[FeathBow]

Why

invar use prepare and invar use run accepted only execution plans that name the MLX backend, so a use contract on a CUDA host could not be prepared or run with Invar's own commands. The run itself is not specific to MLX: for each group it runs invar infer batch with the plan's worker, configuration and adapter, and admission replays those logs through the same checks.

What

• An execution plan may name the vllm backend beside mlx. Every other part of preparation, running and admission is unchanged.
• The README's command table and the walkthrough say how a plan for a CUDA host names the vLLM batch worker and launch file.

Validation

The core suite, formatting, lint, type fixtures and the shared, Hugging Face and MLX worker suites passed. On an RTX 6000 Ada, a contract over 128 problems with two seeds each compared two GRPO checkpoints of a Qwen3.5-4B student through the vLLM batch worker: invar use prepare prepared it, invar use run recorded both sides, and invar use admit computed its decision from the recorded runs.