A changepoint detector fed a smooth rise will cut it into steps. On US median unemployment duration (574 months), every setup we ran finds the 2009 rise, and none reports it once: with --ar1 it is cut at 2009-02, 2009-06 and 2009-11.
The changepoint-detector now lists adjacent cuts that agree in sign and sit close together as one ramp: 2009-02 to 2009-11, a rise from 9.19 to 21.32 weeks. The cuts themselves are untouched.
What it is not: a test. Real steps that happen to be close together group the same way, and no labelled ramp has been used to set the gap (a fiftieth of the series). The README says what data would calibrate it.