HamiltonZero v0.1.0

Initial research release with learned-router training, compiled fine-tuning, compiled evaluation, public datasets, standalone compiled inference, scientific observables, and the HamiltonZero v1 foundation checkpoint.

Updated 14 August 2026:

• restored production explicit multi-device shard-map, JIT, placement, and donation boundaries across learned-router training, compiled fine-tuning, and contest/singular evaluation;
• added model-only checkpoint warm starts for learned-router training;
• isolated staged optimizer caches per optimizer instance.