fix: add embedding regeneration for stale/missing vectors

[johnathanneals-dev]

Summary

Closes #60. Adds the ability to regenerate embeddings for memories that were stored with NULL embeddings (provider failure at store time) or that need re-embedding after a model/dimension change.

I flagged this gap in my review of PR #15 — the underlying issue (no remediation path for failed or stale embeddings) is still present in the current codebase.

Changes

• DB layer (src/db/attribution.py): regenerate_embedding() — reads memory, checks embedding state, UPDATEs the embedding column. Returns None for missing ID, raises ValueError if embedding exists without force=True.

• REST endpoint (src/api/main.py): POST /memories/{id}/regenerate-embedding with JSON body {"force": false}. Status codes: 400 (bad UUID), 404 (not found), 502 (embedder failure), 409 (has embedding, no force).

• MCP tool (src/main.py): memory_regenerate_embedding with memory_id (required) and force (optional boolean, default false).

• Tests (tests/test_embedding_regeneration.py): 8 unit tests covering all status paths (missing memory, existing embedding without force, null embedding update, force overwrite) plus MCP tool schema and handler tests. Follows existing mock-cursor test conventions from test_agent_attribution.py.

Design notes

• The SELECT-then-UPDATE in regenerate_embedding is not concurrency-guarded, matching existing conventions in the codebase (no locking patterns elsewhere).
• force=False default protects against accidental re-embedding of valid vectors. force=True is the explicit operator gesture for model-migration scenarios.

Test plan

• Verify tests pass in CI environment (tests require psycopg2 which is not available locally but is in the CI/Docker environment)
• Manual test: store a memory with embedder stopped, verify NULL embedding, call regenerate, verify embedding populated and memory appears in search results
• Manual test: call regenerate without force on a memory with valid embedding, verify 409 response