Fix PV metric to match paper's orchestration-focused evaluation

[harsh-kr11]

Summary

• Fixed Parameter Validity (PV) metric to align with the paper's intent (Section IV.C, IV.D.3). The paper explicitly treats SQL-level parameter differences as "minor errors" — PV should measure orchestration decisions, not content generation.
• Categorized parameters into three types:
  • Orchestration (format, channel, source_step, mode, etc.) — exact match
  • Content (query, body, subject, title) — structural/semantic match (right tables, right domain terms)
  • Identifier (recipient, target_name) — lenient key-term match
• Updated README with reproduced live benchmark results (gemini-2.5-pro, pgvector) and a detailed "How the Agent Learns" section explaining the feedback loop architecture.

Reproduced Results (gemini-2.5-pro, pgvector)

Metric	Paper	Live Run
TSA	83.3%	86.7%
PV	84.0%	82.2%
PCR	63.3%	80.0%
ESA	83.3%	86.7%
McNemar p	0.004	0.039

All results within the paper's 95% bootstrap confidence intervals.

Files Changed

• src/behavioral_memory/evaluation/metrics.py — Core fix: parameter categorization and semantic matching
• README.md — Reproduced results table, learning architecture diagram
• examples/run_live_benchmark.py — pgvector flag for paper reproduction
• docs/GETTING_STARTED.md — Updated pgvector setup instructions
• Makefile — Added benchmark-pg target
• CONTRIBUTING.md — Updated test count

Test plan

• All 104 existing tests pass (pytest tests/ -v)
• ruff check and mypy pass
• python examples/validate_pipeline.py passes all 30 checks
• Gatekeeper ablation script uses the corrected metrics via compute_metrics

Made with Cursor