It might not be testable with our current API, but how hard should we work to preserve non-ASCII strings when passing down to our underlying APIs?