feat: add Ollama as local embedding provider

[TerminalGravity]

Adds OllamaEmbeddingProvider using Ollama /api/embed endpoint. Fully local, no API key needed. Defaults to nomic-embed-text, 768 dims. Closes #6

[TerminalGravity]

Clean implementation — follows the same pattern as OpenAI, batch support via /api/embed is the right call. Defaults (nomic-embed-text, 768 dims) are sensible. CI green ✅

One thought for later: a health check on init (ping Ollama, verify model is pulled) would give users a clear error instead of a cryptic fetch failure. Not blocking though.

[TerminalGravity]

Clean implementation — the factory pattern extension is exactly right, and good call using /api/embed with batch support from the start.

A couple thoughts for follow-up:

• Consider adding a connection health ping so users get a clear error if Ollama isn't running (instead of failing on first embed)
• The node engine bump to >=20 in package-lock seems unrelated — intentional?
• Worth adding a README note about supported Ollama models

Closes #6. Merging. 🎉

[TerminalGravity]

Clean implementation. Batch endpoint usage avoids N+1 round trips — good call.

A few suggestions for follow-up:

1. Timeout on fetch calls — if Ollama is down or slow to load a model, the MCP server will hang. Even a 30s timeout + clear error would help.
2. Node engine bump — the package-lock bumps to >=20, worth noting since it's technically breaking for anyone on 18.
3. Health check — a quick /api/tags ping would make it easy to surface 'Ollama not running' vs 'model not found' errors.

None block merge. Nice close on #6.

[TerminalGravity]

Nice and clean. The Ollama provider looks solid — using the batch /api/embed endpoint for both single and batch is smart.

One small thing: if Ollama isn't running, the fetch will throw a connection refused error that'll be hard to debug. Worth wrapping with a friendlier message like "Cannot connect to Ollama at ${this.baseUrl} — is it running?" in a try/catch around the fetch calls.

Also, the node engine bump from >=18 to >=20 in package-lock.json — was that intentional? If so, should match in package.json too. If not, might want to revert that line.

[TerminalGravity]

Superseded by a newer PR.