Run the census on the 2026-09-14 snapshot
A fresh metric 1 rendered its best-ranked column without thousands separators: Spark hands the property row's untyped NULL over as an object column and the display formatting skips it. The cache helper now reads a computed result back from its CSV so a fresh run and a cached run share dtypes. The README no longer promises an hour for the flatten pass, which took five minutes this time.
Bug: T437880