Having multiple definitions of package annotations for a given package (via split packages) is equivalent to having the same class name used twice in a split package. The first version found in your class path is the one used by the VM.
I don't think we should use a different rule. And to me the same class define twice is of the same level of error as a package-info.java defined twice for the same package.
I think folk not being used to package annotations was a big part of Gerhard's objections, but what I personally find more problematic is the fact that multiple jars could conceivably include annotations for the same package. We might respond to that with a rule that a package annotation is only applied to the archive in which it is found, but then, AFAICT, an implementor would have to resort to bytecode analysis in order to sort out which setting applies to which jar. That could be more cleanly done by using a resource-based mechanism for package configs per containing jar, but then we're left with the IMO somewhat ugly situation of disparate mechanisms for package config vs. everything else.
Matt