My assumption about what that outcome tells us is that Weld is not ready to use at the point WildFly currently thinks it is ready to use (the point where it calls back into Hibernate through ExtendedBeanManager).

If you can get a solution working via afterDeploymentValidation I think that would be best because that would work with older Weld versions rather than just relying on newer versions if/when CDI/Weld ever adopt something like ExtendedBeanManager. It will be interesting to see whether that is possible though: specifically understanding the specific ExtendedBeanManager instances that a given afterDeploymentValidation call should trigger.