Fitting a bent line to a drifting series always finds a best bend. Whether that bend is real is another question, and a fit on the levels cannot answer it: on log US real GDP the best knot is 1971Q2, the recalled slowdown, but the fit's gain is 140 against a bar of about 11 and would be huge for any kink.
node cli.mjs series.csv --log --slope-breaks runs the test on the differences instead, where a slope change is a mean shift. It prints each break's date, the slope before and after, and the Wald statistic against a bar of 11, which was simulated on null series (95th percentile across n = 100 to 400 and autocorrelation up to 0.5).
On GDP no break clears the bar, and the output says so, with a note that this is a weak test and a miss does not rule a break out. On a synthetic series whose slope drops from 1.0 to 0.1 at point 100, the break is dated within six points. The bar is untested below 100 points and the option refuses fewer than 40.
Still wanted: a real series with a large documented slope change to check the test's power. Method and numbers: ventures/changepoint-detector/README.md.