Module split 

One graph as several kernels, for graphs too large for one shader (a traced contact step’s adjoint is hundreds of thousands of nodes, which shader compilers take minutes and tens of GB to optimize, or reject).

Structs

SplitKernels

WGSL kernels that together compute a graph’s outputs, run in order over the same environments: each reads inputs (n_inputs per environment) and a per-environment scratch (n_scratch floats), and the last writes outputs. Bindings: 0 inputs (read), 1 scratch (read_write), 2 outputs (read_write), 3 the uniform count; entry point main.