Fix #5: Derive SO-ARM units and safety bounds from use_degrees

[theonlypal]

Description

This PR fixes #5 by supporting LeRobot's normalized motor mode end to end instead of rejecting use_degrees=False.

Previously, the package always declared joint_pos as deg+normalized and used degree-mode joint bounds of ±180. Under LeRobot's normalized mode, arm joints are limited to ±100 and the gripper remains 0–100, so the declared state units were incorrect and the default clamp no longer provided an effective backstop.

Changes

• Replace the fixed module-level STATE_SPEC with a state specification derived from use_degrees.
• Derive automatic safety bounds from the selected mode: ±180 for degree-mode joints or ±100 for normalized joints, with the gripper at 0–100 in both modes.
• Preserve explicitly configured calibrated limits in either mode.
• Add use_degrees to LeRobotPolicyConfig and thread the setting through both policy and embodiment observation contracts.
• Keep homing guidance unit-correct by describing max_relative_target in native motor units.
• Document the matching policy/embodiment configuration requirement and normalized-mode safety behavior.
• Add regression coverage for configuration, packing, policy metadata, embodiment clamping, compatibility, and complete evaluation flow.

Verification

• Ruff lint and formatting pass.
• Strict mypy passes.
• 92 tests pass on Python 3.11 and Python 3.12.
• Line and branch coverage remain at 100%.
• Pre-commit and pre-push hooks pass.
• uv lock --check passes.
• Runtime-only import hygiene passes without torch or LeRobot installed.
• LeRobot 0.5.1 accepts the propagated use_degrees=False driver configuration.
• The full suite passes under a fresh compatible dependency resolution.
• Wheel and source-distribution builds pass.

Fixes #5

[jeqcho]

@theonlypal happy to take a look again once the tests are fixed!

[theonlypal]

@jeqcho Hi Jay, the implementation is green. The macOS 3.12 job was rate-limited during setup-uv, before dependencies or pytest ran; ci-ok only mirrors that setup failure. Could you rerun the failed jobs?

[jeqcho]

Merging — thanks @theonlypal, this was a pleasure to review.

Two things stood out. The generator-expression trick in test_soarm_normalized_mode_preserves_explicit_degree_sized_limits, keeping the ±180 tuples from aliasing the sentinel, is exactly the care identity-based defaults demand — constant folding would have silently broken a naive version of that test. And the derived ±100 defaults with per-side swapping restore a real safety backstop: the old ±180 clamp simply never fires in normalized space, so this closes an actual gap rather than just widening an input.

One tiny thing I won't hold the merge for: the module docstring in policy.py (around line 21) still says chunks are "(degrees for the joints, 0..100 for the gripper)", which is stale now that normalized mode exists. If you feel like a one-line follow-up it's yours; otherwise we'll sweep it up.

And an apology: the "failing test" I pointed you at earlier turned out to be our CI runner getting rate-limited, not your code. Nothing was ever wrong with the tests. Hope to see more from you.