Implement Phase 1 + tool calling

[Defilan]

What

Builds mlx-server out from the Phase 0 scaffold into a working OpenAI-compatible inference server on mlx-swift-lm.

• mlx-swift-lm v3.32.1-alpha wired in; package split into a testable MLXServerKit library + a thin MLXServer executable
• Model loading from a local MLX directory or a HuggingFace id
• GET /v1/models
• POST /v1/chat/completions — streaming (SSE) and non-streaming
• OpenAI tool calling via mlx-swift-lm's chat-template parsers (tools / tool_choice in, tool_calls out)
• Inferencing protocol so the HTTP layer is testable without a model
• 20 tests: OpenAI type encode/decode, mapping, SSE framing, HTTP routes (stub engine)

Why

Phase 0 was a scaffold serving only /health. This is "where we left off" — the README's Phase 1, plus tool calling pulled forward from Phase 3 so the server is actually usable by agentic clients (opencode).

How it was verified

End-to-end against two local MLX models — Qwen3-4B-4bit and Qwen3.6-35B-A3B-8bit (Qwen3.5-MoE): model load, non-streaming completion with token usage, streaming SSE, and tool calls.

Notes

• Running requires an xcodebuild build, not swift build: SwiftPM cannot compile mlx-swift's Metal shaders, so a swift build binary fails at runtime with Failed to load the default metallib. CI now builds with xcodebuild to verify the real artifact; swift test runs the (model-free) suite.
• Reasoning models currently emit <think> content into the response content — routing it to reasoning_content is a planned follow-up.

Roadmap

Phase 0 and Phase 1 (+ tool calling) are now done. Next: reasoning/content split, then LLMKube runtime: mlx-server integration (Phase 4).