Day 8: Streamlit UI + dotenv config

Day 7: refusal-hardened prompt (7/7), strict citations; faithful 0.89, correct 0.91

Weighted RRF (2:1): no recall recovery; hybrid still 0.86 vs dense 0.91

Day 5 ablations: hybrid recall@20=0.97, bge-reranker-base hurts MRR (0.71->0.63)

Baseline v0: recall@5 0.91, MRR 0.71, correct 0.89

Day 3: embedding index + baseline ask pipeline