The first number in the write-up is one you can check without a device. Raxxo Studios puts the first-pass offscreen render at roughly two and a half minutes per frame [3], and describes that as a little over two hours of compute for one second of 30 fps footage [7]. 150 seconds times 30 frames is 4,500 seconds, or 75 minutes [1]. The pipeline was unusable at either figure. A number quoted in prose does not tell you what to budget a render farm against.
The cause was the launch line. The tool was starting Chrome with `--use-gl=angle --use-angle=swiftshader --enable-unsafe-swiftshader` [4]. SwiftShader rasterizes on the CPU, which is the correct default for a CI box with no GPU and the wrong one for a workstation that has a GPU [5]. Switching to `--use-angle=metal --enable-gpu` gave 38 to 60 fps on the same code, machine and browser [6]. That is a frame time of 17 to 26 milliseconds against 150 seconds, so roughly 5,700x to 9,000x [2]. A software renderer does not throw an error; it simply takes longer, and that is how you lose days optimizing geometry that was never the problem [8].
For the 38 to 60 fps to transfer, your capture host needs a GPU that ANGLE's Metal backend can address. On a Linux CI runner you get SwiftShader, and 2.5 minutes per frame is the figure that applies to you. Both configurations still live in the author's tools directory on purpose, because determinism and speed are different jobs [9]. Keeping the slow one and demoting it from default is better engineering than deleting it.
The WKWebView failure is the one worth copying. iOS kills the web view's GPU process when it decides memory is tight, and the app does not crash: the game loop keeps ticking and the score keeps counting while the world renders black [10]. Observed on device, no `webglcontextlost` event was dispatched at all [11]. So the shipped build stops waiting to be told. `isContextLost()` is polled about every two seconds [12], and a draw-call floor treats 120 consecutive frames under ten draws, well past boot, as a dead pipeline even when the context object still reports itself healthy [13]. At 60 fps that window is two seconds of wall clock, four at 30 [3], so the floor trips inside roughly the same window as the poll and does not depend on the context object telling the truth.
The ten is workload-specific. It holds because this game's post-processing chain alone submits more than ten draws per composed frame [13]. Port the detector to something that composites with one fullscreen quad and it either never trips or trips at the title screen.
Recovery banks the run so the score survives, then reloads [14]. Reload is also the most expensive thing the app does, replaying parse, scene build and shader warm-up on a device that just ran out of memory, which can trigger the same kill; the author's term for the unbounded version is "a hang with extra steps" [15]. Three recoveries inside ten minutes ends the cycle [16].
The world-fold bug is arithmetic. Folding every 3,000 m over a city that repeats every 540 m leaves a 300 m offset on every fold [20]. 540 times 5 is 2,700 and 540 times 6 is 3,240 [4], so the fold distance was never free to be a round number.
