Max,
I've been following the ZD colour blotches issue for a long time, and from all the evidence I've seen, I don't believe that it is a case of reflections. In your park sunset image (MMFC0079-2.jpg aka attachment 37702), the geometry is all wrong to invoke sensor/IR filter/lens reflections as the cause of the purple blotches. They should be strongest on a line passing through both the brightest thing in the image (the sun) and the central axis of the image/system. They are nowhere near that line. They are also too irregular in outline. Nor do they mimic the sky in either size or shape. Sorry, but that theory does not fit the data!
It's also unconnected with the bit depth. Lots of backs have 12 bits depth (including my Kodak Proback 645M), and they don't show this effect. In my astrophotography, I sometimes have to push at least as much as the 2 stops you and Steve have shown, and my Kodak does not behave like this. Such as, I shoot at ISO 400 (2 stops above base ISO of 100) and then push it another stop or two!
I'm afraid that's all I can do regarding this colour blotch problem: I can tell you that it's not some of the things you thought it might be; but I cannot tell you what it is. It is the oddest behaviour that I've seen from a CCD sensor. How does one explain colour correlations over such large scales?
Pixel to pixel variations in the flatfield response are normal, but it's not that either - if it were, then the blotches would be just as visible at high signal levels, but of course they're not.
The same reasoning rules out any other multiplicative/percentage-of-signal line of thinking, such as localised variations in the Bayer filters.
No: the problem is only visible at low signal levels, and that points to additive/subtractive effects. The reflections idea falls into this category which is why it would indeed be logically sensible, but as I've already said, it just doesn't seem to work.
My guess (and that's all it is) is that if a bias calibration frame is internally subtracted, it is "awry" over moderate scales: too much or too little bias is being subtracted from groups of pixels. Just a few counts of bias error could upset the low-signal RGB colour ratios over these scales.
Since bias calibration would be under Mamiya's control and embedded in their firmware, it follows that other backs (like Leaf) with the same sensor would not show the same effect, and that the problem could have been solved by Mamiya by the time they released the second generation of ZD backs. In other words, it's a theory which fits the data, but better data might well prove it wrong!
Ray