It seems I lied about the %62 business. It can be restored to "b" as part
of URI normalization and [RFC3986] says un-%-encoding of codes corresponding
to unreserved US-ASCII characters can be done any time. But anything that
would become a reserved character (or an unrecognized character) in the
US-ASCII set in a URI should not be touched (and IRI encoding mapping and
inverting shouldn't touch it).