Unfortunate taxonomic results aside, I think the first problem is actually with the merging step. Your data has more taxonomic resolution than the single-end (probably anyways), and so the ASVs that DADA2 picks for each study are going to be different (by virtue of length). That means your merged feature table can never actually shares features between HMP and your study, making downstream analysis a little bit futile.
Assuming you have the same primer position on the forward, you could instead only look at the forward strand and use identical trim/trunc parameters as the HMP data. Then your ASVs would start at the same position and be exactly the same length, allowing the merge to find overlapping features between the two studies.
Alternatively, this is use case can be a good reason to use closed-reference OTU picking (which can be done after denoising so your data looks better). Or you could also collapse your data by taxonomy using the same database for both studies (bringing us to your current issue). But in either case we are supplanting the ASV for some other feature which acts as a shared language between the studies.
As for why your taxonomy looks so stark for your project, I can’t really say. What kind of samples are you collecting, and which taxonomic database are you using? Additionally, what is your primer pair?