Introduce SmolVLA-style VLM prefix encoder backbone and IMF-AttnRes conditioned agent. Add episode-level train/val split, action MSE validation in training, and headless eval support.