This allows us to be more selective about the streams we derive durations from
(specifically, ignoring text streams with outlier end times) in the common case