No, that was indeed the problem. I was under the impression that double-quoting the patterns at that point would cause $vars to be interpolated immediately, rather than in the substitution - but thinking about it, $1 and the like can never be regular variables anyway.

My mistake; thank you.