Ralph Loop: Candidate Review Sleeve Tie-Break

Change

• Updated research/candidate_review.py so “best sleeve” selection prefers fewer hard semantic blockers before total blocker count.
• Hard blockers now include sleeve disablement plus mismatches on signal_action, model_source, market_family, category, neutral direction, and missing live price.
• Candidate packets now expose sleeve_hard_blocker_count and sleeve_soft_blocker_count for easier debugging.

Why

• Candidate review used raw blocker count with alphabetical tie-breaks.
• That could label a row with the wrong sleeve family when one sleeve had a category/model mismatch and another sleeve only failed on tunable thresholds.
• Mislabeling hurts candidate review quality because Ralph can end up chasing the wrong sleeve family.

Validation

• pytest -q tests/test_candidate_review.py tests/test_research_status.py
• Result: 6 passed in 0.08s
• python3 main.py research-status
• Confirmed status still renders normally after the classifier change.
• Smoke script over the latest 6 live runs
• Result: {'rows': 36, 'changed': 0}
• Interpretation: the fix closes a real tie-break bug without changing the current live packet mix.

Outcome

• Bounded PRD 003 quality fix: candidate review sleeve attribution is now more semantically faithful, even when no live deployable trades exist.