feat: memory embedding router (local + API) (by Lumen)

[conoremclaughlin]

Summary

Adds a first implementation pass for semantic memory embeddings with a provider router that supports both local and API backends.

What’s included

• New embedding router service with vetted model catalog:
  • packages/api/src/services/embeddings/router.ts
  • packages/api/src/services/embeddings/vetted-models.ts
• Environment/config support for embedding provider selection and recall tuning
• MemoryRepository.remember() now attempts to embed each new memory and persist the vector metadata
• MemoryRepository.recall() now attempts semantic recall via pgvector RPC and falls back to text search
• New migration:
  • supabase/migrations/20260308092624_memory_embedding_recall.sql
  • adds idx_memories_embedding (HNSW)
  • adds match_memories(...) RPC for filtered cosine similarity search

Design notes

• Router supports both local (Ollama) and API (OpenAI) with fallback behavior.
• Current schema is still memories.embedding vector(1024), so runtime dimensions are enforced to 1024 for now.
• This keeps implementation compatible today while leaving a path to model-aware indexing later.

Verification

• yarn workspace @personal-context/api test src/data/repositories/memory-repository.test.ts
• yarn workspace @personal-context/api test src/mcp/tools/memory-handlers.test.ts

Follow-ups

• Add async janitor/backfill re-embedding job
• Add model-aware dimension/index strategy when we commit to Option B
• Expose vetted model list + active embedding profile via MCP/admin endpoint