Correct contraction, but prediction ranges remain too wide
All six generated identities pass the correctness checks, under three models and both processing cases. The known true parameters survive both training-noise assumptions, and every future truth remains inside its calculated latent-position range. No physical limit or nuisance coordinate was removed.
However, only two of 27 coordinates narrow. After 1.5 seconds of generated training observations, the 1.8-second future ranges still span roughly 770 m north, 174–596 m east and 512–522 m in height. Even a declared 1 mm training allowance does not make the complete broad-box calculation useful for model discrimination. These are overly broad calculated ranges, not measured aircraft errors or survey-IMU drift.
The test shows why a small box around a fitted answer cannot simply replace the full region. Initial orientation, lever arms and other calibration possibilities remain coupled; bounding them separately loses information. Further work needs to preserve more of that joint structure or establish stronger training-feasible constraints. Another observed prediction fit does not follow from this result.
The declared arithmetic allowance is also not a global numerical certificate. The shared receiver-error assumptions and absolute withheld adequacy remain open. The original 53 numerical comparisons are unchanged.
The sealed study used 30 native calls, 21,600 reserved bound steps and 51.60 seconds, with zero archive reads or optimizer starts. Do not restart it or reuse its unused allowance.