createKernelGPU builds a compute pipeline from the emitted WGSL and runs a pool's rows through it with an asynchronous readback; one device is shared by every kernel in the worker. (gpu kernel from [to]) is the same map as run, dispatched to the device: the results of a dispatch land in the pool on the next frame that calls it, one frame late by design, which the source says by using gpu rather than run. Without WebGPU, or while the device warms up, gpu is run. kidlisp/examples/kernels/starfield.lisp: four thousand stars on tilted orbits, one kernel a frame, swap the word on the DISPATCH line to compare. Worker bundle rebuilt.