Stability policy

Torch Lattice is still pre-1.0, but the current direction is intentionally narrow: CUDA training and export should match the MLX artifact contract rather than preserving every historical TorchSparse behavior.

Stable expectations

The following surfaces should be treated as long-term design anchors:

• coordinate rows use (batch, x, y, z);

• SparseTensor aligns coordinate rows with feature rows;

• sparse support behavior is explicit in module class names;

• artifacts are MLIR plus safetensors, not Python pickles;

• conformance compares CUDA output with MLX replay output.

Allowed breakage before 1.0

Breaking changes are acceptable when they remove ambiguous legacy behavior, reduce exporter boilerplate, or align Torch semantics with MLX replay semantics. When such a change affects migration from original TorchSparse, document the new mapping and update migration/conformance checks in the same change.