Add adaptive harness receipt eval

[uncfreak1255-code]

Summary

• add an eval-only five-field adaptive-harness receipt for the DF-03, DF-08, and fuzzy-task failure shapes
• add fail-closed custom-eval, overlay, strict-case, provenance, and frozen-score-floor support to the benchmark path
• update cold-start continuity so future agents recover the rejected candidate and evidence hold instead of repeating this task

Decision

The focused three-sample run passed with +0.4250 weighted delta, 0.9750 candidate score, zero boundary violations, and all DF-03/DF-08 candidate samples strict-pass. The retained sealed run scored 0.9281, below the frozen v0.2.0 floor of 0.9673, so the behavior candidate is rejected and the shipped skill remains unchanged.

Proof

• npm test — pass
• npm run benchmark:adaptive-harness — pass; focused receipt at results/2026-07-15T00-11-39-008Z-gpt-5.5/
• npm run benchmark:sealed — intentional failed decision gate; relative comparison passed, absolute 0.9281 < 0.9673
• npm run benchmark:trajectories — pass, 6/6
• npm run smoke:cold-start — pass
• Codex autoreview — clean, no accepted/actionable findings

No global install, hook, permanent agent, skill-text change, deploy, or runtime mutation is included.