I am not sure that DADA2 is the right tool for this, though it might help to hear more about your experimental design. I'd recommend reading up on it or looking at materials like this.
It seems that the mutation rate is relevant to your analysis, such that you are calibrating expected diversity $X$ based on it. Depending on the expected frequency distribution of your variants, DADA2 might be filtering out most of the variation that you're interested in as sequencing errors.
See this note on the ASV approach used in DADA2 (reference):
ASVs are inferred by a de novo process in which biological sequences are discriminated from errors on the basis of, in part, the expectation that biological sequences are more likely to be repeatedly observed than are error-containing sequences. As a result, ASV inference cannot be performed independently on each read—the smallest unit of data from which ASVs can be inferred is a sample. However, unlike de novo OTUs, ASVs are consistent labels because ASVs represent a biological reality that exists outside of the data being analyzed: the DNA sequence of the assayed organism. Thus, ASVs inferred independently from different studies or different samples can be validly compared.
In brief, you will only be able to infer the presence of mutants that occur at a rate detectably higher than the rate of sequencing error.
While I am not a DADA2 expert, what DADA2 is expecting is more along the lines of multiple modes in the sequence distribution- each of which is surrounded by a small scattering of sequencing errors resulting in variants nearly identical to the parent mode. It then uses some statistics to evaluate how likely each variant is to be an error off of the nearest mode. So variants that are very close to an existing mode are likely to be lumped into that mode, unless they are sufficiently high abundance to justify establishing a new "species" mode. This has led to a lot of success for DADA2 in accurately identifying variation in samples of relatively diverse samples with somewhat distantly related organisms each at some meaningfully high relative abundance (>=0.1%).
See for example this figure from the original DADA2 paper:
Every point is a variant identified by DADA2. you'll notice that the variant frequency has to be quite high (~0.1%-1%) to identify distinct variants from sequencing error at a Hamming distance of less than 10 or so from the nearest other variant. Unless you expect your mutant frequencies to be in that range, you are unlikely to detect them. Maybe they've gotten better over the years, but I expect that there are some fundamental statistical issues here for this approach that mean they can't get much better.
If you really want to use DADA2, any parameter that loosens the statistical thresholds will make it better, but I'd guess you've already tried that. I don't know the QIIME parameters, but if you look at dada2 docs, you will see that the dada() function has a number of error parameters that you could play with to try to tune down the probability of errors to increase the likelihood of recovering variants. See the learnErrors() function that you might be able to play with to get reasonable values for this, if you had e.g. a pure reference virus sequenced sample. Note that these functions are for denoising a diverse metagenomic community, not a low-diversity viral mutation experiment! Finally, see setDadaOpt() as ways of globally controlling various thresholds.
I would recommend doing this within the R package rather than through QIIME, to give you better control. If you can't do that, then maybe re-estimating your FILE-stats.qza or your p-min-fold-parent-over-abundance parameter might let you loosen thresholds a bit (though it looks like you are already doing this last somewhat).
outside forums
see also some similar questions on forums.
