For bundle adjustment optmization why is reprojection error minimized instead of traingulation error or ray-to-3d point error? Disadvantage with reprojection error is: for perspective camera, its highly dependent on depth of the camera from the point in 3d right.
Triangulation error is not measurable in most cases - unless you have an alternate way to get at the 3D points being triangulated, you normally have no ground truth about them. Also, in normal situations where global scale is unknown, triangulation error is only defined up to scale.
Reprojection error is the gold standard error measure because it is directly measurable and it is unaffected by global scale,