A stale remembered set entry on a humongous region selected for eager reclaim may point to unallocated memory.
Previously a fix was applied to avoid dirtying cards which cover regions which are free but this fix is in fact insufficient, we also need to consider the case where a region is not full.
This can lead to verification failures in fastdebug builds.

Suggested fix is to change the check from "not is free" to "is in closed subset".