MCP: vector chunk matches with unified match interface

[endolith]

Summary

Adds within-message vector chunk scoring and harmonizes match excerpts across MCP search tools.

Changes

• ChunkScoringBackend.ScoreMessageChunks — scores every embedded chunk of a message against the query vector (sqlitevec + pgvector)
• search_messages (mode=vector|hybrid) — each hit includes matches[] (up to 5 chunks, best similarity first, optional min_score)
• search_in_message (mode=vector) — paginated semantic chunk matches for one message (same matches shape as keyword mode, with score on each)
• search_message_bodies — context_snippets replaced by matches[] with char_offset, snippet, and line (keyword matches, no score)

Usage

// Hybrid search with chunk excerpts
search_messages { "query": "quarterly budget", "mode": "hybrid", "min_score": 0.3 }

// Within-message semantic chunks
search_in_message { "id": 123, "query": "project deadline", "mode": "vector", "min_score": 0.2 }

Each vector match includes char_offset (byte offset into body_text when the chunk is in the body), snippet, line, and score (0–1 similarity). Use get_message with center_at=<char_offset> to read more context.

Snippets reflect preprocessed embed text (subject prefix + cleaned body), matching what was embedded.

Stack

Stacks on PR #9 (split/pr4-search-message-bodies).

  

[endolith]

Close in favor of #15