A compaction is a marker, not a cut: tier full carries the whole transcript, summaries where they were made. Panoma compacts a conversation itself: the connected model reads it whole in windows, from the source's newest readable summary on, with the calls counted before the first is paid; every tier says its tokens, and a missing model is said beside the box. The list scans a big file once for its compaction.