Hi @nounou,
The heterogeneity spacers are wonderful to have in your design as they greatly increase quality in sequencing but they are a bit of a hassle to deal with bioinformatically. You certainly need to remove those as well since those are also biologically meaningless and lead to false new ASV picking.
The loci-specific part of the primer can technically be left in but I know others claim higher accuracy with their taxonomy assignment when those are removed. Those reads are more or less the same across all reads and don't provide any additional biological information and in fact they may reduce the accuracy of alignments later.
Unfortunately qiime2 doesn't have an easy method to deal with variable primers, and this has been brought up in other posts. You may want to search those and see if there were any solutions posted. So you can either find a way to do this outside of qiime2 or as you suggested trim a constant number of bp from your 5' that is larger than your longest HS lengths. The downside in this method is you are losing some resolution but also you will still create some artificial ASVs since the same exact taxa can have a number of different ASVs just based on how many bp were trimmed from it (I wouldn't suggest this method).
You may also be able to hack something with q2-cutadapt by running your data a bunch of times with each time searching and cutting one HS spacer. I have no idea if this will actually work, I've never tried it, just throwing some ideas around:P
Lastly, I'm still very confused about the whole metadata inclusion part. Neither denoising step with dada2 or the feature-classifier actions even have an option to include a metadata file, so I'm guessing there is something else you did different between the runs. Regardless, you'll have to re-do them anyways with HS spacers/primers removed so let's see how the new run turns out and go from there.
Sorry I didn't have any perfect solutions for you. Keep us posted!