ci: split the release-green sweep so a hosted runner can finish it

[LMPrado-DZ23]

Four consecutive full sweeps died at exit 143, and no full-CI verdict exists for release/v3.8.55. #76 blamed a missing timeout-minutes; #86 records the measurement that disproved that. This is the fix.

What is actually wrong

07:17:26  job starts
07:23:23  ✅ every static and drift gate is green
          ▶ the four serial suites start
08:17:37  ##[error] Process completed with exit code 143   ← 60m11s

A runner on this repository stops at ~60 minutes — run 35468579833 said so in words: "The runner has received a shutdown signal". The suites' own ceilings are 80 + 15 + 40 + 20 = 155 minutes serial. One job cannot carry that, and no timeout-minutes value changes it.

The shape

Job	Does
resolve	resolves the release branch once and outputs the exact SHA
slow-suite (×7)	unit ×4, integration ×2, vitest — each ≤55min
release-green	static + drift + full-ci + package artifact + boot smoke, then merges the shard reports and owns the verdict

Every job checks out the SHA resolve produced, so the merged report belongs to one commit instead of to whatever each job happened to fetch at its own start time.

The shards measure; they do not judge. A red suite still exits 0 so its report reaches the aggregator — and the aggregator runs on !cancelled(), not success(), so a shard that died still gets its verdict pronounced.

The part with teeth

Splitting a verdict across seven jobs creates exactly one new way to be wrong: a job reporting green while measuring nothing. That is the same defect class I have already fixed three times on this branch, so it is guarded up front, not afterwards.

• --expect-slow names every shard id the matrix produces — not just the suite names. With a bare unit, one surviving shard would satisfy the expectation and the other three could vanish. A suite with no report is recorded as a HARD failure reading "it did not run, so it is NOT green."
• --slow-gates=<typo> throws instead of selecting nothing.
• A run that records zero gates is a HARD failure for that reason alone.
• A test derives the expected ids from the matrix and compares the two. It fails closed: a stale list goes red rather than quiet.

Verification

Check	Result
tests/unit/validate-release-green.test.ts	47/47
red-first: drop 2 unit shards from the expect list	test 46 fails, as designed
--no-static --slow-gates=none	❌ ran zero gates — nothing was measured, so nothing is green
--slow-gates=unit-tests	refuses to start: unknown suite 'unit-tests'
--merge-slow without --expect-slow	refuses to start
merge smoke: 2 unit shards green, integration absent	❌ NOT release-green
YAML parses, all budgets < 60	✓
prettier --check	clean

I am dispatching the sweep against this branch to produce the verdict itself; I will post the result here rather than claim it in advance.

Honestly still open

main-green carries the same 155 minutes in one job and will keep dying at 60 minutes. It validates the released state rather than this release candidate, so it does not block this line — but it is the same fix and it is not done.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 2972d5fa-7b91-4d00-85e9-8776053c3caf

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[LMPrado-DZ23]

Pushed one more commit: main-green.

I said in the PR description that it was still broken and untouched. It is still broken, but leaving a comment there that described sharding as an unreachable future option — when the job right above it is now sharded — would have been the same kind of stale explanation #86 had to correct.

So it now says plainly that it cannot finish, why it is knowingly left that way (off by default; this fork ships from release/v* and never merges into main, so main is a stale upstream snapshot), and what the fix is when main starts to matter — the slow-suite matrix above.

The one real change: timeout-minutes 180 → 55. It will still fail if someone enables it, but it will fail saying so, with its log, instead of at an opaque exit 143. The budget test now covers all three jobs.

47/47 tests, YAML parses, prettier clean.