d94eb8f70b
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.