Interpose, absorb, recover — strike where it is, not where it was

Four scenarios on one engine. Watch a robot fighting league and you see the same defect: a unit commits to a strike at an opponent that has already moved, because its recognition loop is tens of milliseconds behind. Fix that, and the same competence becomes protective. Every run begins the same way: an infrared marker — the only human input — designates either a protected entity or a threat; the LiDAR layer then gives the machine a fast, predictive, defensive response. LiDAR sees 360° and predicts between frames, so the unit commits to the intercept point. Vision sees only a ~60° cone in front of itself — everything else is untracked, which is precisely why it reacts too late. And because a fall is cheap for a machine it commits at full speed and runs pre-load → detect → absorb → reacquire to get back up.