feat(agent-comparison): promote autoresearch live eval hardening to main

[notque]

Summary

• promote the live registered-skill autoresearch hardening from fix/autoresearch-recommendations
• include the blind body-eval guardrails, holdout/report correctness fixes, and routing/docs alignment
• include the measured socratic-debugging instruction-body improvement and the short benchmark fixtures

Included Work

• isolate real registered-skill live evals in worktrees
• harden blind body scoring to require trigger evidence and reject fallback contamination
• support body-only optimization with documented short runnable benchmarks
• fix holdout score attribution, final best-by-test selection, and final report consistency
• route autoresearch requests through agent-comparison instead of the older skill-eval path

Validation

• local validation passed on the promoted head before merge into fix/autoresearch-recommendations:
  • 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 proof artifacts were generated during the feature PR for socratic-debugging body optimization

Notes

• this PR promotes merged work from PR feat(agent-comparison): harden autoresearch live evals #207 into main
• unlike the stacked PR, this PR targets main, so the repository's Tests workflow should apply here