Perhaps it could be beneficial for you to ask the Renderman developers to explain how their adaptive sampling works in RIS. It's not splotchy at all and it works really well. Even if they don't spill the secret sauce, you might get some ideas from them you can use.

On your patch I get a segfault flattening all the faces of the monkey head. Can you reproduce?

Here's a Windows test build for your testing enjoyment.

Blender build with D808 applied causes Cycles experimental SM35 CUDA kernel to fail to start with an out of memory error on my GTX 570. It works fine on my GTX 780, and the supported kernel can run the same scene on the GTX 570. In addition the experimental SM35 kernel fails to start on an empty scene on my GTX 570. If there's a fundamental reason why this shouldn't work on Fermi there should be a way to automatically disable building it for SM35 and under, and if there isn't a fundamental reason why this shouldn't work on Fermi and under then it shouldn't crash on startup. Can anyone confirm? Running Ubuntu 14.04 LTS.

Great, thanks a bundle! I'm a big fan of your work on this FBX exporter. It's a big improvement over the last one.

My memory usage dropped to 486 MB when I joined all the bricks together when baking a 1024x1024 image (I opted not to also join the plane in, as it doesn't have the same modifier stack.) It's a lot better, the memory usage is nearly on par with XNormal (343MB for similar baking parameters.) Unfortunately, Blender ended up crashing at the end of the bake. Perhaps I will try and build from git and see if things have improved any in that regard, or if I can get a backtrace.