I have similar problems (CAD to GDML conversion), but so far I was thinking that the overlaps simply “shift” the boundary crossing depending on the direction of the step. (I.e. the boundary crossing would happen on the edge defined for the current volume.)
In our case every volumes’ mother is the world volumes, and there were some problems when calculating the number of geantinos hitting 6 identical volumes (the hit numbers were not consistent).
After some tests it turned out that many geantinos missed the boundary crossings: