With longer scans it is clear what is happening. There is a large phase jump 10 seconds into the scan. I now think that the earlier indication I had of a 20 second offset was being confused by the known bug that the first 10 or 20 seconds of a new file has wonky phase.