fix(newton): support newton 1.5/1.6 (removed attributes, target layout, tiled camera)

[geng-haoran]

Summary

MetaSim's pyproject installs newton @ git+https://github.com/newton-physics/newton.git (currently 1.6.0.dev0), but NewtonHandler was written against newton < 1.2 and could no longer construct a scene: Model.joint_target_pos was removed in 1.5, the tiled-camera sensor API changed in 1.4, Model.num_worlds was renamed in 1.6, and — the one that bites silently — position targets moved to the joint_q layout. Four changes, all version-guarded so older newton keeps its current path:

1. Compat shims (_newton_compat.py): Model/Control joint_target_pos/joint_target_vel → joint_target_q/joint_target_qd (read/write properties over the 1.5 RemovedAttribute descriptors); Model.num_worlds → world_count.
2. Position-target layout: newton ≥ 1.5 lays joint_target_q out like joint_q (a free joint takes 7 slots, not 6) and publishes joint_target_q_start. The handler indexed position targets by dof (joint_qd_start) everywhere, so in any scene with a free-floating object every env after env 0 received its targets three joints off (env 1's panda_joint1 was driven to panda_joint4's target). All six write/read sites now use joint_target_q_start when the model has it and fall back to the dof layout otherwise; velocity targets and gains stay dof-indexed as before.
3. Tiled camera on newton ≥ 1.4: model-wide SensorTiledCamera(model, default_render_config=RenderConfig(...)), rays/outputs/light from sensor.utils, rendering through sensor.update after bvh_refit_shapes; textures that the sensor cannot load (RGB where RGBA is expected) fall back to untextured rendering with a warning instead of disabling cameras.
4. CameraState.depth is (N, H, W); the handler produced (N, H, W, 1) and CameraState.__post_init__ rejected it.

Verification (RTX 5090, newton 1.6.0.dev0, warp 1.17)

• pytest metasim/test/ -k newton: before — construction failed outright; with 1+3+4 only — 76 passed, 6 failed (test_set_dof_targets_per_env, test_parallel_envs_actions, test_parallel_envs_simulate_consistency × {2, 4} envs, all "env 1 didn't move towards target"); with 2 — 82 passed, 6 skipped, 6 xfailed, 0 failed.
• get_started matrix on --sim newton (the four scripts that declare newton): 4/4 pass (0_static_scene, 1_control_robot, 2_add_new_robot, 3_parallel_envs); rendered frames/videos in the session report.
• metasim/test/test_newton_compat_general.py (new, general): the shims against a synthetic newton module with the 1.5 descriptors and the pre-1.5 attributes; 2 passed.
• RoboVerse mjlab.velocity_flat_go1_v2 on Newton with 2 envs (for RoboVerse #817) runs on this branch.

Not verified

No newton < 1.5 install was available here, so the legacy branches are exercised only by the synthetic-module test; they are the unchanged code paths behind the version guards.

🤖 Generated with Claude Code

https://claude.ai/code/session_017i6VtKoovBNed815mWFqxw