Remove duplicate class analysis from memory We no longer store the class analysis of a compile task's output in that compile tasks's own cache. Instead, we compute it right before we need it when the next compilation starts. This removes a lot of unnecessary memory pressure, because no one else was reusing that state.
It would be even better if we could reuse the analysis that downstream tasks did on that classes directory (which is an input for them). To do that we'd need to ensure consistent hashing strategies between the two.