Calibrating the gates to the aircraft we actually fly

A safety filter added in July — the velocity D-term low-pass that contains estimator-step thrust spikes — quietly cost settled waypoint precision, 0.05 m to 0.08 m, and the 1.5 m sim gate never noticed. A pinned-environment bisect found it; halving the filter time constant kept ~13x spike attenuation and returned 0.03–0.07 m. Every validation gate across the three sims is now calibrated to the measured seed-sweep envelope at roughly 3x the worst observed run — margins that mean something, recorded next to the assertions they bound.