the normalized device coordinates is still a 3D space, it is just rectification that transforms the view frustum into a cube such that all rays converging to the eye become parallel to the z axis.
Occlusion is handled in this space through z-buffer: each point is mapped to normalized device coordinates and then the value stored on the z-buffer at its transformed x,y coordinates is confronted with the transformed z coordinate. If the value on the z-buffer is bigger the point is drawn and the z-buffer updated, otherwise the point is occluded. The transformation I propose comes before occlusion detection and you still need two z-buffers, but you can compute transformation for each triangle only toward one eye and then transform the coordinates to those of the other eye.

the last entry in the first row should be d(f+n)/(2fr) ;)

nah, you are making it too complicated by thinking in world coordinates.
You can transform points in normalized device coordinates (after applying the projection matrix).
The transformation from one eye's coordinates to the other is just a 3D homography, so you can map a point in left eye's coordinate to one in the right eye's coordinate (assuming a symmetric frustum) by multiplying it by the matrix

[ 1,0,-d(f-n)/(2fr),d(f+n)/(3fr)]
[0,1,0,0]
[0,0,1,0]
[0,0,0,1]

where r is the location of the right plane (and -r of the left plane), f is the location of the far plane, n the location of the near plane and d is the eye displacement.