i wanted to add to this thread my perspective as developer/maintainer of the DADA2 software package.
We think DADA2 works fine with binned-quality-score data such as NovaSeq, at least in the vast majority of cases. While there are some oddities when inspecting the error model that DADA2 learns from binned quality scores, there don't seem to be oddities when evaluating the results of denoising on NovaSeq (with the caveat that this is based on a relatively limited number of test samples). Furthermore, the oddities in the DADA2 error model with binned quality scores often appear more important than they are -- typically the error model is well-fit at the few binned quality scores that dominate the data, and the visual deviations that appear in diagnostic plots occur only at quality scores that barely appear in the data. That's a consequence of the simple but fairly robust loess fitting used to learn the DADA2 error model.
To sum up: As far as we can discern, DADA2 works fine with binned quality score data like NovaSeq. Undoubtedly it could be optimized further, but I can say that as the developer of DADA2 I use it on NovaSeq data without concern. That said, as always -- do your sanity checks!