ginkgo

That looks as if the time at which you render the frame doesn't coincide with the time you render the shadow map.

Did you render the shadow map in a previous frame or update the camera settings between sm and frame rendering?

Might it be possible to do raycasting along vertical lines but then setting the resulting pixels along diagonal lines using Bresenham's algorithm?

The Bresenham pattern would be the same for all lines, so there should be no holes. It could also be stored and reused. The angle of the lines should be the inverse of the actual roll. You also need a height offset, otherwise you just get a skewed image. I guess there might be some artifacts due to the Aliasing in the Bresenham, but that should be acceptable.

Z is the z coordinate of the vertex.

If z = 0, then you have a problem with your perspective. A vertex is located on the eye-plane.

This division by zero problem can't happen in regular graphics pipelines because the depth-division happens after near-clipping.

I guess my first attempt for avoiding this problem would simply be adding a very small bias value to z if it's 0.

You can do perspective correction yourself by interpolating iz=1/z and iuv=uv/z as varyings and reconstructing uv in the fragment shader as iuv/iz.