Agreed on all accounts. It really depends on what you are trying to do.
While I think most agree by now that denoising methods are superior to q-score filtering/trimming what you do after for example cluster or not cluster is as you mentioned specific to the research question.
Not clustering approaches do have one particular advantage that makes them more desirable than OTU clustering in that they are comparable across studies (as long as they are of the same region) and don’t require re-clustering of your whole data every time samples are added.
I collected some readings and opinion pieces regarding this debate on another post here that might complement this discussion.