Ollama wired in as a local LLM provider with routing and model discovery

stakwork/mikeoss adds Ollama as a first-class provider alongside Claude and Gemini. The change covers the backend adapter, request routing, and a new endpoint that enumerates installed models from the Ollama server.

The adapter fits into the existing provider pattern. A new "ollama" provider type gets registered, and model ids use an ollama/ prefix so the routing layer can dispatch them without touching the existing Claude or Gemini paths. The Ollama adapter talks to Ollama's OpenAI-compatible /v1 endpoint using the standard openai npm package rather than any Ollama-specific client.

Model discovery goes through a new GET /user/ollama-models route that proxies Ollama's /api/tags endpoint. This lets the frontend enumerate whatever models the Ollama server has installed - useful because local Ollama installs vary, so hardcoding model names isn't viable.

The PR was opened and merged in seconds, consistent with Hive-driven automated authoring landing straight to main. Two commits cover the work.

One thing to know before adopting: Ollama models get tools stripped at the routing layer. The backend passes an empty tools array and emits a tools_unavailable SSE event when Ollama is selected. Doc editing and citations don't work with local models. The matching frontend PR (c629bb00) handles the UI side of that signal.