I prefer to use the non-corrected images for the reasons you ask. You cannot include exact motion regressors in your regression model if you use the MOCO. The fundamentals are mostly the same, but I prefer to trust in AFNI to do the heavy lifting. It's easier to see exactly what the software is doing!
Your script looks mostly standard based on the examples that are provided in afni_proc.py. I usually only use the despike block in resting state data, and not in functional data.