Somebody told me the front page felt like it snagged where the point cloud turns into the photograph. Not a number, not a repro — a feeling. Those are the reports worth taking most seriously and the easiest to answer badly, because there is always a plausible cause lying around and no way to tell whether fixing it fixed anything.

So before changing a line I drove the page with real wheel events and recorded, every frame, how long the frame took and how far the hand-off had got. The result was not subtle. The walk ran at about 19 ms a frame. The hand-off ran 98 ms on average and 367 at worst, and the fade itself went from nothing to resolved inside ninety-eight pixels of a four-and-a-half-thousand-pixel runway. Three frames. A crossfade that completes in three frames is not a crossfade, it is a cut, and it was landing on the most expensive frames the page draws.

The expense had a clean cause. As the cloud converges onto the photograph every point grows to 4.6 times its size — about twenty-one times the area to shade — and the thinning that paid for that was keyed to the photograph fading in, which does not start until the growing has already finished. There was a window with every point at full size and nothing covering it. Thinning on the right clock took the worst frame to 117 ms.

Then a second report arrived and it was the same bug in a different hat: a black layer that appears when the photograph does. The film is pinned for the length of the walk and then released, and the walk ends exactly where the pinning ends — so the photograph reached full opacity about twenty pixels before the page let go of it, and the band below slid up and ate the thing six screens of walking exists to reach.

One fact, three symptoms. The hand-off had no room: expensive because the cost had nowhere to spread, abrupt because the fade had nowhere to play, guillotined because the arrival had nowhere to land.

What I keep taking from this one is how badly the first explanation would have aged. The obvious suspect was the image — a full-bleed 2160-pixel photograph fading up, surely a decode hitch. Good theory, ten minutes to implement, and the numbers say it was not the problem. The measuring was not the slow part. The measuring is what stopped me spending an afternoon on the wrong thing.