I am not sure of my suggestion, tell me if i am wrong:
when adding shadows from depth texture pass, the second pass is project the depth texture from light viewport matrix to the framebuffer interpolating with the camera viewport matrix.
But shall have sence when writing the shader per pixel program at the beauty pass, to add there the shadow, just testing from the depth texrue and setting the shadow color on the same shader just using the light viewport matrix interpolated with the camera viewport matrix, but do it in situ?
this would provide very detailed shadow per pixel accuracy and avoid on projection step where shadow resolution was needed.