Phase 2: embedded store + context injection

[Copilot]

Implements Module 2 / Phase 2 from the super-ollama plan: a local SQLite-backed data store with semantic vector search, wired into every inference path for automatic context injection.

internal/store (new)

• SQLite via modernc.org/sqlite (pure Go, no CGO) — auto-migrates a 4-table schema (events, emails, todos, contacts) on Open()
• Embedder interface — store.Embedder decouples embedding generation from the store; wired post-open via SetEmbedder()
• Vector search — embeddings stored as JSON BLOBs; Search() / SearchSummaries() run in-process cosine similarity ranking (no external vector DB)
• SaveInteraction() — persists every prompt/response pair as a source='interaction' event for future retrieval
• Full CRUD: InsertEvent, InsertEmail, UpsertTodo, UpsertContact, UpdateEventEmbedding, EventsWithoutEmbedding

internal/engine — context-aware inference

// Generates a float32 embedding via the in-process EmbedHandler
emb, err := eng.Embed(ctx, "nomic-embed-text", text)

// Retrieves top-10 similar events, prepends as context, saves interaction
resp, err := eng.GenerateWithContext(ctx, model, prompt, prompt, nil, store)

ContextSearcher interface keeps engine independent of store (no import cycle).

internal/config — new fields

embedding_model (default: nomic-embed-text) and db_path (default: ~/.super-ollama/data.db) with ResolvedDBPath() expanding ~.

CLI changes

• learn command (was stub) — iterates EventsWithoutEmbedding, calls Embed(), persists results; respects SIGINT
• ask / chat — open store on startup, inject semantic context per turn; degrade gracefully if store is unavailable
• config show — now prints embedding_model and db_path

Tests

12 unit tests in internal/store using in-memory SQLite (:memory:), covering schema migration, all CRUD paths, cosine similarity correctness, SearchSummaries, and SaveInteraction.

[dkritarth]

Verified locally: go test ./... -count=1 passes on macOS (PR branch). Internal/store + engine + server packages OK.