Description:
Create an OpenAI API-compatible HTTP server that wraps mellea, allowing any OpenAI-compatible client to use mellea backends.

Key deliverables:

• using m serve serves an OpenAI API compatible endpoints
• FastAPI/Starlette server implementing OpenAI chat completions endpoint
• /v1/chat/completions endpoint with streaming support
• Model listing endpoint
• Configuration for backend selection

Acceptance criteria:

• OpenAI Python SDK works against the server
• Streaming works correctly
• Tool calling works
• Can be deployed as standalone service

Notes:

• This enables "any framework that supports OpenAI" to use mellea
• Consider using existing libraries like litellm proxy as reference

Task breakdown

P0 — Breaks Compatibility
These are things that cause OpenAI SDK clients to fail or return incorrect results.

• Add finish_reason to response choices. Default to "stop"; set to "length" if max_tokens was hit; set to "tool_calls" if tool calls are present.
• Add usage to response.
• Map standard parameters to ModelOption sentinels.
• Fix sync/async mismatch. Wrap module.serve() in asyncio.to_thread() or detect and await async serve functions.
• Add OpenAI error response format. Return {"error": {"message":..., "type":..., "code":...}} on errors instead of FastAPI's default HTML/JSON.

P1 — Core Features Expected by Clients

• see sub-issues
• /v1/models endpoint. Return the model(s) configured in the serve script.
• Health check endpoint. GET /health returning {"status": "ok"}.