ci(coverage): give the instrumented lane its own per-test budget so the coverage gate can run

[claude]

Part of #9271. ⛔ Deliberately NOT Fixes — this closes the proximate half only; the structural half named at the bottom stays open on the card.

What was measured

⭐ The blind window is 116 consecutive ci.yml push runs on main — runs 17230 → 17738, 2026-09-11T23:17:11Z → 2026-09-14T07:09:07Z, 2 days 9 h 42 m. Test (coverage) is failure in 116 of 116. Last run whose coverage lane was green: run 17227 (a94e4d073, 2026-09-11T21:27:02Z), which also published the only coverage-report artifact in the window.

⛔ objectui#8479's 66-run streak is NOT this streak. Population: the 500 most recent branch=main, event=push runs of workflow 223145786 (runs 15611 → 17738, from 2026-09-07T06:45Z). It holds 221 successes, the newest being run 17227 — so the 2026-09-07 streak healed and this is a second, later one. ⇒ the card's written re-grade trigger ("longer than roughly a week") does not fire; priority:p2 stands on duration.

The break in the chain, read rather than inferred

⭐ The blobs are not missing. On red run 17738 all four coverage-blob-N artifacts exist and are live (7.8–8.7 MB each); the green run 17227 carries the same four plus coverage-report. blobs=skipped in the sentinel is the counting step being skipped, not an absent blob.

The break is that coverage-report gates every step on needs.test-coverage.result == 'success'. One timed-out row reds a shard, so checkout, download, blob-count and merge are all skipped — and the merge is the only place coverage.thresholds is enforced, because the shard legs override all four to 0. Verbatim from run 17738:

shards=failure blobs=skipped merge=skipped report=skipped — no complete merged report was
produced for db6aa19a94b4893f4e78d0120db7a571a3c033c6, so the coverage thresholds were
never evaluated on this commit

Instrumentation isolated as the single variable

The same-commit CI comparison could never settle this: test-coverage is if: github.event_name == 'push' and test is if: github.event_name != 'push', so instrumentation and trigger event are confounded by construction. Isolated locally instead — same tree, same container, same command, --coverage the only difference:

population: the 146 rows of the two heaviest sweep files, tree db6aa19a9	no coverage	coverage
rows exceeding the production 15000 ms budget	0	5
sum of all row durations	30.3 s	115.5 s (3.81×)
heaviest row	6.6 s	25.3 s

The five rows that blow the budget under instrumentation are exactly the five CI sites in those files, and the ordering corroborates: the marginal row (15.0 s locally) is the one that fires 11× in CI, while the 24–25 s rows fire 64–94×.

⚠️ The rows are NEW, not degraded — the corpus-growth hypothesis is false

Same instrumented command, same container, two trees, rows paired by name:

population: the 106 check-i18n-dead-keys rows present on BOTH trees	a94e4d073 (last green)	db6aa19a9 (today)
total	1.4 s	1.3 s (0.9×)
rows over 15 s	0	0

⇒ the shared rows did not slow down at all. All three of that file's timeout sites are rows that did not exist at the last green run, added 2026-09-12; and the single most frequent site overall, check-object-metadata-write-doors.test.ts, is an entire file that landed 2026-09-12T04:00Z. The tracked-file corpus grew only 3.2% (7492 → 7733) across the window, nowhere near enough to matter.

The repair, and what it does not weaken

A lane-scoped --testTimeout=60000 on the instrumented shard command — 4× the uninstrumented budget, 2.4× the heaviest row observed under instrumentation. ⛔ Not a per-row bump: the wandering site is exactly why it is applied to the lane.

• vitest.config.mts keeps testTimeout: 15000, so the uninstrumented lane — the one that actually runs before a merge — is untouched. gates_weakened: NONE.
• No job-budget risk: test-coverage has timeout-minutes: 40 against observed shard wall clocks of 698–1053 s.

Evidence the gate is restored

⭐ A green shard is not evidence a report was produced, so here is the report and the thresholds. CI cannot host this demonstration — the coverage lane runs only on push to [main, develop], never on a PR — so it is reproduced locally over the five CI timeout-site files:

leg	shard exit	merge	thresholds
--testTimeout=15000	1 (10 timeouts, 0 assertion failures)	never runs — gated on shard success	never evaluated
--testTimeout=60000	0 (379/379 pass)	runs	evaluated

The merged report from the passing leg, produced by ci.yml's own merge command:

Statements   : 75% ( 1924/2565 )
Branches     : 75.69% ( 1501/1983 )
Functions    : 74.83% ( 232/310 )
Lines        : 77% ( 1634/2122 )

⭐ Firing control that the thresholds are genuinely enforced and not merely printed — same blobs, same command, floor raised above the measured value:

$ pnpm test:coverage --merge-reports --coverage.reporter=text --coverage.thresholds.lines=99
ERROR: Coverage for lines (77%) does not meet global threshold (99%)
exit 1

⛔ The half this PR does NOT fix — it needs a decision above this card

The coverage lane runs only on push to main/develop. An author adding a 6-second repo-sweep row cannot see this lane before merging; the first signal is after the row is already on main, on a red that blocks nothing. That is why the class recurs — this file's own comment records objectui#8545 pricing two earlier episodes at 84 and 87 consecutive unevaluated pushes, and this one ran to 116.

⇒ A budget alone postpones the next episode, it does not prevent it. Making the lane visible pre-merge is the structural fix, and the card rules out making it blocking — so the fork is the maintainer's, not this PR's.

Verification

• pnpm exec vitest run --project unit scripts/__tests__/ — 157 files / 4625 tests passed, exit 0 (this is the surface holding every gate that reads ci.yml).
• node scripts/check-changeset-presence.mjs — exit 0, "no changeset is owed".
• node scripts/check-governed-queue-guard.mjs --test .github/workflows/ci.yml — NOT GOVERNED (control: AGENTS.md returns GOVERNED, exit 3).

🤖 Generated with Claude Code

https://claude.ai/code/session_013VGeMu3p6qEFWR6K6GGLaW

---

Generated by Claude Code