I don't think HHH-5417 is accurate any more. I've created a new Jira issue to reflect the current behavior:
I've pushed the fix to 4.2 so it's ready for 6.3.1.
for now verified on EAP 6.3.1.CP.CR1, it's going to be reverted in CR2 because it indroduces new regression. For CP02 it is tracked under BZ1134738