Fix MLX serving for agent harnesses.
- Add --max-tokens 32768 to the MLX server command; mlx_lm defaults to 512 tokens, which truncates thinking output for clients that do not send max_tokens (e.g. Hermes Agent).
- Note that MLX needs KRAFTON/Raon-1-mini-preview-mlx as the model ID in Agentic Use; mlx_lm server loads the model named in each request, so Raon-1-mini-preview returns 404.
Drop the precompiled kernel commit pin from the vLLM install
`pip install git+...` checks out a detached HEAD, so the fork's setup.py cannot resolve its upstream base and installs nightly kernels, whose op signatures no longer match (NVFP4 fails in gptq_marlin_repack). VLLM_PRECOMPILED_WHEEL_COMMIT selects the matching kernels.
