@IVX, understood.
FWIW, just yersterday night I found an unbelievably simple and reliable break-through tweak to get a THD(-N) of -140dB (pure H2) from 2 AK4493's summing the four channels, consistent vs. level and frequency. With a bit of care I might even reach -150dB, APX555 territory. Or get the -140dB with only two channels....
Will post an article once I have it fully fleshed out.
Short update... As to be expected, things at these precision levels are not so easily handled.
It turned out my initial measurements were better than real life due to a damn lucky combination of cancelling mechanisms at work. It also turned out (again not unexpected) that this is very sample dependent, with some chips I got excellent results, with others not so excellent ones (I have only two AK4493 and two AK4490 to test with, as incorporated in the RME ADI-2 Pro FSR and FSR models, resp.). Only two channels are needed, using two chips does not give that much improvement and just opens to many variables for optimization.
At least I now know the limits of my AP better than before and ways to improve resolution.
@IVX, rather than using the limited resolution built-in FFT of the AP I'm now using the monitor out for FFT and set the input range to about 10x the auto level detect would normally select, in order to reduce internal voltages as much as possible. Input set to DC to avoid that cap's influence as well. By this I think I can get valid results down to -150dB for 3rd and -140dB for 2nd. 2nd still has variation depending on signal polarity but now it is within small bounds so that I can use the median value as an estimate.
From what I've seen now I think -135dBc is realistic with a "typical" chip, at -6dBFS. 3rd dominating, it only reduces a little by paralleling channels with different DC bias (not necessarily 50%). Even order products are supressed quite well, though.
All this will be valid to the specific circuit and layout of the RME devices (with the only tweak of connection the Vrefs together, atm). In different implementations one might get different results.
I've learned a lot about the specifics of the chips wrt t Vref and Vcm characteristics and after a heavy web search I now found a key document, a presentation from AKM that allows insight why the chips are so damn sensitive to VrefH/L supply voltages and impedances, same for Vcm pin impedance. It's a bit dated and dealing with the previous generation AK439x models but I'm pretty sure the basics of the Switched Capacitor DAC core have not changed much: