I am not sure until I run some tests, but my strong belief (based on good analysis) is that that you are hitting memory latency in your CPU with the 2MB Scrypt, but the GPU will not be. Thus it is going to toast your CPU. This is not a minor problem, rather I conjecture it is total failure to get what you intended. The reason I suspect this, is because the GPU can run 6 x 1024 MB ÷ 2 MB = 3096 threads and thus entirely mask away the latency. Whereas you do nothing to mask the latency on the CPU as far as I can see in your code.
Well it's anticipated that specialized hardware will be able to run parts of the algorithm more efficiently. However to run the whole hash efficiently you'd need to run all three parts efficiently (four counting SHA512), and specialized hardware to do that would look a lot like . . . . . . a CPU.
You appear to have a serious fail here, but I would need to build a CPU miner to test my hypothesis.
Interested in any diagnostics you're seeing on the CPU miners we've got available.
If instead you don't want to share synergy or simply have other priorities demanding your attention, then that is fine too.
I would have been interested in a detailed discussion on these points 2 weeks ago when there was more scope for re-engineering. As it stands now, the show is on the road so I don't want to be switching the engine unless it's clear the one we've got is failing. If you can demonstrate that you will have the pleasure of my undivided attention.