[ancongui]

Queries constrained to source_ids could lose admitted chunks because BM25 and dense retrieval chose unrestricted candidates before filtering. Apply nonempty source membership together with authenticated tenant/workspace before each ranking window and fusion; keep hydrated checks as defense in depth. Omitted, null and empty-list selectors retain their existing semantics, and unmatched nonempty selectors return no global fallback.

Prepare 26.10.6 server and Python/Java SDK metadata and release notes. Wire schemas, SDK methods, runtime dependency pins, database migrations, source intake and EDA schemas remain unchanged.

Validation:

• Built server wheel: 1,230 unit tests passed; 5 Docker-dependent tests skipped.
• Built Python SDK: 61 tests passed. Java SDK: 62 tests passed and Maven verify passed offline.
• Ruff check/format passed; Pyright: 225 files, zero errors or warnings.
• Server and Python SDK wheels/source archives and Java SDK JAR built locally.
• Native PostgreSQL 17/pgvector 0.8.2 checks used real local 768-dimensional embeddings: independent BM25 and dense-only filtering, lower-ranked admitted content behind distractors, actual HNSW execution, tenant/workspace intersections, missing-source emptiness and authenticated HTTP retrieval passed.

Native runtime checks used the published 26.10.5 dependency set with an explicitly inventoried private overlay whose two runtime files match this change. This PR does not claim published-artifact or full framework-corpus acceptance. No remote LLM or RLM answer calls were used. Release artifacts are not published yet.