Qualify the reference coefficients before native use

The source preflight demonstrates a tighter value range while preserving the actual training bounds. The next task is to qualify the reference linear map for the same A=R f and B=R(P−I)f source, retaining all 27 physical columns.

Compute the full reference coefficients independently and compare them with the coefficients used by the native body-rotation calculation. Preserve the clock's exact branch convention and record distance to packet boundaries. Unused coordinates remain explicit zero columns. A boundary-specific reference map must not be described as a unique two-sided derivative at a kink.

For any verified reference row J and value interval [l,u], a conservative affine-error radius about reference value c is

max(|l−c|, |u−c|) + sum_j |J_j| max(|lower_j−center_j|, |upper_j−center_j|).

This bound works even when the linear map is not an exact derivative. However, the row must be the one consumed by the propagation method, or an additional coefficient-discrepancy allowance must be retained. Compare the resulting radius with the original remainder before any expensive integration test. A narrower value interval need not produce a narrower affine-error radius.

If that comparison improves the actual baseline, develop the dynamic candidate with componentwise operations and preserved training-conditioned information. Keep native centers, all shared errors, original physical bounds and acceptance limits. Qualify independent motions and the original native control matrix before a separately frozen complete comparison. No observed fitting follows from this preflight alone.