Testing by paintingThe app learned by being painted with
Instead of testing features one at a time, the app was used to paint real pictures from a reference while I coached, the way I'd teach a student: commit to the stroke, find the shadow shapes, use the background to define the object, let it dry and sand it down.
Every sitting turned up something to fix. Wet-into-wet passes turned to mud, so the app now shows wet paint and suggests letting it dry. Colours came out too dark, which led to a colour fix that took mid-tones from 35% too dark to 7%. Blending was too weak, sanding too rigid, and the darkest paints couldn't get dark enough. Each one got fixed and painted again.
The biggest lesson was about judgement, not code. One painting that chased the reference's numbers region by region looked worse than an earlier, lighter version with one unified, warm tone. You judge a painting by squinting at the whole thing, not by the meters.