the test imports './lint-startup-registry-verdict.js', which vitest resolves to the sibling .ts in src, and the corpus is read from the filesystem at run time -- there is no dist/ leg that could hold a stale mutation. Each leg was mutate, PROVE ON DISK BY MARKER COUNT, run, restore, prove absent (--absent) by marker count; every python edit script exits non-zero on an anchor miss rather than silently writing nothing. LEG A (plant a #4771-shaped start() verdict into the corpus; predicted: only case 2 reds) OBSERVED: '× no package records a verdict the boot can still contradict', both rules emitted with path and line, '✓ no ledger entry is stale', 1 failed / 3 passed. LEG B (one LEDGER key no live finding produces; predicted: only case 3 reds) OBSERVED: '× no ledger entry is stale -- the site is fixed, delete the line: packages/ablation/src/gone.ts:1::startup-open-vocabulary-verdict', '✓ no package records a verdict...', 1 failed / 3 passed. A and B together are the proof that neither case silently stopped testing off the shared array. LEG C (drop the sweepResult publish in beforeAll; predicted: both corpus cases red LOUDLY, not green) OBSERVED: both red with 'Error: the corpus sweep did not run -- this case would otherwise report a clean audit over a corpus it never swept (#10838)', 2 failed / 2 passed. LEG D (with leg A's plant present, mutate findings[0].message; predicted: frozen object throws) OBSERVED: 'TypeError: Cannot assign to read only property message'. Worth flagging: TypeScript did NOT catch leg D statically -- the accessor returns a readonly array whose ELEMENTS keep mutable props -- so the runtime freeze is what actually holds the invariant. RESTORE PROVEN BYTE-IDENTICAL by regenerating the file deterministically from the pristine HEAD copy and comparing sha256: S1 (post-ablation restore) = S2 (regenerated) = 5529ce217d459212b811db26fbec78e9395460b0dc3ce94b1228224e7d910db9; git status --porcelain then showed only the one intended file and no untracked ablation residue. GATE UNION re-derived AFTER the final commit with `node scripts/pm/dispatch-gates.mjs` (no hand-built path list) and run at b97a70c9a7 -- 12 families, all green, each quoting the gate's own verdict line: cross-package-test-inputs ('OK: 13 package(s) read outside themselves, all declared' + 104 self-test cases), slot-lookup ('✓ ratchet holds: 107 unswept site(s)... none new'), test-source-alias ('OK -- 72 packages with tests scanned'), type-source-resolution ('OK -- 76 packages... scanned'), check-ci-filter-parity ('OK: all 82 declared cross-package glob(s)'), check-plugin-teardown-shape ('✓ ... 57 Plugin implementation(s) across 4408 source(s)'), check-affected-docs ('✓ affected-docs self-test: 323 cases pass.'), query-options-erasure ('✓ ratchet holds... baseline key set verified against 409077e: no files added'), type-check-coverage ('OK -- 64/77 workspace packages type-checked'), engine-double-contract ('OK -- 376 pinned, 133 in the DEBT ledger, 2 exempt'), where-matcher ('✓ conformance holds: 275 matcher(s)... none new'), check-nul-bytes ('OK (scanned 6301 text file(s)... no raw ASCII control bytes)'). DECLARED NARROWING: `check:type-check-debt --re-measure` was NOT run locally -- it refuses to measure on an unbuilt worktree and needs a whole-workspace turbo build, which would hold this container's shared verify lock for a long time with three other agents queued on it. Its trigger here is 'a new test file that does not typecheck cleanly'; this PR adds no new test file, edits one that already typechecks, and the package typecheck is green at this commit. CI runs it regardless. RELEASE DECLARATION: no changeset -- the change is confined to a *.test.ts that tsup never builds into dist and package.json#files never publishes, so nothing releases. `skip-changeset` applied ADDITIVELY via the REST POST (it worked from this agent seat, confirming #10902's correction of os-dev.md's whole-set-PUT premise) and READ BACK: ['size/s', 'tests', 'skip-changeset'] -- the labeler bot added 'tests' after my write and my label survived. PR body byte-check: 9263 sent, 9262 stored, sole difference the stripped trailing newline; the session-URL footer survived. No other agent has touched packages/lint since my merge-base (git log 409077e93c..origin/main -- packages/lint/ is empty at origin/main 58563be069).",
 "open_questions": [],
 "out_of_scope_findings": [
 "filed as #10913: 'no ledger entry is stale' in the same file is green by construction while LEDGER is empty -- Object.keys({}) makes `stale` always [] for every corpus and every findings set, so that case never consults the sweep and is the one case in the file with no standing non-vacuity proof (the existing 'the sweep can still fire' proof covers the sweep, not the staleness comparison). My leg-B ablation shows it CAN fire, but that proof lives in a PR body and nothing re-runs it. Filed unassigned, labels finding+tests, not fixed here: different defect class from this card, and the repair is a shape decision. Searched the 41 open tests/finding issues first; no prior art."
 ]
}

---

Generated by Claude Code