That's tricky, for sure. My 'workaround' has long been converting codepoints into byte sequences and creating a character dictionary from that. Based on the source corpus, this dictionary can be further expanded/compressed and used for downstream processing.