With cr2hdr, the flicker comes from this:
Linear fit : y = 7.9666*x - 0.65
Linear fit : y = 7.9500*x - 0.64
That is, the two exposures were matched slightly differently in the two images. That's because cr2hdr was written with still photos in mind, so it analyzes every single frame.
Solution: the algorithm should be updated to operate in two passes: one to analyze all images, and another to process all images with the same parameters.
Why did this large difference appear?
These two frames are underexposed - by at least 4 stops in the high-ISO frame. That is, the algorithm didn't have good data to perform a robust exposure match. Running cr2hdr with --iso-curve, to debug this process, shows some weird graphs.
One of the green channels from the bright image saturates in an unusual way. Not sure if that's a problem only in this particular clip, or in all clips recorded at the same settings. This could be a bug in crop_rec, or a quirk of the 1x3 readout. In any case, it's an issue with the raw image data (at capture time).
BTW, there's no point in using Dual ISO 100/800 in very underexposed scenes. You will get much better results with plain ISO 6400.
TLDR: edge case, can be fixed, not a priority for me atm.