feat(quality): add evidence-backed task review and bounded reviewer evaluation

[adithya-s-k]

Task reviews can miss an inactive grader or accept forged test reports even when the reference and baseline checks look correct. This adds an opt-in evidence-review-v2 profile that preserves grounded findings, records the exact evidence and reviewer identity, and derives its disposition from supported defects.

The existing reviewer remains the default. This is an independent implementation inspired by micro1's Flow-QC article, not a reproduction of its private code or reported accuracy.

Included

• Seven evidence-backed criteria, stable finding IDs, attribution and investigation records; existing quality repair/rollout integration is preserved.
• A private, family-aware benchmark with paired profiles, immutable inputs, provisional/adjudicated labels, heldout gates, costs and unresolved-result denominators.
• Feedback recording and metered second opinions without silently changing task labels.
• Bounded OpenAI/Sonnet guidance search: equal model allocations, development-only inputs, no per-model regression, stock on ties, source pinning and rejection of exact benchmark literals.
• A standalone isolated-value worker and a reproducible migration example for the confirmed Accelerate report-forgery case. This does not automatically migrate arbitrary pytest suites.
• RFC 0033, CLI usage, pipeline diagrams, canonical prompt reference, credit and measured limitations.

Validation

• All 11 GitHub CI checks pass on efde0d53, including Python 3.12–3.14, Windows, docs, runtime contracts and distribution builds. 379 focused local quality, compatibility, migration and pipeline contract tests pass; Ruff, wheel build and documentation checks pass (11,016 internal links).
• Expanded comparison: 20 HF ML families, 40 cases, 80 scheduled attempts each for GPT-6 Astra and Sonnet 4.6. Evidence v2 located 18/20 and 8/20 planted defects respectively; practical v1 located 6/20 and 0/20. Incomplete reviews remain counted. Original labels are provisional, so blocked originals are not asserted false positives.
• Modal: ten isolation checks reject the confirmed forgery and other invalid controls while accepting reference/alternative implementations. Full Harbor oracle/nop controls return 1/0 for the migrated task and the data-only calibration task.
• The initial calibration image omitted its entrypoint; full Harbor execution caught the defect despite two model source audits. Failed attempts and corrected revisions are retained.

Scope and remaining promotion gates

Static local inventory found the same report-trust pattern in 650/1,330 local research/Tasksmith exports, including all 50 HF ML exports. This is exposure evidence, not 650 executed exploits or current Hub verification. Published datasets and acceptance labels are unchanged. The isolated example has an observable-value contract and does not prove arbitrary Python object identity.

The 20-family independently audited heldout gate is unmet. The prompt-search pilot completed stock and OpenAI-guidance comparisons equally on both models, then stopped a partly started Sonnet-guidance evaluation because its proposal copied an exact benchmark marker. No guidance was selected or promoted. The shipped pre-evaluation guard now rejects exact defect quotations; the stopped run and its uncertain requests remain recorded.

Final pilot ledger: $70.868357 accounted + $27.10 held = $97.968357 committed against the unchanged $150 cap. No paid controllers or validation sandboxes remain running. This PR is ready for review as an experimental, opt-in feature; it does not claim default-promotion readiness or a fully audited 20-family accuracy benchmark.