I'll echo Gang's answer about using the residuals from 3dMVM as likely the best option.
Years ago, 2014 apparently - I did a similar approach to what you proposed of estimating smoothness with 3dFWHMx in this paper: [
www.ncbi.nlm.nih.gov]
That was of course early in the entire cluster exploration process and would still say is a valid approach, especially if you're conservative in your cluster sizes and p-value thresholds. I had also evaluated these data using permutation methods and found the results to be rather similar.