Thanks Hummel for your effort!
I forgot to mention that it already worked with a tracing in world coordinates but all the matrix multiplications inside the loop were way too slow on my notebook, that's why I switched to screen space tracing. My results looked pretty bad compared to your version, though (they can be seen in the WIP thread), so your effort hasn't been in vain.
Note that I changed the depth output in Depth.fx to linear depth.
Hm what does that mean exactly? The issue with the clinched reflection has to result from something that is "not linear" (as one can easily see on the crates) and I've always doubt that my approach will work because of non-linearity somewhere, but the source I've linked in my first post proves me wrong.
I don't achieve linear depth when I remove the multiplication with the Projection matrix, do I?
Multiple reflections and the like have very low priority, but thanks for the hints.
