This is not exactly the same as the previous version, now we create a new instance each time, even if the entry existed in the map.
For getMapper and getAccessor it doesn't matter that much, because they're only called directly by the client (and in general they will probably only do it once), but getUDTCodec is also called internally, each time a new class with an UDT field is parsed.
Checking the map first is simple enough, so I think we should do it in all cases:

Mapper<T> old = (Mapper<T>) mappers.get(klass);
if (old != null) {
    return old;
}
EntityMapper<T> entityMapper = AnnotationParser.parseEntity(klass, this);
mapper = new Mapper<T>(this, klass, entityMapper);
old = (Mapper<T>) mappers.putIfAbsent(klass, mapper);
if (old != null) {
    mapper = old;
}
return mapper;

It can still create an unnecessary instance, but the window is much shorter.

🐼 my bad, I was looking at the diff view on github and I only focused on the changed lines.