feat(agent-comparison): harden autoresearch live evals

[notque]

Summary

• harden live registered-skill autoresearch evaluation and blind body scoring
• add guardrails/tests for trigger evidence, contamination rejection, holdout selection, and final report consistency
• apply a measured socratic-debugging body improvement and document the short runnable workflows

Validation

• pytest -q scripts/tests/test_agent_comparison_optimize_loop.py scripts/tests/test_skill_eval_claude_code.py scripts/tests/test_passk_eval.py scripts/tests/test_eval_compare_optimization.py
• python3 -m scripts.skill_eval.run_eval --eval-set skills/agent-comparison/references/read-only-ops-short-tasks.json --skill-path skills/read-only-ops
• python3 skills/agent-comparison/scripts/optimize_loop.py --target skills/socratic-debugging/SKILL.md --goal 'Improve the first response so it asks exactly one question, avoids direct diagnosis, avoids code examples, and does not add tool-permission preamble.' --benchmark-tasks skills/agent-comparison/references/socratic-debugging-body-short-tasks.json --optimization-scope body-only --max-iterations 1 --beam-width 1 --candidates-per-parent 1 --revert-streak-limit 1 --output-dir /tmp/cck-autoresearch-pr-proof2 --report /tmp/cck-autoresearch-pr-proof2/report.html --verbose

Notes

• base branch is fix/autoresearch-recommendations because the autoresearch infrastructure is not on main.
• this includes the accepted socratic-debugging instruction-body update proven by the short blind body benchmark.

[notque]

Current status on 98ecf1a587affda2e09611359dd6dd30132b48cd:

• Local validation passed:
  • pytest -q scripts/tests/test_agent_comparison_optimize_loop.py scripts/tests/test_skill_eval_claude_code.py scripts/tests/test_passk_eval.py scripts/tests/test_eval_compare_optimization.py
• Short live body benchmark still works on this branch and produced an additional accepted socratic-debugging body tightening that is now included in the head commit.

Workflow note:

• GitHub reports no checks for this PR because .github/workflows/test.yml is configured only for push/pull_request on main.
• This PR targets fix/autoresearch-recommendations, so the Tests workflow does not trigger for it in its current configuration.

PR is merge-clean, but there is no branch-applicable GitHub Actions signal available for this stacked base branch.