One fixed repair of original-observation probes

Use the four previously saved scale-1/10 generated probes: positive and negative north directions under zero and profiled removal. Preserve the original nonzero truth, all five perturbed training observations, 96000 packets, 27 physical bounds, ±1-second clock and 20/20/50 m future width targets. Earlier failures stay unchanged.

For each saved point, use its existing native Jacobian in one minimum-norm linear solve. Fifteen rows target the known generated truth at the five earlier epochs; one additional row preserves that seed's computed future north prediction. No future observation enters the solve. This constructs a counterexample probe, not an honest held-out prediction or a confidence interval.

Use all 27 columns, SVD rcond 1e-12, maximum normalized correction 0.25 and maximum linear row mismatch 0.001 m. Reject out-of-box points without clipping. These are proposal checks; only a new nonlinear evaluation against the actual unchanged training observations can establish training feasibility within ±1 m per axis.

Save each proposal before querying it. One new native batch per valid proposal returns five earlier predictions and the 438-second prediction. Preserve failures and report sampled spreads including the original truth. Small spreads are not outer bounds; a feasible spread above a width target would disprove that target for this generated observation problem only.

Freeze source, input references, environment and compiler before launch. No iterated repairs, alternate scales, retries, signals, observed aircraft fits, original-data reads or allowance increases. This does not change the running long-history study, qualify native execution error or establish an aircraft error law. Existing scientific results remain unchanged.