Determine whether the learned geometric gate is reliability-calibrated

Determine whether the learned gate that blends the corrected-ray intersection candidate with the learned pose update is calibrated as a reliability estimate for individual joints, and identify which candidate is more accurate for each joint.

Background

EgoRayFormer combines a geometric candidate obtained by intersecting corrected stereo rays with a learned pose update through a learned scalar gate. The analysis reports that the intersection candidate has slightly lower mean error in every distance bin and seed, while the gate retains substantial weight on the learned update.

However, the reported averages do not establish whether the gate makes reliable joint-level decisions or whether either candidate is consistently preferable for individual joints. Calibrating the gate would clarify when the geometric branch should be trusted, particularly under weak depth constraints and varying viewing-ray angles.