Act II — Dynamics of optimization

Act I built the arena’s instruments: a tail class that survives a union bound, the price of a continuum, two singular edges, a calibrated bulk, and a dimension that counts spectral mass instead of coordinates. What it cannot do is move. Every object so far was measured at a point chosen in advance. Training chooses its points as it goes: the derivative traverses a graph, the estimator is drawn mid-flight, and the landscape is evaluated exactly where the trajectory drags it. Act II turns the dynamic lens forward. The geometry does not retire; it becomes the control group.