The library currently implements a Cache that can be used to speed up the compilation. But that implementation is quite limited as it forces a filesystem-based cache.

I suggest providing a pluggable cache API instead, where the underlying storage can be customized. We could keep shipping a filesystem-based implementation (the one we have today, basically) for people using the self-contained library so that they can have an implementation with no dependency. But I would suggest that we also provide an implementation which relies on a PSR-6 or PSR-16 cache, so that projects can plug whatever cache storage they want there.