feat(agent): strengthen headless contract verification

[minixalpha]

Headless runs now identify deliverables, interfaces, and constraints; verify output behavior through the real entry point with proportionate independent or boundary checks; diagnose and retest failures; and report observed evidence and remaining limitations. The prompt permits task-supplied source, executables, and tests while restricting externally obtained task-specific answers and hidden evaluation tests. Only HEADLESS_SYSTEM_PROMPT changes in runtime code.

The PR includes frozen experiment configurations, prompt snapshots, all trial results, reference-use audits, a tool-limit diagnostic, and synchronized development notes.

Experiment evidence

Measure	Main (6664f7b)	Frozen 368-word candidate (03cfd53)
Raw official passes	18/20	19/20
Common eligible tasks	17/19	18/19
Estimated model cost	$1.457520	$1.898565
Summed agent seconds	11160.416	13570.056

The candidate adds a regex-chess pass with no observed per-task regression. Both variants still fail pipeline backward-activation checks. The main MTEB pass involved reference-answer and hidden-test retrieval, so its raw score is retained and the whole pair is excluded from the eligible comparison. The candidate's MTEB pass used task-authorized public data. Costs rise 30.3% and summed execution time rises 21.6%; usage records are complete.

The final 385-word prompt (e61558f) clarifies permission to use supplied workspace files. Its separately registered path-tracing follow-up passed in 340.510 seconds with $0.114759 estimated cost and used the supplied executable. The full twenty-task score belongs to the frozen 368-word candidate; the final clarification was checked on one selected task only. This adaptive, single-sample pilot has no held-out cohort and does not establish a stable general gain. Abandoned-candidate results and historical contamination findings remain separately recorded.

Validation and documentation

• Core tests: 351 passed, 4 skipped; Harbor workflow tests: 91 passed. Python 3.13/3.14 and Harbor CI pass at the PR head.
• All 40 trajectories, 152 task-file hashes, 20 image IDs, eight source/wheel snapshots, installed wheels, and native verifier retry policies were validated. QEMU's registered dependency preflight passed; the other 19 profiles are explicitly not_configured.
• The changelog records contract verification and permitted references. English development notes match the complete Chinese source. Both root READMEs and research notes are unchanged.

Experiment report · All paired results and audits · Clarification follow-up