A small, closing window
A dynamically stable legged robot is, by definition, constantly operating on the edge of falling and relying on active control to prevent it. That means the feedback loop, sense the robot's state, compute a correction, apply it, echoing the same PID and feedback control ideas from earlier in this track, has to run extremely fast. A robot whose center of mass has drifted outside its support polygon has a small and rapidly closing window to place a foot or adjust a joint torque before gravity finishes the job and the robot actually falls.
Unlike a statically stable system with margin to spare, a dynamically stable one has no buffer for a late correction. A control loop that senses state, computes a fix, and reacts even a fraction of a second too slowly isn't just slightly less optimal, it can be the difference between a smooth recovery step and a fall, since the error keeps growing at gravity's pace the entire time the controller is behind.