Screenshot testing in Python: why I ended up writing my own library

Yes, with one limitation. A failure message has an "also" line: what was set aside, grouped by reason, plus the first three regions with their position and the rule that dismissed them. The HTML report lists up to 50 of them under the reported changes, each with its size, coordinates and what the engine measured, e.g. 40x12 at (216, 473): antialiasing — the baseline as it is reproduces 99%. Passing checks get the same list, so a green run doesn't hide anything silently either.

The limitation: the diff image only outlines the reported changes, so for now you match a dismissed region to the picture by its coordinates. Marking them on the image too, in a muted style, is a fair ask — I'll look into it.