Jeff Pohlmeyer did some marvelous debugging to track this one down. We MUST

NOT free the existing hash entry when we try to add a new one that matches
an existing entry. We now instead free the new one, and make the parent
function use the old entry's struct instead.