What the saved evidence narrows down

The two affine solvers agree: both propose a decrease of about 0.00112024 in the primary's local objective and 0.00088518 in the repeat's. Column scaling changes those values by less than 0.000000000001. Neither local correction is worse than the feasible zero step. A basic failed affine solve is therefore not the explanation supported by this audit.

The preserved full-model histories show the opposite behavior for those complete corrections:

The largest normalized correction coordinate is time offset, accompanied by initial horizontal position. The full corrections imply about +3.98 ms and −3.46 ms of time-offset movement within the unchanged ±0.15-second allowance. The saved endpoints themselves have almost identical fitted offsets: 0.1131511098 and 0.1131512051 seconds. Their close agreement is numerical repeatability, not a measurement of timing accuracy.

This points toward examining the time-alignment direction and how its local derivative behaves over a finite correction. Both saved endpoints record zero receiver epochs exactly at packet boundaries; that does not establish smoothness through a correction spanning several milliseconds. The audit does not establish an interpolation defect, a processing correction or a cause in the physical data.

The next targeted check should compare the saved time-offset derivative with bounded full-model perturbations at several small scales, and inspect the same packet/time branches on both sides. That requires a new finite native-prediction freeze, but no optimization. No further fit of this target is justified merely by increasing the evaluation cap, and none is launched by this saved-only audit. The unresolved label remains unchanged.

The audit measurements, original matched-trial results and retained source artifacts provide the supporting record. The separate six-target assessment can proceed under the already qualified hybrid method without changing this target's checks.