VUH-2460 — landing keeps timing proofs for release

Tracks: https://linear.app/vuhlp/issue/VUH-2460

The six issue-named timing-sensitive files are listed once in scripts/testing/landing-test-policy.ts. They are on main at bd1ab027bc3370f4f094a28ffa04f2ebbbe0d600. The shared Vitest configuration applies the landing policy to parallel and serial projects. Plain pnpm test and release checks retain every moved file. Existing assertions, deadlines, product defaults and eval boundaries are unchanged.

Verification

Real native Vitest discovery proves all six remain in manual selection and leave landing, while cheap runtime-canary, fleet-refresh and remote-queue controls remain. Eval and opt-in computer tests stay excluded in both modes. All 13 discovery/revalidation cases pass; the typed policy also passes its settings import compiler and native discovery test. Discovery executes no eval. The actual six-file candidate's exact-tree root gate is green: 28.15 s, stable source, settings compiler passed. A preceding JS declaration error was fixed with a typed policy; no compiler check was weakened. Full logs are retained in same-change-comparison.json; initial red/green discovery is discovery-proof.json.

Same feature, before and after

The after-policy proof places the policy in its base, then applies the exact original VUH-2462 feature patch. Git patch ID is identical: 9dfb08d05509342b706bb2b2fbe5016fa32626f4. Product/test source under apps/clankie, apps/tui and packages is byte-identical. Comparison refs are local proof refs, never pushed. The actual policy candidate differs only in evidence docs.

The only selected-file difference is fleet-lead-round (27 cases). Before: 652 pass/8 skip after one supported runtime-terminals retry. After: 625 pass/8 skip after one supported cache-reclaim retry. The latter correctly kept an unverified cache when native process census failed with errno 1; its unchanged isolated case passed. The full gate saving is largely warm compiler cache, not all caused by this policy. The observed first-test phase difference is 40.86 s; load and different retry conditions prevent a controlled causal speedup claim. Raw gates, selected files, retries and exact patch checks are in same-change-comparison.json.

Measured history decision

The read-only snapshot has 722 history records: 533 gates, 272 nonzero outcomes, starting October 10. A bounded metadata scan found 138 completed, deduplicated gate reports; 15 identify file failures, including recovered attempts, across 13 files. This does not establish complete two-week attribution. The classified failures, source references and observed slow-file samples are in history-review.json.

No additional exclusion or transitive integration cap is justified by this partial evidence. Runtime-provider has a contract assertion and unresolved cleanup error; Codex has a mock import failure; provider, bridge, power and SSH assertions are not established flakes. A Discord timeout shares a HEAD with passes but has different uncommitted source fingerprints. Runtime-terminals has a supported recovery but a normal median of 1.584 s and protects pinned routing. Native process-census refusal is a safety result. Keep those checks. Recent remote-queue regression coverage and sparse slow-file samples also do not prove absence of useful catches. The roster lifecycle regression required the real source fix bc4025758; preserve its broader coverage.

Missing per-file history remains unknown, not a flake or proof of no catches. The app's stale clankie-ops baseline is setup drift and needs no test removal. No full release suite was rerun: native discovery proves unchanged manual/release selection. No eval, release, deployment or live retirement was performed.