Changelog

Changelog#

Declare SciPy, PyTorch, tqdm, and Warp 1.10+ as core dependencies; add USD support to the simulation extra; preserve caller-managed Warp caches during import.
Apply shared validation to graph, loss, MLP, rendering, and training inputs across Torch and JAX.
Align histories, selected models, trajectories, and checkpoint refinement with post-update states; restore matching Torch optimizer state and initialize JAX Optax state from loaded weights.
Prefer Warp’s public JAX kernel API while retaining compatibility with Warp 1.10’s experimental FFI.

Fix Chamfer loss to use Euclidean directional means; validate squared and Chamfer input shapes; preserve explicit GeomLoss options and reject unknown keys.
Fix growth and division to respect target and capacity bounds; preserve inactive emulator rows and their identity gradients.
Fix categorical rendering point counts and polarity validation.

Document the purpose-built simulator and fixed-agent emulator, backend-specific contracts, model units, CUDA acceptance, and runtime decisions in ADR 0001.

Initial release of waxMorph.