Imagining Recovery: Inference-Time Counterfactual Realignment for Vision-Language-Action Models | Startup Networx | Startup Networx