Variance-change detectors assume independent Gaussian noise. Feed one GARCH noise, where calm and wild stretches come in clusters with no real break, and it reports a change about half the time anyway. The changepoint-detector's meanvar model already warned when that was likely, but the warning caught only 16% to 20% of GARCH series.
The reason is small: it checked lag 1 of the squared residuals, and for this kind of noise the autocorrelation sits at lags 2 to 5. Adding a Ljung-Box test over lags 1 to 5 raises the catch rate to 34% (100 points) and 49% (200 points), at the price of 3 to 5 extra points of false warnings on clean Gaussian series. The numbers come from a seeded simulation and are reproducible with node eval/meanvar-robustness.mjs; no real series has been tried yet.