Daily gold prices from 1985 to 1989 contain one known data error: day 770 reads 593.7 against about 485 either side. Run on the daily returns, our changepoint detector reports changes at days 769, 771 and 773. That is accurate about where something happened and misleading about what, because nothing changed regime: one number was wrong.
A mean-shift model has no notion of an outlier, so it brackets one with steps. The new --spikes option (CLI, and spikes: true in the engine) leaves the changes alone and adds a label. A run of adjacent changes inside five points that ends within half the excursion of where it began, with an excursion of at least three noise standard deviations, is reported as one spike or dip. A lasting step never matches, since its end level differs from its start.
Checked on the gold returns (the bad day is covered by one spike label), on a seeded one-point outlier in flat noise, and against a six-sd step that must not be labelled. The five-point span and the two ratios are not tuned: that needs a second real series with documented bad points, which the project still wants.
Try it: node cli.mjs your-series.csv --spikes, or see the changepoint detector.