[huster-wgm]

The new DeterminismType.ALL option can be mistaken for a guarantee of bitwise-identical trajectories. However, _sort_contacts and _sort_constraints are still TODO/pass hooks on main, as noted in the future work of #1687. The README now makes that limitation explicit and distinguishes model-level controls from Warp's unsupported global deterministic mode, referring readers to #562 and #1635.

This is a documentation clarification, not a physics fix. In a downstream G1 GPU replay on MJWarp/MuJoCo 4.0.0 and Warp 1.19.0, model-level ALL still diverged between native runs; enabling global RUN_TO_RUN raised WarpDeterminismError in the convex CCD kernel. Those observations motivated checking the implementation, but are not presented as an isolated engine bug reproduction.

Validation: checked the statements against current main (4d1c64b), #1687 and the maintainer response in #1635; git diff --check passes. No runtime code changes. Full pytest/pre-commit checks were attempted but could not start in the isolated checkout because the test tools were not installed; keeping this draft pending CI/review.

[huster-wgm]

Closing this documentation-only proposal in favor of issue #1925, which records the model-level ordering gaps, evidence, validation limits and acceptance criteria. Implementation is split between #1923 (contacts) and #1924 (constraints/address remapping); both are needed. The unsupported global Warp RUN_TO_RUN path remains a separate concern, as discussed in #1635.