hybrid-retrieval-usagelisted

# Hybrid Retrieval Usage Combine two or more retrieval methods with reciprocal rank fusion (RRF), and decide when a single method is enough versus when to fuse or contextualize. ## What it is `retrieval/fusion.py::reciprocal_rank_fusion(rankings, k=60)` takes a list of ranked lists (each an ordered list of document indices, best first) and returns a single fused ranking, scored by `sum(1 / (k + rank + 1))` across all lists a document appears in. `LexicalRetriever` already uses this internally to fuse its own TF-IDF and BM25 rankings; the same function fuses rankings **across** retrievers — e.g. `LexicalRetriever` + `TurbovecRetriever` — to get the benefit of both token-matching and semantic-matching signals in one ranked list. The same lexical + dense fusion also ships prepackaged as a first-class strategy: `retrieval.retrievers.HybridRetriever` (REGISTRY key `hybrid`, built via `build_retriever("hybrid")`) indexes a `LexicalRetriever` and a `TurbovecRetriever` over the same corpus and fuses their rankings with RRF at search time — this is what `retrieval index`/`query --retriever hybrid` use. It needs the same `turbovec` + `local` extras as `TurbovecRetriever` (indexing raises their guidance `RuntimeError` when absent). Use the class when you want the standard lexical+dense pairing; use the manual `reciprocal_rank_fusion` recipe below when fusing a different pair (e.g. lexical + Lucene) or more than two arms. Contextualization is a complementary, index-time lever (not a