Check denser training before another long bound run

The next diagnostic should use the existing generated eight-minute inputs for all three models and both processing assumptions. Preserve the broad physical box, ±1-second clock and prediction at 438 seconds.

Compare the original five training epochs with 292 earlier epochs at seconds 2 through 293. The five-epoch set is a subset of the dense set. Keep the same ±1 m affine training-change allowance and all 27 physical columns. This is a new generated geometry diagnostic, not the earlier aircraft prediction split.

One batch native call per case can supply all 293 nominal values/Jacobians; one separate endpoint call checks batch agreement. Verify the aggregate packet-boundary counter against the timestamps. Then compute exactly checked dual support bounds for each future axis under both training sets. Preserve the sparse constraints and certificates; do not infer nonlinear coverage from the linear calculation.

Proposed separate ceiling: six identities, twelve kernel calls, 1,764 queried epochs, 169,344,000 conservatively charged packet visits, 72 LP proposals and 216 exact support checks. No interval integration bounds or aircraft fits. Estimate a few minutes, with a 600-second cooperative ceiling and five seconds per LP. Use one numerical thread after checking other allocations.

Before execution, test the batch/clock and certificate paths, freeze inputs, sources, environment and limits, and verify the prerequisite native controls. If the dense linear ranges are still too wide, diagnose that result before spending hours on complete interval propagation. If they are useful, that is only evidence to prepare the next nonlinear qualification, not permission to claim useful aircraft prediction or weaken any existing gate.