mrmrs-eec7ec

Closed into main

4 commits · 53 files +23 added ~30 modified ·Seen by zach 1 month ago

Description

Preserve measured agent failures and validate campaign copies before admission

Fix the fourth QA gate on mrmrs-eec7ec at fcd8a785ee1a. Scripted workflow summaries, including workflow.skipped, remain strict publication checks. Fully measured failed agent trials remain reliability evidence; trusted successes still require independent persisted-state evidence. Producer regressions exercise skipped workflow subjects, mixed success/failure agent results, and trusted worktree failures before HEAD validation.

Campaign fixture copying and validation precede the durable trial start. Copy faults leave all pending trials retryable without new journal event types. Cache files are read-only while directories remain owner-writable for deletion and rebuild; chmod no longer requires nofollow support. Mid-trial payload or record drift invalidates the measurement, retains rejected evidence, names the discarded trial on stderr and in reports, and stops admission. Completed resumes do not regenerate pruned fixtures.

Oracle calls pass resolved instrument identity through the read and reuse observed binary hashes, removing duplicate timed work. Evidence classes distinguish tests/status, repository metadata and failed committed-read attempts. Existing journal recovery, full timeout admission, exact payload checks and unknown-outcome rules remain.

This round changes only the named implementation files, their existing tests and the existing QA invariants document. The Low devloop transport-diagnostic heuristic/consolidation finding is deferred. Native Linux was not run; chmod portability is covered by the unsupported-flag regression. Same-UID cache permissions are deterrence, not OS containment. Independent Oak persistence, calibrated tokens and cache qualification remain open. No real-provider performance claim and no merge.

Evidence: parent qa5-* logs and docs/qa-invariants-2026-09-12.md. Author worker /Users/mrmrs/o/worker-mrmrs-20260911T191215-bench-regate-fixes/bench; unchanged parent main b297e8a25d08.

Full instrument suite: 1,102 tests, three skipped, OK in 197.279 seconds (qa5-tests.log).

Identical-binary default-threshold core: two repetitions REGRESSED on one latency-only check (+78.1%, threshold65.6%); preserved qa5-devloop.log. One eight-repetition/eight-AA follow-up PASS exit0 (qa5-devloop-8.log), no intervening code/threshold changes. High A/A noise remains; no speed claim.

Commits 4

2e081b95a118