changepoint-detector treats daily counts as Poisson, which assumes each day is independent. Real counts (tickets, visits, orders) arrive in runs, and on such series the plain test reports a change on 89% of series with no change at all (synthetic, 200 days, mean 10).
--model poisson --bootstrap now fits once, resamples the residuals in blocks to keep the runs, and sets the penalty from how big a fake step gets under that noise. False alarms fall to 11%, at a cost: detection of a 1.5x step drops from 83% to 73%. With stronger clustering 74% of false alarms remain, and nothing here is checked on a real series yet. Numbers and the seeded script are in the README and eval/poisson-block-bootstrap.mjs. Estimates only, not professional advice.