A shared framework for video-action programs and local-context dynamics, evaluated through robot control, frozen-component transfer, and action-conditioned prediction.