In a previous post I mentioned the less-than-ideal performance when Carbon is powered by a ±400V supply, and I suspected that the lower Vce on the PZTA42 is the culprit. Now it's been proven. The PZTA42 being a high voltage transistor, has a non-linear region at low Vce, as the slanted curves you can see on the upper left side. With 407V on the negative rail, the transistors on my board works at Vce=8.4V and Ic=20mA, right around the knee. The global negative feedback would have a hard time correcting that non-linearity. It also explains why some people prefer setting the Carbon at a lower current when powered with ±400V supply, as it also improves the linearity of the PZTA42, albeit to a lesser degree.
I guess Kevin chose the high voltage PZTA42 to deal with the power-on transients. I have a quick and dirty fix. Just bias the SiC MOSFET a little higher to give the PZTA42 more headroom. The SiC MOSFETs are biased by two 175k and an 20k at the gate. Reducing either 175k or increasing the 20k would do. The goal is to move the PZTA42 operating point to the right, well into the constant-current region (parallel lines). I would use Vce=14 to 15V. Pushing it even higher would increase the power dissipation on the PZTA42, eat into the max output voltage swing and have diminishing return. What I did was to put a 260k resistor in parallel with one of the 175k resistors. YMMV because it has to do with the operating point of the PZTA42 in your circuit, the Vgs(th) of your SiC MOSFET, etc.
After the quick fix, one of the channels now measures as good as with the ±450V supply. We can see that the max output voltage is slightly less compared to with ±450V supply. The difference is subtle with the log scale, though.
Now I'm continue to work on the other channel and see if I can find something else.