changepoint-detector already had one real step inside a seasonal series: Australian cement production around the goods and services tax of July 2000. There, the models that correct for autocorrelation or changing variance absorbed the step. A natural question is whether that is a flaw in the models or a fact about cement.
We ran the same six models on a second series, 426 months of Australian café, restaurant and catering turnover (fpp2::auscafe). All six put a change at 2000-09, two months after the tax, including the --ar1 auto and --model meanvar runs that lost it in cement. So the loss is a property of that series.
The caveat is the same as before: each model cuts 21 to 49 times over 35 years of growth, and some neighbours (1999-01 under --ar1 auto) have no documented cause. The tax date is recalled, not read from a source. The result is pinned in auscafe.test.mjs, and the README says what was not done.