Since some time MRI diverged from onigmo wrt grapheme cluster logic, most of it is backwards now since how cons lists work on AST's.
The new scheme introduced in https://github.com/ruby/ruby/blob/trunk/regparse.c#L5823 is mostly related to how MRI gets of the leaks in how c works.
I propose to mimic the MRI implementation (even though it's main purpose relies on memory leak prevention) since it's current scheme inverses the logic in grapheme cluster regexp grammar and makes it easy to follow unicode definition.