ci: split slow Verify jobs with exact file coverage and strict summaries

[Ancienttwo]

Verify can exhaust the existing 20-minute macOS / 60-minute Windows caps while tests keep passing. Split only the four measured timeout targets—macOS15/Node24, macOS15/Node25, Windows/Node22.22 and Windows/Node25—into three isolated whole-file test jobs. The other five targets retain complete verifiers.

This CI delivery is independent of #351 / #354. #354 remains frozen at 468df532; no product commits are included. Base: d6788c4b565aa043530b2f6407234b3ed108fa00. Review HEAD: d0fcb921b5fe088cf4406ad7280d5bf441170236. Keep Draft; no merge or serial integration yet.

• First-round evidence is retained: 38637184856, HEAD a10fafde, ended with 27 successful jobs, macOS25 and Windows22 execution timeouts/cancellations, and two correctly failed summaries. Neither cancelled execution log contained a (fail) marker. Windows22 finished main tests (3104 pass / 5 skip / 0 fail, 3352.16s), explorer and native smoke; the full packaged CLI smoke and later gates were unfinished. macOS25's main suite did not finish. This was not all-green.
• Cost: first round measured 282.87 raw runner minutes, 302 in the per-job rounded model (not an invoice), and 19 installs / 6.68 minutes; the extra eight shard/post installs consumed 3.87 minutes. Four-target topology now has 25 execution jobs, 27 installs and 39 total jobs. The additional eight installs model 2.93 raw minutes at the observed 13s macOS25 / 31s Windows22 rates. Provisional total planning is 295–310 raw runner minutes, including an explicit allowance for more setup and previously unfinished work; it is not a proven bound or cost-saving claim. Actual second-round totals remain pending. No Governance reuse savings are assumed.
• Order/state: pre → all shards → exact coverage proof → post. Local bun run verify remains byte-for-byte unchanged. Post reinstalls dependencies and rebuilds its native bundle/Git/daemon fixtures. Only run-bound discovery/inventory/JUnit metadata crosses machines; tests keep real isolated processes, protocols and every product assertion.
• Discovery: pinned Bun supplies the actual current census, including empty files; weights never whitelist paths. Added/deleted paths, spaces, duplicate basenames and empty shards are checked with real Bun fixtures. Fresh native shard inventories must cover every discovered file exactly once for the same run/attempt/checkout.
• Summary: always() with current-attempt paginated job evidence; every expected worker must uniquely be completed/success with matching subject. Missing, skipped, cancelled, failed, duplicate or stale jobs fail. All nine FG6 summary names, IPC artifact semantics, permissions and caps remain unchanged.
• Governance is unchanged. The runbook records its 15-minute pressure and a deferred concrete receipt/dedup proposal. That authority change is not approved or implemented in this push; fix(runtime): preserve selected failure metadata (#328) #356/D6-A is untouched.

Second-round delta is limited to .github/verify-sharding.json, scripts/verify-sharding-workflow.test.ts and the active runbook. The approved policy expectation expands from two to four targets. A real topology CLI regression checks four pre/post targets, twelve shards, five unsplit workers and 25 execution jobs; strict rejection cases cover each expected worker.

Validation: 27 focused tests pass, 0 fail, 344 assertions; typecheck, package-boundary audit and diff checks pass. Workflow/controller/package bytes match the first candidate; actionlint1.7.12 passed on that unchanged workflow. The original native discovery census was 260 files, including 3 absent from old weights and excluding 1 deleted weighted path. No product/protocol assertions or historical verification artifacts changed.

Second-round CI: Verify 38647011093, queued when created; budget and results are pending. Detailed state/dependency audit, first-round failures, measured cost and next-run assumptions: verify sharding runbook.

Coordinator acceptance (2026-10-14 23:10 UTC): exact HEAD d0fcb92, Verify 38647011093 attempt 1 completed successfully; all 39 jobs and all nine named summaries passed. Independent static and incremental reviews found no blocking findings. Observed total: 273.23 raw runner minutes, 298 summed per-job-rounded minutes (model, not invoice), 27 installs / 12.28 minutes; sixteen extra shard/post installs / 7.22 minutes. First-to-last job wall clock 48m45s. One successful run supports acceptance, not a stable cost-saving claim. Prior pending language above describes the state when the candidate was first published.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-14T23:14:16.414364Z	d0fcb92	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.