now I have a guess:

the depth-view without NOLOD applies lod models of a model if exist,
but depth-view looks in opposite direction of the camera,
so what is close to the camera is far from the depth-view's "camera",
if the depth-view finds no lod model for lod2-3 (when camera can see lod0),
it has no reason to cast a shadow by an empty model.

is it right?