Disjoint spatial prediction test for coherent surfaces

Pre-registered 2026-10-30. Compare the frozen21px estimator with an additional measurement-only model test, same27 input frames:54 rows. Retain plane depth, normal, confidence and original selection. Candidate only removes points.

Split each21x21 patch by local checkerboard parity:221 fit pixels,220 holdout. Fit affine disparity on the221. Predict the disjoint220. Under independent .5px Gaussian noise and a correct plane, prediction covariance is .25(I+Xh(Xt^T Xt)^-1*Xh^T). Use the exact Mahalanobis statistic via Woodbury, with220 degrees of freedom, accepting at chi2_220(.99). This incorporates common coefficient uncertainty rather than treating all predicted residuals as independent. Verify10000 exact-plane noise draws (seed20261031), acceptance .98–.997.

After this validation rule, report the original full-patch21px estimate for accepted locations. Full-patch refitting uses holdout pixels; this is a diagnostic model validation gate, not a claim of independent post-selection confidence coverage. Checkerboard noises are independent only under the frozen synthetic noise model; actual stereo errors may be correlated. Curved patches can predict held-out pixels well without having the correct center normal. Selection must be judged against analytic center geometry, not just prediction scores. No truth enters the gate.

Keep all clean/noisy/frame conditions and the common domain. Same joint precision

=80%, retention>=10%, >=8 cells and normal singular ratio>=.05 in EVERY noisy frame for continuation. Also report angular coverage. Reproduce archived21px baseline and hashes; missing data explicitly reject. No post-result threshold tuning, real filtering, ICP rerun, changed accuracy verdict or SOTA claim.