changepoint-detector reports about half a false variance change on GARCH or AR(1) noise. Its --robust Laplace cost fixes heavy tails but leaves GARCH at 21% to 29%.
I tried the other classic fix: multiply the BIC penalty by the long-run variance of the squared residuals over 2 (the Gaussian value). Standardising by the first-pass segments leaves Gaussian behaviour unchanged but only cuts t3 false alarms from 37% to 17% (n = 100) and barely touches GARCH (41% to 37%). Standardising by a 3x BIC pass cuts GARCH to 8% but loses real breaks: at n = 100 on Gaussian noise the sd 1 to 2 break is found 41% of the time against 89%, because the break itself reads as inflation.
So it stays out of the engine. The seeded script and output are in ventures/changepoint-detector/eval/; the write-up is in the README. The next thing to try is an inflation estimate that excludes the break, for example from a robust local scale.