droboticflow/docs/verification/predeploy-gate-repair.md

3.3 KiB

Unified candidate release gate repair

Base: codex/recent-integration@64d7400c. This change repairs test environments and release checks; business handlers, permissions, source admission and customer data are unchanged.

Failures and corrections

  • Dashboard race run originally measured 121 concurrent synthetic cohort reads at p95 9238 ms; the same non-instrumented scenario measured 706 ms. Timing under race instrumentation was incorrectly treated as the production timing target. The release command now requires both the race suite (including concurrent correctness and source-read bounds) and explicit -race=false timing. The 1 s sequential and 2 s concurrent thresholds remain enforced without instrumentation. Compile-time race flags prevent accidentally accepting an instrumented timing run. Both required test names must execute; missing, skipped or failed phases fail the gate.
  • Dashboard browser runner still expected the replaced annotation-output endpoint and old layout controls. It now runs the existing production-quality browser suite against its own random-port in-memory fixture: five roles, supplier/person isolation, project filtering, role-switch snapshot identity, CSV, mobile layout and zero unexpected writes. Old layout tests are retained for the legacy component, not presented as coverage of the new entry point.
  • Rework browser fixture had no source validation processor, so admission correctly kept the page read-only while validation was pending. It now runs the real processor over generated local video/PTS/IMU. After this correction a stale 重做 locator failed; it now uses the approved 恢复 name. No timeout or rapid-history assertion was relaxed.

Final evidence

Logs: local-data/verification/predeploy-quick/.

  • stability-final.log: race and timing phases both passed, no skipped required cases. 121 synthetic warm reads over 5000 parts: p95 648 ms, maximum 689 ms, zero source rebuilds. This is not 120-person full-work capacity acceptance.
  • dashboard-final2.log: five roles passed; no browser errors or unexpected writes. Includes current project filter and CSV checks.
  • rework-final.log: real-browser persistence suite passed (41.56 s). Covers legacy recovery preserving newer draft, naming, boundary edits, archives, reload, rejected/pending/lost writes, immediate undo/restore, and description edits. History request starts remained under the existing one-second bound.
  • release-final.log: 51 release unit tests passed, including failure of either performance phase blocking the gate. Format, script syntax and Go formatting passed.
  • Other unchanged integration checks from this session: source recovery and failed-generation synchronization browser passed; authorized review pool search browser passed; build/web boot, permission, decisions, drafts, slice input, pool and delivery/worker/API checks passed. Frontend and Go broad gates had only their known failures (13 and 68 respectively), with no new failures. No known-failure baseline was edited.

No push, PR, image upload, deployment or production write was performed. The extra validator/Worker/normalizer release remains subject to the concrete component plan in docs/release/recent-integration-components.md; no release record or immutable candidate images have been prepared. 120-person full work, eight-hour soak and node failure remain unaccepted.