RAG: retrieval quality evaluation set

[qiuethan]

Sub-issue of the RAG retrieval epic (#125).

Scope

• A small fixed question set with expected source documents, checked into the repo.
• A runnable harness reporting retrieval hit rate at k, so chunking and model changes can be compared rather than guessed at.
• Deliberately small and hand-built: enough to catch a regression, not a benchmark suite.

Acceptance

• The harness runs locally against a seeded corpus and reports a score.
• Changing chunk size produces a measurably different number.

Why this is in the epic

Wiring the bot (#81) to retrieval before knowing whether retrieval is any good makes bad answers look like a bot problem.