Pretty useless if it does, since you'd be limited my the DR of the gamma-compressed, DCT-compressed JPEG files.
Considering where the problem resides, I'm fairly sure 3 gamma-encoded 8bit channels trump 1 linear 16bit channel any time of the day, even with all of the heavy JPEG compression. Still, what you might try is how the camera reacts to severe shadow punishment as much as it allows before it spits out that jpeg. Another issue of course is blackframe subtraction which I believe Fuji allows with a high iso noise reduction setting some where IIRC.
How can you tell? The pixels in the raw file are interpolated over.
Because I'm looking at the CFA data, see attached. It represents the bottom of the DPR test file, generally containing the problematic pixels. Pushed the levels up quite high, but would be hard pressed to say I see significant banding of any relevant kind. Admittedly, I need to write some code to get a graybalanced representation, which can sometimes reveal more problems. I'll do that tomorrow. If you have a raw file you want to specifically look at for those pixels, I might be able to run it here.
The PDAF pixels are not used to form the image in the raw file. Their locations are filled by the camera using interpolation.
There is something in Fuji's fixing up of the PDAF pixels in camera that gets the mean wrong. This is not the case with Sony OSPDAF MILCs. On those cameras, the sigma changes in the PDAF rows, but, absent flare, the mean is correct.
Okay, I might give that a check tomorrow as well if I get around to it.
In 16 bit mode, those channels have 16 bit precision, just like the green channels. Read noise is independent of raw plane assignment, to a first order.
Yes, I understand, but apparently the 16bits of the green channel do not provide increased DR, at least as I understand from the discussion. The question though is what happens for the lower values of the R and B channels? If those values are encoded with actual more precision (not just rescaled from 14bit) then there is a significant advantage for raw converters to compute crosschannel mixes. This can reduce the noise produced by the converter (not by the camera).