changepoint-detector warns when its drifting-season mode finds three or more steps within two cycles: that pattern often means the seasonal shape is changing, not that the regime changed three times. The threshold was a guess checked on one real series (UK gas).
We simulated 200 seeded series per case, quarterly (n=120) and monthly (n=144). With a fixed season and no change the note appeared on 2% and 3.5%. With two real, well-separated level steps it appeared on 2% of quarterly series but 11.5% of monthly ones. With three real steps packed into two cycles it fired on 95% to 100%, which is right only if they are one drifting shape; the rule cannot tell the two apart. A smoothly tripling amplitude produced almost no steps at all, so the note seldom fires for that kind of drift.
Synthetic data, one profile shape, no autocorrelation. The threshold is unchanged; a tighter monthly rule is the next thing to try. Script and output: eval/drift-warning.mjs in the venture.