There are two kinds of files we care about; source files (which are cheap to re-read, but re-reading would require us to unwind the graph somewhat in ways I don’t think currently exist), and output files (which have much more vague lifetimes).