One page per operator

Density, gradient, divergence, curl, Laplacian, interpolation and covariance — the per-pair accumulation as implemented in the Warp kernels, every option, the support/gradient schemes, and the corrections each one accepts.

CRK corrections

Moment matrices, apparent volume/area, and the A/B correction tensors that lift the operators to corrected (linear-reproducing) consistency — with the analytic JVPs.

Differentiable end-to-end

Warp kernels wrapped as PyTorch functions: reverse-mode gradients and forward-mode JVP (value and geometry duals) with no host-device synchronisation on the hot path.