RayRAG v0.4.83 — search legs + unified SCA review (W-4 continued)

What's in this release

• Search legs (RAGFlow rag/advanced_rag/harness/tools/search.py → src/harness/tools/search.rs): retrieval settings resolution (top_n / top_k / rerank-candidate floor), the 400-character effective_query (weighted retrieval query first), the per-request search cache key, and the three retrieval legs behind an injectable HarnessRetriever — hybrid (0.3 vector weight, degrades to keyword-only when no embedding model), vector (weight 1.0, pure-cosine 0.2 floor) and BM25 (weight 0, threshold 0) — each followed by keyword narrowing. Also is_table_chunk (HTML / >=3 pipe rows), compact grep-term extraction (10-term cap), regex-query-to-terms conversion, grep_search (BM25 candidates + whole-table passthrough + prose narrowing with a one-line lead window, falling back to the raw candidates), list_chunks (80-chunk cap), load_specific_chunks and rank_chunks_by_terms, plus the tool-name and compiled-tool-mode tables.
• Unified Sufficient Context Agent (orchestrator/sufficient_context.py → src/harness/orchestrator/sufficient_context.rs): the sca_select prompt verbatim, per-claim context rendering (draft + up to three bounded evidence anchors per claim — tables keep their full text, prose keeps a 300-char window), the overall draft assembly, coerce_dict format-drift tolerance, full parsing of the verdict (sub_queries with missing_fact/search_hint, per-claim groundedness and missing information), the insufficient-with-empty-claims failsafe (_global from top-level missing pieces or a coarse fallback derived from the drafts), and the to_boost / to_grounded adapters.
• Ledger: both rows closed as full; worklist W-4 queue 10 → 8; ragflow-coverage --check green (5977 files, unmapped=0).
• Tests: 10 new unit tests (search 5 + SCA 5); cargo fmt --check, cargo check --locked --features postgres-backend,zvec-backend and the interface gate clean.

Build: cargo build --release --locked (default features: postgres-backend, zvec-backend).