Ralph Loop 2026-03-16: Candidate Review Event Stream

Change

Added a candidate_review_done event to the continuous research loop.

Files: - research/continuous_loop.py - research/README.md - tests/test_continuous_loop.py

Why

The loop already wrote the live candidate-review packet into research/latest_state.json, but the step-level event stream stopped at replay evaluation. That meant Ralph could not see whether live review quality improved from one cycle to the next without opening the latest-state handoff.

Result

Each successful cycle now emits a compact candidate-review summary with:

1. deployable, review, and research candidate counts
2. deployment-gate readiness
3. latest scan timestamp
4. top invalidation reason
5. top deployable/review/research tickers

This keeps candidate-review quality observable in the same place as backfill, replay, and search diagnostics.

Validation

• pytest -q tests/test_continuous_loop.py
• python3 -m py_compile research/continuous_loop.py

Bottleneck After Change

This improves operator visibility, not candidate quality itself. The larger remaining issue is still weak eventful coverage and stale loop snapshots versus current replay/live scans.