Runtime correction
- Default prequantized Diffusers artifacts to BF16 compute when dtype is omitted, enabling native packed W4 dispatch.
- Preserve every explicitly requested compute dtype.
- Leave Triton version selection to the installed PyTorch build; the kernels extra no longer pins a second Triton.
- Pin and select CMake 4.4.2 in Windows ABI3 CI to avoid the CMake 4.4.0 SABIModule detection regression.
Validated on MiniMax H3 with 300/300 eligible transformer linears using the packed W4 path. PyPI: https://pypi.org/project/orbitquant/0.9.1/