Add CrossSource to 5c · RAG / retrieval evaluation

[zoeb-nomi]

Adds CrossSource to 5c · RAG / retrieval evaluation — an open evaluation harness measuring citation accuracy in RAG over public court opinions.

Why it clears the bar (real data + code, not a generic "you need evals" take):

• Validated LLM-as-judge: 15/15 blind human–judge agreement on a stratified sample.
• The validation pass itself caught a real harness bug the judge couldn't see — consecutive citations producing punctuation-only claim spans scored as citation failures. Fixing it moved citation precision 0.981 → 0.994 under strict citation discipline.
• Claim-level citation precision & recall, plus a failure taxonomy that separates prompting failures from retrieval failures.

Repo (canonical, URL verified): https://github.com/zoeb-nomi/crosssource

One addition, placed in 5c immediately after Tonic Validate; entry annotated per CONTRIBUTING.md.