I stopped looking for replacements and set up an experiment that removes components from attention one at a time. I was able to load the training code from the previous theme as-is, so I only added about 40 lines.
Translated from Japanese
commit a7664ca