test: pin #240 fail-fast timing with real encode/httpx rows

[Gagaismo]

Summary

#240's assumption holds on real GitHub data: in a default fail-fast matrix, every sibling GitHub cancelled has a completedAt later than the failing job's startedAt. Two tests now pin those real rows, so vigilar and rebote keep reporting the failure. No source change.

Source (read-only gh pr view https://github.com/encode/httpx/pull/3796 --json statusCheckRollup): PR encode/httpx#3796, run 37009481166 (workflow Test Suite, matrix with no fail-fast key, so GitHub's default true, checked on the run's head 7aee79f4).

name	conclusion	startedAt	completedAt
Python 3.12	FAILURE	2026-10-02T12:53:52Z	2026-10-02T12:54:34Z
Python 3.11	CANCELLED	2026-10-02T12:53:52Z	2026-10-02T12:54:36Z
Python 3.13	CANCELLED	2026-10-02T12:53:54Z	2026-10-02T12:54:56Z
Python 3.10	CANCELLED	2026-10-02T12:54:17Z	2026-10-02T12:55:11Z

Earliest cancelled completedAt (12:54:36) is later than the failure's startedAt (12:53:52): the failure is classified as real. The same shape showed in the jobs of three more httpx runs read with gh run view <id> -R encode/httpx --json jobs (35071320524, 22379030162, 22365498386): in each, the earliest cancelled job completed after the failing job started. They are context only; the fixture uses PR #3796.

Evidencia

Rule, legacy RED item, or touched public entry	Test (exact name)	Result
1. Real data	Not a test: PR, run id and the four rows above, read with gh pr view 3796 -R encode/httpx --json statusCheckRollup	Recorded above
2. Fixture test (vigilar)	test/vigilar.test.mjs › Issue 256 Rule 2: real fail-fast rows emit ci-roja with the failing row	✔ Issue 256 Rule 2: real fail-fast rows emit ci-roja with the failing row
2. Fixture test (rebote)	test/despacho.test.mjs › Issue 256 Rule 2: rebote bounces real fail-fast rows as a failed check instead of asking for gh run rerun	✔ Issue 256 Rule 2: rebote bounces real fail-fast rows as a failed check instead of asking for gh run rerun
2. Tests catch a regression	Same two tests with #240's filter in src/checks.mjs reverted to [] (local mutant, not committed)	Both ✖; restored file, both ✔
3. If the assumption fails	Not applicable: the real rows satisfy #240's rule, so src/checks.mjs is unchanged	—
4. #218 rows unchanged	Issue 240 Rule 2: red row started after first cancellation completed is not reported and emits ci-cancelada and Issue 240 PRESERVE: reboteCommand retains rerun refusal when cancellation has no real failure (run 38843085964, 02:58:16 / 02:59:09)	Both ✔
5. Existing tests unchanged	All #225 and #240 tests in both files, untouched	npm run verify: ℹ tests 703, ℹ pass 703, ℹ fail 0; mods suites 0 fail

Touched public entries: none touched (tests only).

Merge Danger

Two-way door. Tests only; blast radius is the test suite.

Desvíos del ticket

ninguno

Implementador: claude
Closes #256

🤖 Generated with Claude Code