Hey there @ben!
Thanks for the examples, that makes a lot of sense!
I think either approach (filtering your FeatureData[Sequence] or not) makes sense to me - if you were staying within QIIME 2 that wouldn't cause any kind of mechanical issues with things like diversity metrics or taxonomic assignment, since features that aren't present in your FeatureTable[Frequency] would just be dropped from the tree or sequences. I am not 100% how this will work in phyloseq (if extra tree tips will cause a problem or not). Perhaps it is worth running things both ways for a subsample and comparing the results? The other aspect that might be worth looking into is the actual tree-building process - you will most likely see different trees, depending on the features present, or the tree building method utilized (right now the q2-phylogeny plugin uses fasttree, but @SoilRotifer added in RAxML support which should be out later this month; as well @Stefan has a q2-fragment-insertion plugin, which uses a fragment-insertion technique instead of a de novo one to build a tree). Anyway, looks like you found a few resources that will help you with the actual process of filtering FeatureData[Sequence] using a FeatureTable[Frequency] - thanks for linking to those here! Keep us posted, and let us know if you have any more questions! 
:qiime2: