Add eval analysis scripts and precision sample

[CodeNinjaSarthak]

Summary

• Adds the analysis and statistical scripts used to produce the benchmark results in the README, plus the human-annotated precision sample backing the 52% precision claim.

Files

File	Purpose
eval/analyze_multihop.py	Per-error-type breakdown for multi-hop failures (wrong_speaker, outdated_fact, reasoning_fail, missing_fact)
eval/analyze_sh_drop.py	Dev vs. held-out single-hop comparison — isolates selection effect from run-to-run variance
eval/bootstrap_ci.py	Bootstrap 95% CI + paired significance testing against RAG baseline
eval/sample_facts_for_precision.py	Stratified sampler over user_ids → produces the CSV below
eval/results/precision_sample.csv	Human-annotated 100-fact sample; is_relevant column filled during manual review (52% precision)

Test plan

• python eval/bootstrap_ci.py runs without error against existing results JSONs
• python eval/analyze_multihop.py runs without error
• python eval/analyze_sh_drop.py runs without error
• precision_sample.csv opens cleanly and is_relevant column is present