Corrected JUNGFRAU data in burst mode (16 cells) was getting a mask which was as expected for the first memory cell used, but in the other 15 cells, every pixel was marked with WRONG_GAIN_VALUE. The mask is used for the ROI reduction, which was therefore giving nan.
The issue is in the BadPixelsDark10Hz constants. For medium & low gains in burst mode, all but the first memory cell are marked as invalid. Then a check for the gain stage of each pixel gets nan, because there's no data, so it thinks all pixels are in the wrong gain stage. Finally, the code added in !914 (merged) 'reflects' WRONG_GAIN_VALUE into the high gain stage, where we do have valid data.
The solution here is to skip the processing entirely when no data is found for a memory cell. Where this is the case, I set the bad pixel constant to NO_DARK_DATA. It also reports which cells are missing.
Run the dark notebook interactively, working with p6616 runs 272, 273 & 274 (JNGFR02) and checked the resulting constants against the existing ones from this report.
offset[..., 0]) and the same for the first memory cell used (offset[:, :, 15]).offset[:, :, :15, 1:]), the offset changes from nan to 0. I can put it back to nan if we prefer.badpx[:, :, :15, 1:]) go from bad pixels WRONG_GAIN_VALUE | OFFSET_NOISE_EVAL_ERROR to now NO_DARK_DATA | NOISE_OUT_OF_THRESHOLD | OFFSET_OUT_OF_THRESHOLD, because of 0s replacing NaNs.NO_DARK_DATA | OFFSET_NOISE_EVAL_ERROR.