Inspect the position-propagation remainder

The six-bound comparison is finished. Its original labels remain failed because the adapter compared coefficients in different units. The saved-only unit audit finds exact agreement for every coefficient after conversion. No numerical run needs repeating to resolve that discrepancy.

The useful outcome is that recomputing the same method on the tested smaller domains changes the height allowance by less than 0.003%. Do not expand this study or apply its outputs as new training constraints.

Next, trace the position-propagation remainder through the existing dynamics, separating included force terms from lever and clock terms. Check whether direction-preserving bounds can avoid spreading horizontal uncertainty into height. Such an improvement needs an explicit mathematical derivation and independent known-motion controls before a new finite prediction comparison. Do not assume that a smaller reported radius is valid merely because it helps the width target.

The original 53-stretch findings, physical parameter bounds, observation assumptions and independent prediction/adequacy gates remain unchanged.