Hi Kevin,
I think your assumption regarding the amplicon length is correct. But for training the classifier I wouldn't go with too tight interval params. You may be risking removing biologically relevant sequences due to differences in 16S rRNA gene in different prokaryotic groups (as you mention also in your previous post).
I am pretty comfortable with --p-min-length 400 and --p-max-length 500 as this gives me enough error margin, and certainty that I am recapturing relevant sequences and excluding some larger artefacts 