feat: add mlx-server runtime to the metal-agent

[Defilan]

What

Add mlx-server as a metal-agent inference runtime, alongside llama-server,
omlx, ollama, and vllm-swift.

Why

mlx-server is a native
OpenAI-compatible MLX inference server for Apple Silicon. This lets the
metal-agent supervise it as a first-class InferenceService runtime — the
same lifecycle the agent already gives the other runtimes.

No linked issue.

How

• MLXServerExecutor (pkg/agent/executor_mlx_server.go) implements
  ProcessExecutor: resolves the model directory, spawns mlx-server, polls
  /health, and stops it with SIGTERM plus a grace period. Unlike the
  vllm-swift executor's per-spawn ephemeral port it binds a fixed port
  (default 8080) so clients keep a stable base URL across respawns.
• runtimeMLXServer wired into the executor-selection switch and
  validateRuntimeFormat (MLX directories and HF safetensors accepted, gguf
  rejected — same as vllm-swift).
• --mlx-server-bin, --mlx-server-port, and --mlx-server-startup-timeout
  flags on the metal-agent, with binary auto-detection.
• mlx-server-specific flags such as --tool-call-format and --reasoning
  ride through InferenceService.spec.extraArgs — no new CRD fields.

Checklist

• Tests added/updated — executor_mlx_server_test.go covers arg construction, the fixed-port executor, health polling, and process teardown
• make test passes locally
• make lint passes locally
• Commit messages follow conventional commits
• All commits are signed off (git commit -s) per DCO
• Documentation updated (if user-facing change) — the macOS Metal guide's runtime list is a small follow-up

[codecov]

Codecov Report

❌ Patch coverage is 28.02548% with 113 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
pkg/agent/executor_mlx_server.go	40.74%	63 Missing and 1 partial ⚠️
cmd/metal-agent/main.go	0.00%	33 Missing ⚠️
pkg/agent/agent.go	0.00%	16 Missing ⚠️

📢 Thoughts on this report? Let us know!