Hi @mycol,
Your overall strategy appears fine.
I assume when taking into account reading through into the opposing primer at the 3' end of the reads you are doing so by feeding cutadapt the reverse-compliment of that opposing primer sequence? I ask because, it is often forgotten to take the reverse compliment. If not, this may explain why so few reads are merged/returned via DADA2, as the opposing primer is contained within both of the R1 and R2 reads. Also, make sure you've not trimmed to much off of the end of your reads. I've had problems when many of my ITS reads were to long to merge regardless if the data were clean. I refer you to this paper to help you decide if you should use only the forward reads or attempt to merge them.
One thing of note, it has been recommended to me a long while ago by one of the ITSx authors to try leaving the primer sequences in your data prior to using ITSx. Note, for this particular case it means that you'd only try trimming the opposing primer that has been read through, and you'd leave the 5' primer in each of the reads. Then merge them.
The primer sequences make it easier for ITSx to find and extract the region of interest, potentially retaining more reads from ITSx. I am thinking it may be possible to merge the reads via vsearch after you trim the 3' primer sequences, and then run ITSx on the resulting FASTA output. Anyway, I only wanted to mention this for an added means of sanity checking other steps should you need to. Otherwise continue as you've done, trim all the primers, then use ITSx. Then use vsearch and de blur.
The short of it, I would suggest remaining within qiime for now (i.e. use the vsearch plugin to merge reads) and then make use of deblur. Then determine how many reads to retain or lose compared to DADA2.
-Hopefully this makes sense.
-Best