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

Change

Added share-based replay metadata checks to research.continuous_loop.assess_confidence.

The confidence gate already required minimum absolute counts for:

• pair_eventful_metadata_rows
• large_move_pair_eventful_metadata_rows

That still allowed a replay dominated by legacy rows to clear the metadata gate once it accumulated a small absolute number of eventful-era pairs. The gate now also requires:

• pair_eventful_metadata_share >= 0.05
• large_move_pair_eventful_metadata_share >= 0.5 when large-move rows exist

These shares are now persisted in the confidence block alongside the existing absolute counts.

Why

Current source of truth shows recent journal rows have eventful metadata, but replay pair coverage remains low because older non-eventful rows still dominate the paired history. This is a replay evaluation quality problem, not a live-trading threshold problem.

Validation

Ran:

pytest tests/test_continuous_loop.py

Focused assertions added:

• low pair-level metadata share fails confidence even with otherwise healthy replay stats
• low large-move pair metadata share fails confidence even when absolute large-move metadata count passes