Module efficacy 

Learning-efficacy evidence (#549, VIS-3).

The learning layers (#538-#544) adapt continuously — this module proves whether the adaptation WORKS, with real telemetry instead of claims:

• Bounce-rate trend straight from the savings ledger (daily_bounce_trend, #507): week-over-week rate of compressed reads that had to be re-read full. Learned thresholds (#538) must push this down.
• LITM placement snapshots: daily cumulative hit/miss counters from the calibration store (#539) so hit-rate movement is visible over time.
• Playbook survival (#541): share of entries that stayed net-helpful past 10 turns — the ACE quality proxy for “facts worth keeping”.
• Prevented duplicate work (#540): lifetime count of rejected scent claims.

Snapshots live in ~/.lean-ctx/efficacy_snapshots.json, bounded to a 30-day ring, captured lazily on every ctx_metrics call and on server shutdown — no timers, no daemons.

Structs

EfficacySnapshot

EfficacyStore

Functions

capture

Capture (upsert) today’s snapshot. Called from ctx_metrics and shutdown.

report

Human-readable efficacy section for ctx_metrics.

report_json

Machine-readable efficacy for the dashboard (#548).