I was actually experimenting with an implementation of this initially. However, it would have been iffy - we obviously cannot emulate it by messing with the normals, as that would mean that light coming from the opposite direction would over-saturate the "shadow". Basically, I knew that it would make things complicated going into dynamic lights, so I skipped it.

In any case, the CPU can't pre-calculate the shadows. I would have to be done on the GPU - so for example we could use the light normal at a point and test a few steps into its direction to see whether we find a higher-placement material. Might work, but will have some probably strange-looking side-effects such as single pixels casting shadows at a distance, plus even more advanced craziness when more than one light is in play.