Expand RAG retrieval benchmark

[lmdixon23]

Summary

Expands the RAG Assistant retrieval evaluation from a five-case smoke check into a deterministic 40-case benchmark over 10 documents.

Changes

• adds stable case IDs, tags, multi-source labels, and dataset validation;
• covers direct lookup, paraphrase, terminology variation, cross-chunk wording, and hard-negative categories;
• reports recall@1, recall@3, MRR, per-case results, and per-tag slices;
• adds a versioned deterministic HashEmbedder baseline and regression rule;
• updates the CLI, smoke and re-ranker evaluation runners, README, and repository results documentation;
• adds benchmark and metric regression tests without changing retrieval algorithms.

Deterministic baseline

• cases: 40
• documents: 10
• chunks: 31
• embedder: HashEmbedder
• recall@1: 0.500
• recall@3: 0.775
• MRR: 0.629

These values are a project regression baseline, not a leaderboard or general retrieval-quality claim.

Validation

• RAG Assistant: 34 passed, no skips
• LLM evaluation bridge: 2 passed, no skips
• Python compilation: 7 changed Python files
• README and RESULTS relative links verified
• whitespace and exact 21-file change boundary verified

Closes #18.