pending is only the backfill QUEUE depth, so it reads zero while in-flight backfills are still writing records. waiting on it alone stopped each arm mid-flight: the first 10k-repo run had baseline at 7,456,623 records and interned at 4,511,412, a 40% gap, which makes the comparison meaningless because the arms indexed different repo sets. the drain now also requires the record count to have stopped growing, for 12 consecutive samples, and marks an arm INCOMPLETE if it hits the timeout rather than reporting a number that looks finished. the summary shouts when record counts differ instead of noting it in passing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>