[raiseCatError]

Product direction

Audit the repo, not the summary. Mutation quality + bounded autonomous repair are strictly opt-in extensions to the read-only freview auditor. Stacked draft PR #20 on draft #19; parent drafts #18, #17 and #16 remain unmerged. No automatic merge, push, tag or release.

Implemented (stdlib-only)

• Mutation quality: bounded Stryker mutation-testing-elements JSON and mutmut CI-stat reports, detected vs undetected mutants, incomplete/pending/no-valid classification, survivor hints, score threshold and minimum valid-mutant gate.
• Trustworthy execution: an exit-zero mutation command alone is inconclusive. The report must be newly generated by the current isolated verifier; stale committed reports, symlinks, oversized, malformed and interrupted reports cannot pass. Standalone run and pipeline follow the same rule.
• Real mutmut format: upstream mutmut's total may include classifications not exported by its CI JSON. The unclassified difference is surfaced as incomplete, never guessed. Score threshold uses unrounded counts.
• Integration: optional mutmut_driver.py executes an already-installed mutmut run then CI-stats export without a shell or installation. Stryker needs a separately installed engine and its JSON reporter. Fake fixture tests only; no real mutator invoked.
• Autonomy: configurable global max subprocess stages (default 12/max 32), max 3 review cycles, stop-on-no-progress (default true), repeated-finding stop, and optional mutation-quality shipping requirement. Shipping still only prepares human handoff.
• Recovery: atomic scratch checkpoints now include mutation quality, evidence-freshness and exact verifier identity; no automatic resume or signed attestation.
• Defaults: every operation module remains OFF unless explicitly enabled; quality mode is advisory, min score 60, minimum 1 valid mutant, maximum 12 survivor hints, require_mutation_for_shipping=false. All policy fields strictly validated.
• Updated the portable two-OS CI, skill layout assertions, default config JSON, SKILL, README and operations manual; no code copied from third-party tools.

TDD / verification

• Initial red tests before implementation
• Stryker/mutmut source parser and stale-report fix, both OS
• Standalone mutation status and legacy compatibility, both OS
• Fake mutmut CLI adapter, both OS
• Checkpoint mutation quality schema, both OS
• README / skill and example policy, both OS
• Final exact-head CI passed on Ubuntu and macOS at d7e991f616ccd3231f667c6a230ac4c349813e68. This is automated source/test verification, not independent real-agent/mutator acceptance.

Docs / recovery

• Implementation specification and checkpoint journal
• Mutation/autonomy tuning and usage guide
• Bounded freview-on-freview self-review

Remaining human gates

• Native mutmut and Stryker real-project mutation runs; interpret survivor usefulness
• Real Claude/Codex/OpenCode autonomy review, account and permissions; physical external scratch / Mac tests
• Human review of stacked PRs; decide whether actual authenticated shipping integration is warranted
• Explicit owner authorization to merge/release; no automatic publication or mutation of main

Passing fake Git fixtures does not prove a reviewer independently read source, a mutation score is correct on production code, an OS sandbox exists, or a production release is safe.