think of the cubemap as 6 stitched scene captures, whose scene depth is computed from the perspective matrix which results in something like a ‘rectangle view’ such as of a camera.
instead of storing the cubemap depth try storing the distance from each pixel position to the origin (as easy as in your capture material replacing scenedepth with vector length of worldposition) and you should get an omnidirectional depth from the light rather than 6 stitched perspective depth captures. then the distance from caster to pixel should match