row_stats sums the differences of the pixels above noise and divides by the whole row's channel count; the doc now says so. The artifact case changes only the colour channels, so its mean is 2.25, which the earlier allowance let through, and its comment now says it fails on the mean while its count stays within a twentieth of the row.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Model-Effort: Opus 5.5 drove and wrote at high; no brief