I spent half of the last night poring over the datasheets and wading through horrible C code examples, but finally I managed to find the code that was used to make this example:

Inside I found a lot of scrambled code, unused files, copy-pasted fragments, but also the thing I was looking for: the LUT tables for the refresh sequence! After adding them to my driver code, the full refresh suddenly takes less than 2 seconds, and the "quick" refresh is practically instant (but of course leaves some mess from the previous image). That is so much better!

I have published my driver code here: https://bitbucket.org/thesheep/circuitpython-epaper/src/