After querying CDS folks about this issue, I got some responses that indicated the problems was likely limit-cycle oscillations due to zero-padding of the data when upsampling. Tobin ran some Matlab tests to confirm this issue.
Starting in RCG 2.5 there is a new "no_zero_pad=1" cdsParameters option turns zero padding OFF. I tried enabling this option c1scy to see how the behavior changed. Sure enough, the 32 kHz oscillations mostly went away. There are no oscillations for outputs held at the half-count value, and the oscillations around the half-count transitions went away as well.
The only thing I could see is a bit of oscillation when converging on a constant half-count value that went away after a couple of milliseconds:
So we might consider adding the no_zero_pad=1 option to all of our coil driver outputs, which might eliminate the need to add notches at the Nyquist in the analog anti-image filters