Hybrid semantic retrieval (core novelty)

[cl117]

Goal
Add semantic retrieval that keyword/VSEARCH can't do (functionally similar but non-homologous parts).

Tasks

• Offline-compute text embeddings (name/description) into Typesense.
• Offline-compute sequence embeddings (DNA model: Nucleotide Transformer / Evo; CDS→protein via ESM-2 as v2), precomputed at index-rebuild (GPU).
• Implement the fusion layer in Flask: multi-recall (keyword + text-vector + seq-vector) → RRF → log(pagerank+1) re-rank.
• Evaluate vs old ES baseline on the Phase 0 harness.

Deliverable
Method figure + main results table (the paper's central contribution).