Sounds good!

Yeah so I have no sophisticated tools, just guessed calculations done mostly at run time (add up how much do gpu compressed textures should take for a model, mesh+blendshape memory estimates, other data structures). And obv for audio and animations just ensuring what’s not needed is never in memory and just knowing how much the biggest ones cost. To come up with an estimate of an upper bound and obv with lots of testing. ios safari doesn’t even publicly expose how much memory usage is allowed to a tab (and it changes) so really the final test is that stuff stops crashing. But my testing put the crash window between 500mb-1gb on an iphone pro. over 1gb usually guarantees a crash. Not really enough to make a proper game by today’s standards but hey they shipped gta san andreas at 32mb ram on ps2 so I’m sure I can push things further.

To answer your webxr webgpu question, no, it still doesn’t mostly work. If you’re interested in reading up on it:

https://github.com/immersive-web/WebXR-WebGPU-Binding/blob/m...

Post by a Chrome dev: https://toji.dev/2025/03/03/experimenting-with-webgpu-in-web...

Webxr webgpu attempts iirc some advanced reprojection for increasing FPS that takes into account the depth buffer. And vision pro at least today has some nausea inducing bugs around that for at least the official immersive web examples hosted on github.io. so its better if a depth texture isn’t supplied today. the fallback reprojection works better more smoothly.

For the story stuff, I have similar things setup in my project to essentially reload the game at a certain point in the story but still I personally find HMR to be much faster wherever possible.