feat(gitree): add vector search for concepts

[Evanfeenstra]

Summary

Adds semantic (vector) search for Concepts. Previously only Clues had embeddings; concepts were only retrievable by ID or plain text matching. This mirrors the existing clue search infrastructure.

What & why

• What to embed: name + description (concise, high signal) rather than documentation, which can be arbitrarily long and would be silently truncated at the model's 512-token limit / diluted by mean-pooling.
• Concept nodes are already labeled :Data_Bank:Concept, so no new schema/index work is needed to store embeddings.

Changes

• types.ts — add embedding?: number[] to Concept.
• store/utils.ts — conceptEmbeddingText() / computeConceptEmbedding() helpers (embed name + "\n\n" + description via vectorizeQuery).
• saveConcept (GraphStorage + FileSystemStore) — computes/stores the embedding on every write, so both create and description-update paths stay current. Falls back to any existing embedding on failure.
• searchConcepts() — added to the Storage abstract and both impls:
  • GraphStorage: brute-force gds.similarity.cosine over :Concept nodes (concept counts are small, so a full scan is fast and gives exact scores — same pattern as searchClues).
  • FileSystemStore: in-memory cosine.
• POST /gitree/search-concepts route + search-concepts CLI command. Results are flat ConceptSummary objects with a merged score.
• Startup migration: backfillConceptEmbeddings() runs at the end of initialize() (alongside the existing backfillPRFileEdges), embedding any pre-existing concepts missing an embedding. Becomes a cheap no-op scan once all concepts are embedded — no manual script required.

API

curl -X POST http://localhost:3355/gitree/search-concepts \
  -H "Content-Type: application/json" \
  -d '{"query": "authentication"}'

Returns:

{
  "query": "authentication",
  "repo": "all",
  "count": 1,
  "results": [
    { "id": "owner/repo/auth-system", "name": "Authentication System", "description": "...", "prCount": 2, "commitCount": 0, "lastUpdated": "...", "hasDocumentation": true, "score": 0.82 }
  ]
}

Testing

• tsc --noEmit passes.

Notes

• Backfill runs sequentially and loads the BGE model in-process; fine for hundreds/low-thousands of concepts. Can switch to vectorizeBatch if a much larger legacy corpus is expected.