Ralph Loop 2026-03-16: Replay Metadata Confidence Gate

Change

Added an explicit replay-quality confidence gate tied to eventful pair coverage.

1. research/prepare_replay.py now reports large_move_pair_eventful_metadata_rows directly instead of requiring operators to infer it from other counters.
2. research/continuous_loop.py now feeds replay-input coverage into assess_confidence() and blocks confidence when: - pair_eventful_metadata_rows is below the configured minimum - large_move_pair_eventful_metadata_rows is below the configured minimum when large moves exist
3. Status and Discord summaries now print the explicit large-move pair metadata count so replay miss quality matches the confidence gate.

Why

The current replay file is still dominated by legacy rows without eventful pair metadata. That means a positive replay score could look better than it really is from a research-quality perspective.

Current local replay snapshot after the change:

• pair_eventful_metadata_rows=2
• large_move_pair_eventful_metadata_rows=0
• confidence blockers now include pair_eventful_metadata_rows<10 and large_move_pair_eventful_metadata_rows<3

Validation

Commands run:

• PYTHONPATH=. pytest tests/test_prepare_replay.py tests/test_continuous_loop.py tests/test_research_search.py -q
• python3 main.py research-status
• python3 main.py research-discord-status
• direct confidence check against research/replay_rows.jsonl

Results:

• Focused tests passed: 19 passed
• Current replay remains non-confident for the right reasons, including the new metadata coverage blockers
• Status outputs show large_move_metadata(entry/next/pair)=.../.../... using the explicit pair counter