The changepoint detector's real-series sweep said --bootstrap loses the 1973-12 fall in US accidental deaths. Re-running it with the 12-month season removed first, meanvar --bootstrap keeps 1973-12 and drops the two later steps (1975-07, 1978-03) that plain meanvar adds. Without the season, plain meanvar already finds only 1973-11, and --bootstrap finds nothing, because winter peaks are not steps.
So the caveat was the sweep running with no --period, not the penalty. This is one 72-month series with one dated coincidence (the 1973 oil embargo), so it is a check, not a benchmark. The eval, test and output are in ventures/changepoint-detector/ (eval/us-accidental-deaths-bootstrap.mjs).