Fixed stereo appearance ambiguity diagnosis

Pre-registered2026-10-30 after reviewing earlier disparity-band/cycle experiments. No matcher, depth, plane, mask or motion changes. All18 original real frames. Recompute exact pinned restored SGBM arrays before probing. Rectify raw grayscale left/right images identically; use frozen forward disparity at nearest rectified pixel corresponding to a source-camera pixel. Do NOT substitute smoothed plane depth for the original stereo correspondence.

Sample the frozen predictive surface domain at every4pixels (top-left crop phase0). Probe5x5 patches at disparity offsets[-8,-4,-2,-1,0,1,2,4,8]. Mean-center each patch and use mean absolute grayscale difference (zero-mean SAD), bilinear sampling in right image. Require positive disparities for ALL candidates (base>8px) and image/ remap-valid support at every candidate; retain sampled versus valid counts. This common-candidate population excludes far disparity<=8 and border samples; no claims about their ambiguity. These local costs are NOT SGBM's regularized matching cost.

Report texture std quantiles, base cost quantiles, alternatives>=2px away within .5 gray units of base (ambiguity), alternatives better by>1gray unit, and texture std<=5 fraction. Thresholds are fixed descriptive diagnostics, not confidence gates. Group all, reference-normal-agrees(<=15deg supported), disagrees, distance>1m; reference labels are scoring only. Stratify original depth(0,10],(10,20],(20,40], (40,80], plus pooled(0,80]. Do not fit or select a gate from labeled groups.

All18 frames, same tests and no exclusions. Check known synthetic shift and flat patch ambiguity, archived raw-depth reproduction and hashes. Compare frame-level statistics, not independent pixels; no causal claim from mere association or posthoc pooled medians. No refitting, ICP, default change or SOTA claim. If simple appearance cannot separate failed surfaces, report that rather than retune it.