Tuesday is the only day the model reads the carrier feed before the overnight reconciliation finishes, so a slice of Monday's late arrivals is counted twice. Start with the ingest window in loaders/carrier_feed.py, then compare the reconciled totals for the last four Tuesdays against the raw feed.

The patch moves the ingest cutoff to 04:00 so it always lands after reconciliation, drops the duplicate rows that the old window let through, and adds a regression test that replays four weeks of Tuesday feeds. Nothing outside the loader changes, and last week's totals fall back in line to within 0.3%.

It does. The fixture parametrises the reconciliation finish time, so a run that lands at 05:10 fails on row counts instead of passing quietly. I also kept the old window behind a flag for one release, which lets you diff both outputs on staging before the cutover.