Add 100-PR eval evidence: flow-guided wins 92%

[benitezfede]

Summary

• Adds a 100-PR blind evaluation comparing baseline (diff-only) vs flow-guided code review across 57 open-source repositories
• Flow-guided approach wins 92% of comparisons with +1.3 avg score improvement and +3.1 on flow awareness
• Adds detailed methodology docs explaining the 3-agent framework and all 5 scoring criteria
• Updates main README with a "Why Flow-Guided Review?" section linking to eval evidence

What's included

• evals/*.json — 101 individual eval results (one per PR), each containing baseline review, flow-guided review, and blind judge scores
• evals/RESULTS.md — aggregated results: win rates, per-criterion breakdown, by-language analysis, full PR table with GitHub links
• evals/README.md — methodology: 3-agent framework, detailed criterion definitions (Completeness, Flow Awareness, Risk Identification, Actionability, Efficiency), file format, how to run
• evals/run-eval.ts — eval runner script
• README.md — new "Why Flow-Guided Review?" section with key results and links to eval docs

Key results

Criterion	Baseline	Flow-Guided	Delta
Completeness	6.7	7.7	+1.0
Flow Awareness	3.9	7.0	+3.1
Risk Identification	6.3	7.6	+1.3
Actionability	6.2	7.1	+0.9
Efficiency	7.1	7.0	-0.1
Overall	6.0	7.3	+1.3

Test plan

• All 101 eval JSON files parse correctly
• RESULTS.md aggregations match individual file data
• README links resolve to correct files
• [na] No code changes to test — documentation and data only

🤖 Generated with Claude Code