[4a] Org-scoped pgvector similarity search

[andrmaz]

What to build

Implement the retrieval query path: embed the incoming user query, run a pgvector similarity search against indexed Chunks, and filter results strictly by the authenticated user's Organization. Return a ranked top-k chunk list ready for Context Bundle assembly.

Acceptance criteria

• Query embedding is generated and similarity search runs against the Chunk table.
• Results are filtered by Organization FK — chunks from other orgs are never returned.
• Top-k ranked results are returned in a structured format (chunk id, content, score, source ref).
• Integration test verifies org isolation: two orgs with identical content return only their own chunks.

Blocked by

• [3b] Async ingestion worker — chunk, embed, index #10 [3b] Async ingestion worker — chunk, embed, index