Separate Sesh answer support from vector-neighbor recall

Benchmarking takeaway: keep two lanes separate. Product quality should use family-deduplicated answer-supporting passage Hit@5, evidence-unit coverage, citation/source correctness, and corpus-scoped abstention, with source-first labels frozen before rankings. Exact-vs-ANN recall is only an engineering fidelity check against the same embedding space; Qdrant FineWeb-10B-style brute-force ground truth does not prove that a passage answers a question. With roughly 12 questions, report per-case counts and coverage gaps, not model-selection or weight-tuning claims. No private questions, passages, or source IDs are included here.