It's not clear to me that errno is defined to be valid except
immediately after an error return.  Can you point to the spec that says
the behavior you don't like is wrong, if only to educate me and others?

That said, I see the point of not modifying errno across non-error
calls, or maybe just forcing zero after them.

Perhaps pthread_create makes internal system calls, and one fails, and
then it copes, but leaves errno.