Sounds reasonable. And running just the forward reads is a great idea just in case (and if nothing else, it is a useful comparison)
Based on all the above discussion, I would think yes subtract the primer lengths from the total amplicon length you have cited, unless if you aren't sure whether that length includes the primers. (to be on the safe side perhaps we should just assume that it does include the primers and you so would not subtract them).
It is lower that I like personally, but really does not sound bad. I would personally opt for fewer longer sequences than more shorter sequences, so long as I have enough reads.
So it is definitely worth running all ways and compare the stats output to decide what works best for you.
You could also experiment with the --p-trunc-q
parameter instead — since you are joining pair-end reads it may prove useful here for getting more length out of higher-quality reads, instead of trimming all at the same length.
Sorry this is such a difficult trial-and-error process! The good news is all steps after denoising are usually easier..
I hope that helps!.