Ralph Loop 2026-03-16: Pair-Level Eventful Replay Metadata

Change

Improved replay evaluation quality in research/prepare_replay.py by carrying eventful metadata from both observations in each replay pair instead of only the entry-side snapshot.

New replay fields:

• _next_eventful_snapshot
• _next_eventful_score
• _pair_eventful
• _pair_eventful_score

Replay summaries now report:

• eventful_rows as pair-level eventful coverage
• entry_eventful_rows
• next_eventful_rows

This fixes undercounting when an eventful observation appears as the later side of a replay pair.

Why This Was The Smallest Useful Change

Current loop state showed replay prep producing many rows but almost no eventful coverage:

• raw journal rows: 1742
• replay rows: 1280
• raw eventful observations: 60
• replay eventful rows before fix: 1

The journal already contained eventful observations, but replay metadata only preserved eventfulness from the current row in each pair. That made replay input quality look weaker than it really was.

Validation

Focused test:

PYTHONPATH=. pytest -q tests/test_prepare_replay.py

Result:

• 3 passed in 0.02s

Smoke command:

python3 research/prepare_replay.py --db data/unicorn.db --out /tmp/replay_rows_check.jsonl

Observed profile after fix:

• total_rows=1280
• eventful_rows=2
• entry_eventful_rows=1
• next_eventful_rows=1
• actionable_signal_rows=0
• review_signal_rows=0

Remaining Bottleneck

This fix improves replay metadata honesty, but the larger issue remains upstream:

• all 1742 current journal rows are still signal_action=watch
• all current replay rows are still model_source=heuristic

So the next useful change is likely in observation or sleeve generation, not replay pairing.