I'm not sure about the "only" part: for usability it's better if the system is completely agnostic to details, or I may correctly enter all diacritics and have my solution rejected for no reason.

When implementing this we're probably going to use some standard Unicode solution for case folding and diacritics/accent folding.

Yes, this is absolutely necessary. Not only diacritics might not be visible, but also some users may not have the keyboard to enter them.

I am not sure how to implement the folding, and it may even be language-dependent. For example, users may enter 'ö' as 'o' or as 'oe', or 'đ' as 'đ', 'ð', 'd' or 'dj'. A possibility is to simply avoid words with diacritics, which should be possible for most languages.

In future, when non-Latin captchas are implemented, the same should apply to alphabets (f.e. it should be possible to enter a Cyrillic captcha in Latin alphabet).

In T65216#653257, @Nikola_Smolenski wrote:

A possibility is to simply avoid words with diacritics, which should be possible for most languages.

That probably will help with Western European languages. Unfortunately it’d also exclude the vast majority of Vietnamese, leaving a word list too short to serve as an effective CAPTCHA word list.

Libraries like ICU have diacritic-folding facilities that should make it possible to accept o for ở and both o and oe for ö, as long as the source language is known (which it is in this case). Other scripts would be supported to some extent, though transliteration is a far messier problem than diacritic folding.

In T65216#2589692, @mxn wrote:

Libraries like ICU have diacritic-folding facilities that should make it possible to accept o for ở and both o and oe for ö

Yes please. (Can we update the task summary? That "only" makes me shudder.)