feat: autonomous scoring weight adaptation

[salishforge]

Summary

Self-tune importance scoring weights based on retrieval effectiveness metrics over time, rather than using fixed weights.

Design

• Track retrieval feedback outcomes per agent (positive/negative/neutral)
• After N feedback events (e.g., 100), compute which scoring dimensions (recency, frequency, centrality, reflection, stability) correlate most with positive outcomes
• Gradually adjust weights toward empirically better values
• Store per-agent weight profiles in agent metadata
• Bounds: no weight below 0.05 or above 0.50

Inspiration

Inspired by MH-FLOCKE (Apache 2.0) autonomous closed-loop parameter tuning and evolved plasticity rules.

Closes #57

[salishforge]

Implementation Decisions

1. Correlation method: Simple count correlation (Option A)

Options considered:

• A) Simple count: For each scoring dimension, count how many positive-feedback memories have above-median values. Increase weight for high-correlation dimensions. ← Selected
• B) Linear regression: Fit a model predicting outcome from the 5 dimension scores. Better accuracy but requires math library or manual implementation.
• C) Bayesian update: Start from priors, shift incrementally per feedback. Most principled but hardest to debug and explain.

Rationale: Option A requires no external dependencies, is easy to debug/explain, and provides a clear baseline. Options B and C should be evaluated via A/B testing once we have data showing Option A's limitations.

2. When adaptation runs: Sleep cycle Phase 0 (before scoring)

Runs once per sleep cycle, not on every feedback call. Batch processing is cheaper and avoids oscillation from individual feedback events. Only activates after 100+ feedback events for the agent.

3. Per-dimension score access: Store at feedback time

When feedback() is called, we snapshot the current importance score and the 5 dimension contributions into retrieval_log metadata. This gives us ground truth for correlation analysis without expensive recomputation. The dimension breakdown is computed from the same SQL used in Phase 1 scoring.

4. Learning rate: 0.01 per cycle

Conservative — takes ~50 sleep cycles to shift a weight from 0.20 to 0.25. Prevents oscillation while allowing meaningful adaptation over time.

5. Activation threshold: 100 feedback events minimum

Below 100 events, statistical signal is unreliable. Agent uses global defaults until threshold is reached. Once active, per-agent weights are stored in agents.scoring_weights JSONB and loaded in Phase 1.

Future A/B testing opportunities

• Option A vs Option B (linear regression) accuracy comparison
• Learning rate tuning (0.005 vs 0.01 vs 0.05)
• Activation threshold (50 vs 100 vs 200)
• Per-domain weight profiles (code agents vs research agents)
• Weight decay toward global defaults when agent is inactive