Action Without Interaction: Probing the Physical Foundations of Video LMMs via Contact-Release Detection