But how many top hits are equivalent and hit other taxa? Remember BLAST will arbitrarily sort top hits that are equivalent. But in this case, it appears that most of the top 500 hits are to Akkermansia using general megablast search, with some other different near neighbors. But using the 16S/ITS db w/ megablast close hits from other taxa are observed too. But again, as I've referenced above there are over 200 exact matches to other taxa in GTDB. So, the tools is telling you that it cannot disambiguate these differences. Perhaps these are database annotation issues? I do not know. You could try constructing (clawback) / downloading environment specific classifiers from SILVA.
Again, it depends on how often close matches are hit and contribute to the taxonomic assignment. Also, many databases might suffer from taxonomic and/or other curation biases, which can alter the assignments. See Figure 5 of the RESCRIPt paper. Which is why, it's a good idea to try different approaches.
I'd avoid the majority method as it does not work like most people assume. See here. At least make sure here are no hybrid taxonomy strings. This method assumes that the annotations are of high quality and rigorously curated to avoid hybrid taxonomy string outputs. But as you set --p-perc-identity 0.99, that might be okay, and mitigate hybrid taxonomy strings. Though, there are many different taxa that might share exactly the same sequence but have differing taxonomies (as we are observing with your query against GTDB above) . Hence the uniq mode. You could try altering the --p-confidence of classify-sklearn and see if that helps. Probably, also try classify-consensus-vsearch or classify-consensus-blast too?
There are many clades that cannot be disambiguated at the genus / species level even if you have full-length 16S rRNA gene sequence. This is expected for the 16S rRNA gene regardless of reference database used. This is why SILVA explicitly does not curate to the species level taxonomy. In fact, we only provide "species" labels as an option in RESCRIPt, and we warn about using this option in the SILVA tutorial. In a nutshell, if you do observe many species-level assignments, this is likely a problem of over-classification. At least I'de be wary... even with full-length 16S rRNA gene sequences. Remember, the 16S rRNA gene is actually a very conserved gene overall. If you'd like to obtain high quality species / strain level assignments, then the use of other markers and/or shotgun metagenomics should be used.
Finally, do not base the apparent quality of a reference database simply because it returns more or less genus and/or species level annotations, these could be signs of under-/over-classification problems. Also, this is why RESCRIPt exists, you can fetch reference data from any database and curate to your needs in order to reduce the assignment problems you are referring to.
Yeah, taxonomy assignment is not as easy as many think. 
I am not sure I have any further ideas at the moment. Perhaps other people smarter that I have insights? But please do keep us posted on what works.
Perhaps I should reference these pipelines in the SILVA tutorial? 
