The LCD solution was the first one that came to my mind too, and I'm surely missing something here but why wouldn't the regular translucent pixels suffer from the same problem as that above?
At least I think this should work, combined with being able to control the translucency of the pixels would allow it to be fairly accurate.
In current commercial HMDs, the additive pixels are statically lensed to a fixed focus distance (not infinity, as a sibling comment asks, but rather a best guess of where the user will be focusing most often, e.g. 1.25 meters).
Using dynamic lenses to put the additive pixels at any focus distance is an exciting topic of current research:
http://www.quora.com/Volumetric-3D
Dynamic lenses could also be put on the forward-facing / physical-reality optical path both before and after the subtractive pixel layer, to put the subtractive pixels at any distance as well, solving the problem. (This would incidentally also allow lensing physical reality itself to any distance too, creating perfect / hyper-real vision correction.)