also there's a slowdown on the criterion benchmarks for directory materialization from #9779 so @witty-crayon-22786's remark about how it would be awesome to avoid duplicating glob matching is still relevant. the duplication of the glob matching code is likely responsible for the slowdown. i think i'm actually going to ask for holding off on review for a bit longer while i separate #9779 into yet another PR, one which separates the changes to

subset()

from the other changes (adding the benchmarks, etc).