OK, that confirms what I suspected was going on, thanks for the clarification. This sort of explanation should be documented…
By the way, does the same apply to mother-daughter overlaps, ie when a daughter sticks out of its mother V1 into another volume V2? In that case, I could also imagine that the geometry tree acts in our favor: since V1 and V2 do not overlap, the overlap region would always belong to V2 and not to the daughter. But since I have no idea how navigation is done, I suspect that is the wrong guess 
I was playing around with voxelizing arbitrary volumes using replicas, and at least the overlap checker did not complain if some voxels were outside the mother.
(And if I may shamefully direct you to this other topic I opened and ask for your opinion…)