Ok, I dug into this some. I think I see the issue.
I made my measurements at maximum level. Here is what it looks like there with and without load:
View attachment 10963
As we see it is clipping with 32 ohm load. Archmiago used 20 ohm which would make this even worse.
I then dialed down to where there is no clipping and now the loaded output is very similar to unloaded:
View attachment 10965
Now the output impedance measures 0.4 ohms.
I was careful in looking at the waveform when I first started but then I stopped doing it. My bad.
I will correct the original post. This also explains the difference in some of my previous measurements if the device was clipping with 33 ohm load.
Thanks for bringing this to my attention.