Evaluation dashboard

This tests the one place in the pipeline where the LLM's output feeds into anything downstream: extractProfile(), which turns free text into a structured profile. It does not test eligibility correctness — the rule engine is the ground truth for that by design, so there's no independent oracle to check it against. This is a 25-case hand-written gold set, run live against the real Groq API — not cached or pre-computed.