Conditional cycle diagnostic after predictive surface selection

Pre-registered2026-10-30. Review found the prior1px cycle mask insufficient in stereo translation refinement (stereo_cycle_filter_comparison.md); do not rerun that optimizer or tune its thresholds. New question: does the exact same reverse consistency test distinguish wrong surfaces inside the NEW predictive21 selected population? This is conditional measurement diagnosis, not a calibration trial.

Reconstruct forward/reverse SGBM with frozen settings/OpenCV and re-rectification from joint_learned_stereo18 frames. Restore masked depths identically to the old cycle code; only archived saved-depth suffix changes from rerectified to sgbm. Require exact raw restored-depth equality for all18 and exact old filtered-depth equality on the12 previously tested frames. The6 other frames are already-spent 0013/0015, not fresh data. Pin original images/calibration/metadata and hashes.

Keep fixed <=1px finite round-trip consistency. Cross it with predictive21 masks from the frozen error maps, without refitting a plane on the masked depth. Record all-selected,cycle-pass,rejected-or-missing counts, domain retention, distance failure rate (>1m among selected), normal failure rate (>15deg among supported pairs), with original reference maps scoring-only. No new masks enter a solver. Rejected and reverse-missing are grouped; do not call them all wrong matches.

Retain all18 frames and emphasize prior3 failures without selecting favorable frames. Compare paired rates/counts and explicit denominators; dense pixels are not independent recordings. A consistent match is not ground truth. If the three failed frames remain below80% normal agreement, do not present cycle masking as a remedy or run ICP with it. No gate/threshold adjustment, fresh-data use, core change, calibration verdict change or SOTA claim. Freeze code/input hashes first.