Learned alias, then the literal key, then the keys and script ids whose own words contain
every word said. No similarity scoring anywhere: a near-match fails silently and confidently,
and here it would write to a key nobody named. Several matches is a question, not a ranking.