[yuecideng]

Description

Add a NeRD neural dynamics solver and differentiable Scene/Gym integration on top of the independent NeRD model/data/training workflow in #855.

Layer 2/2 · Base: codex/nerd-model-data-training · Depends on: #855 · Followed by: none. Review this PR against its base to see only the solver integration layer.

• Register nerd through dexsim.newton_solvers, bind ordered scene joints/contact slots and verify the exported timestep, gravity and physical signature before inference.
• Share one learned transition between ordinary and functional Scene stepping; retain independent histories, explicit reset/detach boundaries and first-order Torch/Warp kinematics gradients.
• Add Cartpole/Ant learned-dynamics tasks, APG deployments and matched live rollout evaluation against the corresponding Newton solver.
• Keep the SDK-free dynamics generate path from Add NeRD model extension and an SL training showcase #855 as the default. Scene-matched presets explicitly set collector: scene; their collector belongs to this integration layer and builds data matching deployment physics. Direct-Newton signatures cannot be relabeled as Scene signatures.
• Reuse Add NeRD model extension and an SL training showcase #855's reference asset interpretation, models, codecs, shard storage and training code. Those shared layers have no additional diff in this PR.

Dependency: #855 plus an extension-enabled DexSim build exposing named Newton solver providers and functional Scene APIs, as provided by DexSim MR !1483. The tested SDK is dexsim_engine==0.5.1; the version number alone does not prove extension availability.

Published datasets and models, under neural_dynamics/nerd/<robot>/v1, retain their recorded Scene domain. Current-model accuracy is measured against its matching analytical solver; no paper-accuracy reproduction claim is made.

Type of change

• New feature
• Refactor
• Documentation update

Validation

• After splitting, 115 focused solver/collection, CLI, task and functional-dynamics tests pass in the extension-enabled Newton 1.7 environment. This includes actual Scene physics, reconstruction of the recorded Ant friction domain, and the default SDK-free pipeline. GPU visibility is restricted to GPU 5; explicit GPU-marked tests in this selection are deselected.
• 42 project-context tests pass. Public API coverage is 3124/3124. Black 26.10.1 and whitespace checks pass.
• The independent layer has its own validation in Add NeRD model extension and an SL training showcase #855: 228 CPU tests, four GPU checks, original/published bundle compatibility and both real SDK-free CLI pipelines. These are not counted again here.
• Existing full-budget checkpoints, published artifacts, videos and experiment outputs are preserved. No training matrix or full-budget run was restarted.

Checklist

• Run black ..
• Review documentation/context and describe both collection boundaries.
• Document public APIs and pass the coverage gate.
• Retain focused behavior and integration coverage without duplicate offline-layer tests.
• Declare the lower PR and DexSim extension dependency explicitly.