Correct completed checks, but the implementation is too slow
Twenty of the 24 controls pass. The four broad globe cases reach the unchanged per-bound time limit, so the complete candidate does not qualify. All six derivative controls, all six packet-boundary controls, all six stationary-motion controls and both broad flat-disc controls pass.
The implementation carries separate coordinate bounds through velocity and position calculation while keeping the existing global norm bounds. It retains the original model equations, centers, training-conditioning path and all physical/shared columns. Local training-region tests also pass. This is useful implementation progress, but it is not a qualified replacement for the complete prediction method.
A single, shortened generated-data performance probe identifies repeated exact rational construction and repeated validation of already immutable arrays as major costs. It used 32 packets and a fixed clock for profiling only; it is not a retry of a failed study identity or a scientific acceptance test.
The next candidate will reuse immutable validated views and cache repeated exact norm calculations. The rational arithmetic and its resulting bounds must remain unchanged. Independent tests compare all returned bounds bit for bit with this reference. Qualification requires a new freeze under the same limits; the completed study and its four failures are preserved.