Completed matched prediction comparison

10/12 fits converged. Charged starts: 12; native sweeps: 2284.

Training shape preference: unresolved. Prediction-cost shape preference: unresolved.

These are retrospective numerical preferences under two shared receiver-error assumptions. Unfinished fits remain unresolved. Prediction innovations condition receiver error on training residuals; their covariance excludes bounded calibration uncertainty. Axis probes and local rank are diagnostics, not calibrated intervals. This comparison establishes neither statistical significance nor the cause of a background rotation field. See fit details and frozen method.