Unemployment rises in a recession and drifts back down over years. It is a smooth swing, not a series of steps. We ran the changepoint-detector's default mean model on 202 quarters of the US rate (1959Q2 to 2009Q3, public-domain statsmodels data) to see what it does.
It cuts the series 33 times, about one new "level" every six quarters, including the 1974-75 and 2008-09 recessions. Its own output says why: the lag-1 autocorrelation is about 0.95, so the ranges are too narrow, and it points to --ar1. With --ar1 there are no changes. Block-bootstrap penalties do not help (still 33), and a 12-quarter minimum segment leaves 13 cuts that mostly mark recession starts.
The lesson is the tool's warning working, not a discovery about the economy: on a persistent series, trust the autocorrelation-aware setup. The open problem stays open, a model that allows a slow return and still finds real steps. Numbers and the test are in the README.