So one or more of these formats resulted in the graph you posted, and you won't tell us which one? And all anyone has to do is run a simple test and one of the formats will act strange like you showed?
A guess before doing any testing would be it is one of the DSD formats.
If you go back, you can see, I posted graphs for 44.1k alt setting 1, 352.8k alt setting 1 and DSD256 alt setting 3 (PCM rate 352800). Mentioned in the graph headers.
Of those, 44.1k PCM was worst performer, 352.8k PCM was a bit better but had picketed noise floor, and DSD256 was cleanest but still abnormal noise floor modulation.
In addition I think I posted graph showing how 19+20 IMD test tone at 705.6k PCM trips up the output.
PCM input is 24-bit TPDF dithered data.