M5: embedding-experiment infra + sub-exp A null result + UniXcoder (B) setup

M3: Hybrid RRF + MS-MARCO cross-encoder reranker

Parallel embed/upsert with .npy vector cache; Qdrant reap resilience

M3 design probe: Recall@K of candidate pools (BM25, Dense, RRF, set-union)

M2 done-criterion: qualitative BM25 vs dense comparison

M2 concept checkpoint: HNSW ef_search sweep

M2 complete: dense retrieval via MiniLM + Qdrant, MRR@10=0.3891 Recall@100=0.7520