Do What? Teaching Vision-Language-Action Models to Reject the Impossible