So MoarVM uses its own database of the UCD. One nice thing is this can probably be faster than calling to the ICU to look up information of each codepoint in a long string. Secondly it implements its own text data structures, so the nice features of the UCD to do that would be difficult to use.
In my opinion, it could make sense to use ICU for things like localized collation (sorting). It also could make sense to use ICU for unicode properties lookup for properties that don't have to do with grapheme segmentation or casing. This would be a lot of work but if something like this were implemented it would probably happen in the context of a larger rethinking of how we use unicode. Though everything is complicated by that we support lots of complicated regular expressions on different unicode properties. I guess first I'd start by benchmarking the speed of ICU and comparing to the current implementation.