I realized that these plots are a little misleading.
The VCS3 filter is highly nonlinear and "bubbles" and so taking a measurement of frequency response is quite difficult since the filter is moving around on it's own accord all the time...!
Here is a more detailed FFT that shows how the filters do match more closely than the graphics in my paper would suggest.
.
Black is the slow accurate VCS3, Red is my optimized VCS3.
Not oversampled.
c4000_fb9_m42__one3_c050_k2.png
c10000_fb9_m42__one3_c050_k2.png
c2000_fb6_m42__one3_c050_k2.png
c10000_fb9_m42__one3_c050_k2.png
When oversampled, I found the proposed model to be perceptually close to the slower accurate model. The resonance and cutoff and gain values differ slightly, but the important aspects of the filter are retained.