Right. On closer look, it seems like there’s some rather fundamental assumptions in the code that locates the position closest to a set of coordinates that don’t hold up in bidirectional text. This will take some more time to fix. Putting it on my to-do list.
I have a fix for this in the block-wrapper branch (see this thread). It will be released properly when that code has had a bit more time to prove itself, but is already available as @codemirror/view 5.39.0-beta.3