At this point I wrote a quick script to match kanji to letters 1:1 in one paragraph, and use replace on the whole text.
This was much faster, and once I had decrypted 10 or so paragraphs like this grepping for rgb yielded the encrypted flag - 璧技丩忄鰏叒讞鸞鸚鱺鑱磒夢洶飳勊淩鼱殆鸝钁厵鬱. However, some kanji had values above 26.
I struggled with this for a while before trying to just extend them on the ASCII table, which worked!