retrievallisted

# /retrieval Compare six retrieval strategies on the invoking project's own files (docs/code under the project root): 1. **Contextual retrieval** (`lexical`) — TF-IDF + BM25 fused with reciprocal-rank fusion. Fully offline, zero required dependencies. 2. **Contextual retrieval + LLM enrichment** (`lexical+ctx`) — the same lexical retriever over text enriched by an LLM (or heuristic) contextualizer that situates each chunk or document before indexing. Opt-in (costs LLM tokens); needs the `remote` extra + `ANTHROPIC_API_KEY` for LLM enrichment. 3. **turbovec** — dense ANN retrieval over embeddings, quantized with TurboQuant. Needs `sentence-transformers` + `turbovec`, installed by setup. 4. **pi-serini** — Lucene BM25 via Pyserini, the reference lexical retriever from the Pi-Serini paper. Needs `pyserini` (installed by setup) + a Java 21 JDK. (pi-serini names the strategy; pyserini names the library — both spellings are correct.) 5. **hybrid** — lexical + dense arms over the same corpus, fused with RRF at search time. Needs the same extras as turbovec. 6. **treesitter** — the same lexical BM25+TF-IDF+RRF ranking over AST-boundary chunks (cAST), carrying an enclosing function/class breadcrumb on each hit. Needs the `treesitter` extra (only for chunking; ranking itself is zero-dependency). The default `lexical` retriever runs fully offline with zero required dependencies. Every optional strategy degrades gracefully — a missing backend is