feat(isaacsim): expose material sharing and task-local G1 fixture

[TATP-233]

Summary

• Adds validated EnvCfg.isaacsim_share_friction_materials and forwards share_friction_materials only when the task explicitly opts in.
• Keeps the default false so existing per-geom/per-env reset-time friction DR is unchanged.
• Enables material sharing for the G1 IsaacSim tensor benchmark fixture and raises cold materialization timeout to 300 s.
• Replaces the inter-environment 100 m floor with a task-local 0.25 m finite box.

This is coordinated with unilabsim/unisim#320 and does not continue IsaacSim performance optimization.

Linked Work

• Issue: feat(isaacsim): expose material sharing and task-local G1 tensor fixture #1682
• Parent follow-up: chore: productionize GPU-centric tensor runtime (M11) #1673
• Historical IsaacSim scope: M9/IsaacSim: device-resident tensor lifecycle via CUDA IPC #1675
• Backend dependency: fix(isaacsim): harden selected reset and material sharing unilabsim/unisim#320
• Governing decisions: UniSim tensor lifecycle ADR and UniLab ADR-0002 backend capability boundary
• Roadmap declared base: develop/tensor-runtime
• Milestone: M10/M11 IsaacSim follow-up
• Base branch: develop/tensor-runtime

Validation

• make test-all passed on the final local head before this PR was created or updated
• Additional task-specific validation listed below

Commands actually run:

# 159b5b27 clean-worktree focused check
uv run pytest -q tests/base/test_isaacsim_backend.py tests/config/test_g1_motion_tracking_owners.py tests/scripts/test_g1_flashsac_backend.py
# 101 passed

# 804406b1 clean-worktree focused check
uv run pytest -q tests/config/test_g1_motion_tracking_owners.py
# 17 passed

# final head
make check
# ruff format/check, mypy, pyright, and focused test-import checks passed

make test
# 1686 passed, 5 skipped, 610 deselected, 6 failed

make test failures are confined to the pre-existing Drake/local-sibling dependency setup and do not touch IsaacSim files:

tests/base/backend/test_drake_batch_pool.py::test_drake_batch_pool_uses_thread_workspaces_not_env_workspaces
tests/base/backend/test_drake_batch_pool.py::test_drake_runtime_stewart_compact_state_matches_mujoco_qpos_order
tests/base/backend/test_drake_batch_pool.py::test_drake_runtime_stewart_ball_collides_with_top_plate
tests/base/backend/test_drake_batch_pool.py::test_create_backend_batch_mode_avoids_pydrake_and_steps
tests/base/test_backend_imports.py::test_unisim_dependency_uses_an_approved_source
tests/envs/test_stewart.py::test_stewart_drake_materializes_or_fails_at_optional_runtime_boundary

A clean worktree of the unchanged base also fails test_unisim_dependency_uses_an_approved_source because this local checkout resolves unisim-core to file:///home/user/ws/unilabsim/unisim instead of the approved Git URL. make test-all remains a merge gate and must be rerun in the integration environment after the sibling checkout/version is synchronized.

Remote CI route:

• Base is not main: not scheduled; local make test-all is the test gate (currently blocked by the unrelated local environment failures above).

Impact

• Backend impact: isaacsim
• Platform impact: Linux x86_64 / NVIDIA CUDA opt-in tensor fixture; default behavior is platform-neutral
• Training effect expected: no default change; the G1 diagnostic fixture opts into shared materials and a task-local floor

Artifacts

• Benchmark diagnostic from the coordinated backend source:
  • m10-isaacsim-box025-shared-materials-angular-correct-direct-reset-g1-2048-5090-gpu7-20260928.json
  • SHA256 b986065843ded85571fa3c4f88d38e0aa541500ae6419b2bd3681533cd2d9971
  • 16,713.83 env control steps/s, 50,141.50 physics substeps/s, 122.53 ms mean iteration
  • single-run diagnostic, not an end-to-end throughput-improvement claim

Checklist

• Added or updated tests where needed
• Updated docs if behavior or workflow changed (backend contract is documented in UniSim PR fix: Flashsac for g1_walk_flat #320; no UniLab user docs are changed)
• Linked the driving issue
• Noted any follow-up work explicitly

[TATP-233]

Fixture provenance note: the committed explanatory-comment XML has SHA-256 fe81401f47d198fe3f8681e1aa3a6726a3bf0481fb75ecf00b70ad9af759ccb6. The recorded benchmark used the semantic 0.25 m box fixture without this comment (SHA-256 07c60d2b94b98f01fcad99929ba65c86224a357e4470d9bb5f2769b46b121888); geometry, pose, mass, timestep, and integrator are unchanged. The hash difference is documentation-only.