From a qualified short adapter to full-history bounds

The native adapter comparison is complete: six cases pass across all 27 parameters, using six native queries and 162 interval steps. It took 299.62 seconds. Its inputs, sources, limits and failures are sealed; do not rerun it. Higher derivatives retain separate known-answer controls because the native oracle supplies only first derivatives.

The compiled midpoint step and packet loop now each pass ten local checks against known answers and the rational reference. A separate frozen 64-packet native comparison covers all six cases and all 27 parameter columns. Measure runtime and memory rather than assuming that these short controls make recording-length ranges practical.

The rotation-switch component bounds both local formulas and their switch differences. It is separate from the completed adapter. Propagate incoming changes and the effects of switches through subsequent steps; do not infer a smooth trajectory from local bounds. Clock branches need complete support and one-sided changes too. Keep native cumulative timestamps, ideal integration periods and execution error distinct.

Then bound the residual after the exact shared-source expression, preserving cancellation. Independently bounding two large trajectory/source expressions may erase the useful information. Test useful width over the original full parameter and clock domain before any wider qualification study.

Whole-period prediction coverage, native execution error and a defensible aircraft error law remain separate gates. No new observed fits, changes to completed studies or expansion of active allowances follow from these controls.