A bound that only held while the grid was coarse
The background
Every coloured cell the map draws is one stored value fetched by a real query — no smoothing, no interpolation. That rule is the point of the page, and it made the map a chessboard: 480 cells twenty-three kilometres across, with the sixty-kilometre eddy the world is built around covering three. Drawing it smoothly is the one thing the panel may not do, so the field had to get denser at source.
The requirement
Four times the values along each horizontal axis: 7,680 cells where there were 480, without the page becoming something you wait for.
The options considered
Sixteen times the values is sixteen times the work: a now-cast pass goes from 18 ms to 274 ms. Invisible at the rate the map is watched; not wound forward, where the fast-clock buttons now deliver about half of what they did.
Then the acceptance check went red, which was the useful part. It recovers the eddy's centre from the stored bytes, and allowed two grid cells of error: 47 km on the coarse grid against 13 km measured, 11 km on the fine grid against 13.9 km. Measured at five densities the error converges on 14 km, so it was never limited by the grid — the world composes four features additively, and the centroid is pulled by the front and the thermocline. A floor under the achievable error had been read as a ceiling over this estimator's, true only while the grid was coarser than the bias. The check now claims what it can — the recovered centre falls inside the eddy those bytes carry — and goes red when the eddy is taken out of the world.
The demo
The eddy is round.