Transcript segments mapped to wrong version when user switches versions mid-utterance (decision made at confirmation time, not speech time)