feat(newton): skeleton NewtonSimulation(SimEngine) + config + factory registration

[cagataycali]

Summary

Add Newton GPU simulation backend skeleton — the first PR in a 7-PR series to land the Newton/Warp physics engine as a SimEngine backend.

This PR adds only the foundation: the class skeleton, config dataclass, factory registration, and 81 tests. No GPU dependencies are imported at module level.

Depends on: #84 (SimEngine ABC) — this PR branches from feat/simulation-foundation.

What's Included

New files (5)

File	LOC	Purpose
simulation/newton/__init__.py	47	Lazy loading, no warp import at module level
simulation/newton/config.py	105	NewtonConfig dataclass — 7 solvers, 5 render backends, 3 broad-phase options
simulation/newton/simulation.py	412	NewtonSimulation(SimEngine) skeleton — all 12 ABC methods + Newton extensions
tests/simulation/newton/test_*.py	490	4 test files, 81 tests

Modified files (2)

File	Delta	Change
simulation/factory.py	+16/-11	Register "newton" backend + aliases ("warp", "wp")
pyproject.toml	+9/-0	Add [newton] extras, warp.*/newton.* in mypy ignore

Key Design Decisions

1. Lazy imports: import strands_robots.simulation.newton does NOT trigger import warp or import newton. Heavy deps load only on first method call via _lazy_init().

2. Factory kwargs → config: create_simulation("newton", num_envs=4096, solver="xpbd") flows kwargs directly to NewtonConfig.

3. Fail-fast config validation: Invalid solver, render_backend, broad_phase, physics_dt, or num_envs raise ValueError immediately in __post_init__.

4. All ABC methods stubbed: Each raises NotImplementedError with a clear message indicating which subsequent PR will implement it.

5. Newton extensions as public methods: replicate(), run_diffsim(), solve_ik(), add_cloth(), etc. are NOT in the ABC — they're Newton-only capabilities.

6. send_action() + step(n_steps): The ABC separates action buffering from physics stepping. send_action() buffers in _pending_actions, step() applies.

7. reset_envs(env_ids): Newton extension beyond the ABC's reset() — enables selective per-env resets for RL training.

Tests (81 passed, 0 failures, 0.56s)

File	Tests	What it covers
test_config.py	30	Defaults, all 7 solvers, all render backends, all broad-phase options, invalid values, edge cases
test_factory.py	15	Registration, alias resolution (newton/warp/wp), create_simulation() round-trip, kwargs passthrough
test_lazy_import.py	4	Verify no warp/newton import at module level
test_simulation.py	32	Instantiation, context manager, cleanup idempotency, all 27 stub methods verified

✅ ruff check — 0 errors
✅ ruff format — all files formatted
✅ mypy — 0 issues in 3 source files
✅ pytest — 81 passed in 0.56s

7-PR Roadmap

1. ✅ This PR — skeleton, config, factory, 81 tests
2. World lifecycle + robot loading (create_world, destroy, reset, add_robot, procedural builders)
3. Step / action / observation (step, send_action, get_observation, get_state)
4. Object management + rendering (add/remove_object, render)
5. Multi-env + soft bodies (replicate, add_cloth/cable/particles)
6. Advanced features (run_diffsim, solve_ik, sensors, CUDA graphs)
7. Documentation + examples

Ref: cagataycali/strands-gtc-nvidia#314 | Roadmap: #94

---

🤖 AI agent response. Strands Agents. Feedback welcome!

[cagataycali]

This is the canonical Newton-stub PR. Closed duplicates: #102, #104.

Tracking issue: #96 (feat(newton): NVIDIA Warp/Newton simulation backend)