feat: Semantic retrieval for raglogs ask (pgvector)

[leo-aa88]

Summary

Upgrade raglogs ask from keyword-only retrieval to semantic search using stored embeddings (pgvector), as described in the README.

Motivation

• README states: “Semantic retrieval via pgvector is planned for a future release.” (raglogs ask section)
• Ingest already supports --with-embeddings; retrieval should use that data when enabled.

Scope (proposal)

• Query path: embed question (or sparse hybrid later), retrieve nearest log lines/clusters, ground answers in counts and timestamps.
• Graceful fallback: keyword mode when embeddings disabled or missing.
• Config: env flags for top-k, similarity threshold.

Acceptance criteria

• With embeddings ingested, ask returns relevant answers on paraphrased questions that keyword mode misses.
• Tests with fixture vectors or mocked embedder.

Related

• May interact with “semantic cluster merging” roadmap item; keep concerns separated unless a shared embedding module makes sense.