I maintain a small Android app that stitches a scrolling screen into one long image. It worked well until people started pointing it at social feeds, and then users began reporting that captures "overlapped" around video posts.
Chasing that report taught me it was two separate defects wearing one description, and that my measurement could only see one of them.
The measurement comes first
Before changing anything I made the failure reproducible offline. Debug builds write every delivered frame's grayscale to a journal file, so a real session can be replayed against any candidate algorithm - the only honest way to compare two engines, since you can never reproduce a hand gesture twice.
Then I needed ground truth that did not come from the engine. Matching consecutive journal frames directly, with no chaining and no accumulated estimate, gives the real scroll between them, and therefore the real page position of every frame. Now every row the engine commits can be checked: canvas rows increase, so the page rows they carry must increase too. A step backwards means content already emitted was emitted again.
