waxmorph.torch.warp_autograd

Torch autograd adapters for Warp mechanics and diffusion.

Each forward call owns a fresh warp.Tape and a frozen neighbor list. Backward seeds the output adjoint and replays that tape. The gradient contract covers the documented Torch state. Torch uses the input’s Warp-compatible device. JAX uses custom VJPs on CUDA. Each bridge provides its documented gradient contract.

Classes

WarpDiffusionStep(*args, **kwargs)	Graph diffusion step with concentration gradients.
WarpMechStep(*args, **kwargs)	Sticky-sphere Euler step with position gradients.