Actually PD4ML caches CSS documents in parsed form, but, you are right, it does not cache CSS documents failed to parse. I’ve added the feature request to our TODO list.
I do not believe you use CSS stylesheets, greater than 1 Mb in size. Otherwise it should not tangible impact HTML-to-PDF conversion performance.
Well, the best solution would be to support the notorious “filter:” properties – we need to check the actual state of CSS Parser project and adapt our library to it (if it already supports the syntax).