Unify ONNX deployment dependencies under policy-deploy

[acrlw]

Description

Replace the NMG-specific nmg extra, which installs CPU ONNX Runtime, with
policy-deploy for both NeuralPlanner and ONNX-backed DexSim Motion Policy Kit
consumers. pip install "embodichain[policy-deploy]" installs
onnxruntime-gpu>=1.20,<1.27; base dependencies remain unchanged.

The upper bound keeps the documented CUDA 12 stack compatible: ORT 1.27+
defaults to CUDA 13 according to the
official compatibility table.
Update NeuralPlanner's missing-runtime hint, installation instructions, and
motion-planning context to use the shared extra.

Migration: replace [nmg] with [policy-deploy] in install commands. Existing
CPU ORT installations need removal before installing the GPU package; the
installation guide includes cleanup instructions for environments with both
packages. NMG functionality and execution-provider selection are preserved.

Related: #520.

Type of change

• Breaking change (existing [nmg] installation commands require migration)

Validation

• Built sdist and wheel with uv build --no-sources; both pass twine check.
• Checked both artifacts' metadata: base requirements unchanged, nmg absent,
  and only policy-deploy selects onnxruntime-gpu>=1.20,<1.27.
• Installed the wheel with --no-deps in an isolated environment and verified
  installed extra metadata and the packaged NeuralPlanner installation hint.
• Black 26.3.1 repository check passes (1,032 files); API docs coverage passes
  (2,067/2,067 exports); context checker and 23 context tests pass.
• Changed MyST pages parse successfully and their local Markdown links resolve.

Validation covers packaging and documentation. Full simulation/GPU inference
tests and a full Sphinx build were not run. Black emitted a Python-version
advisory, and pytest emitted a cache-directory warning; both checks exited 0.

Screenshots

Not applicable.

Checklist

• I have run the black . command to format the code base.
• I have made corresponding changes to the documentation.
• Public API documentation coverage check passes; no public API signatures changed.
• I have added tests that prove my fix is effective or that my feature works. No new runtime behavior; built-artifact checks and existing context tests were used.
• Dependencies have been updated.

[greptile-apps]

Confidence Score: 4/5

The PR appears safe to merge, with a non-blocking gap in automated coverage for its new packaging contract.

Findings

1.  Packaging Contract Lacks Coverage ▶

Fix with agent prompt

### Issue 1
pyproject.toml:54-58
The new public installation contract is only manually validated. Existing metadata tests iterate through extras but do not assert that `policy-deploy` exists, `nmg` is absent, or the CUDA-compatible `<1.27` bound remains. A focused metadata test would prevent future packaging changes from silently breaking the documented migration or CUDA 12 compatibility.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

• Replaces CPU onnxruntime with bounded onnxruntime-gpu>=1.20,<1.27.
• Updates NeuralPlanner’s missing-runtime message and user-facing installation documentation.
• Documents CUDA compatibility, package cleanup, and migration instructions.
• Leaves the base dependency set and planner execution-provider selection unchanged.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  Extra[policy-deploy extra] --> ORT[onnxruntime-gpu 1.20 to 1.26]
  ORT --> Neural[NeuralPlanner / NMG]
  ORT --> MotionKit[DexSim Motion Policy Kit]
  Providers[Configured execution providers] --> Neural
  Providers --> MotionKit

Loading

_{Reviews (1) · Last reviewed commit: "build(deps)!: replace nmg extra with pol..."}

[greptile-apps]

Packaging Contract Lacks Coverage

The new public installation contract is only manually validated. Existing metadata tests iterate through extras but do not assert that policy-deploy exists, nmg is absent, or the CUDA-compatible <1.27 bound remains. A focused metadata test would prevent future packaging changes from silently breaking the documented migration or CUDA 12 compatibility.

Prompt To Fix With AI

This is a comment left during a code review.
Path: pyproject.toml
Line: 54-58

Comment:
**Packaging Contract Lacks Coverage**

The new public installation contract is only manually validated. Existing metadata tests iterate through extras but do not assert that `policy-deploy` exists, `nmg` is absent, or the CUDA-compatible `<1.27` bound remains. A focused metadata test would prevent future packaging changes from silently breaking the documented migration or CUDA 12 compatibility.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!