v4.5.0 — cross-encoder reranking, entity-graph retrieval, self-hosted model support

Retrieval-stack release. Everything is gated and defaults to current behavior — with no new env vars set, search results are byte-identical to 4.4.0.

Highlights

• Cross-encoder reranker (RERANK_ENABLED) — re-scores the fused candidate pool by reading query and document together. Speaks TEI and Cohere-compatible (Infinity / vLLM / Jina / Cohere / llama.cpp) /rerank shapes, probes at startup, degrades gracefully to fused order on outage. On our private bilingual production corpus it roughly doubled MRR (0.375 → 0.771) and lifted recall@5 to 1.0 — measure yours with the eval harness.
• Entity-graph retrieval (GRAPH_RETRIEVAL_ENABLED) — the entity graph the write path has always maintained finally feeds search: query entities expand one hop over co-occurrence edges and join RRF fusion as a weighted third list.
• Weighted RRF (RRF_VECTOR_WEIGHT / RRF_KEYWORD_WEIGHT) — bias fusion toward semantic or exact matches.
• Self-hosted models — OPENAI_BASE_URL for local encoders and LLMs (vLLM / Infinity / TEI / llama.cpp), EMBED_QUERY_PREFIX/EMBED_DOC_PREFIX for instruction-aware encoders, in-place re-embed tooling for encoder swaps, <think>-block stripping + LLM_CHAT_TEMPLATE_KWARGS for reasoning models.
• Opt-in multilingual full-text (BM25_TSCONFIG=zengram_multi) — accent-folding text-search config for mixed-language corpora, self-healing reindex.
• Resilience — an embedding-backend outage degrades search to surviving paths (with honest degraded metadata) instead of failing; MCP tool responses are capped with a truncation envelope (BRAIN_MCP_MAX_RESPONSE_CHARS); startup rejects placeholder API keys.

343 tests (was 287). Full details in CHANGELOG.md.