feat(examples): migrate LIBERO Isaac examples + shared harness from robots-sim

[yinsong1986]

What

Migrates the Isaac-specific LIBERO examples and their shared harness from
strands-labs/robots-sim/examples/libero/ into this repo's
examples/libero/, now that the Isaac backend is vendored in-tree (epic #1144).

New files under examples/libero/:

File	Role
run_isaac.py	LIBERO eval on the Isaac backend via create_simulation("isaac"); two grep-stable output lines + synchronous rollout-MP4 recorder
run_isaac_agent.py	Same eval driven by a Strands Agent (single @tool wrapper)
libero_backend_matrix.py	Subprocess-and-parse side-by-side backend-matrix table
gr00t_server_deterministic_wrapper.py	Container-side strict-determinism wrapper for run_gr00t_server (imports gr00t.*/torch/tyro; runs inside the GR00T container, not on the host)
README.md	Deps (via extras), run commands, env vars

Why

Follow-up to backend-absorb epic #1144 / migration epic #1269. With the
Isaac backend in-tree, create_simulation("isaac", ...) resolves to
strands_robots.simulation.isaac.IsaacSimulation, so these examples no
longer need the external strands-robots-sim package.

Key changes vs the robots-sim originals

• Rewrote imports: from strands_robots_sim.isaac import IsaacSimulation
  -> from strands_robots.simulation.isaac import IsaacSimulation in both
  run files. Updated stale strands_robots_sim/... module-path references
  in docstrings/comments and the strands-robots-sim[isaac] install hints
  to strands-robots[sim-isaac,benchmark-libero].
• Repo conventions (epic [EPIC] Migrate robots-sim examples into this repo #1269): multi-file example in a named subdir
  with its own README.md; deps documented via extras (no per-dir
  requirements.txt); indexed the three host-side scripts in
  examples/README.md; documented the new STRANDS_ISAAC_RTX_PATHTRACING
  env var.
• Lint: applied this repo's ruff rules (UP037 quoted-annotation removal
  • format) to the migrated files.

Test surface

• create_simulation("isaac", headless=True, num_envs=1, render_mode="rtx_realtime")
  constructs the in-tree IsaacSimulation and destroys cleanly on this
  non-Isaac (no-GPU) host.
• IsaacSimulation.is_available() short-circuits with the install-hint
  reason string (no omni.*/isaacsim.* import needed) — safe to
  import/lint on CPU-only CI.
• run_isaac._suite_for_task doctests pass.
• hatch run lint: clean (Success: no issues found in 755 source files).
• hatch run test: 8077 passed, 45 skipped. The only failures are the
  pre-existing OpenGL/rendering environmental ones
  (test_sim_camera_mesh_publish, test_lerobot_e2e::test_record_train_load_loop,
  test_rendering.py/test_policy_runner.py) — confirmed identical on clean
  upstream/main (mjr_makeContext / "an OpenGL platform library has not
  been loaded"). This PR touches only examples/, so it cannot affect
  library tests.

Acceptance criteria (from #1273)

• Migrate libero/run_isaac.py, run_isaac_agent.py as runnable (no skip marker — feat: dataset format bridge — LeRobot recording to GR00T training format #140 RTX fix is upstream)
• Migrate shared deps libero/libero_backend_matrix.py, gr00t_server_deterministic_wrapper.py
• Do NOT migrate run_mujoco.py/run_mujoco_agent.py
• Update strands_robots_sim imports -> strands_robots
• Confirm create_simulation("isaac") path works now that the backend is in-tree
• Place scripts under examples/libero/, no NN_*.py renumbering
• No per-dir requirements.txt; deps documented via pip install "strands-robots[sim-isaac,...]"
• Add rows to examples/README.md

Out-of-scope follow-ups

• run_mujoco.py/run_mujoco_agent.py migration + run_isaac_fleet.py (R23)
  are separate epic-[EPIC] Migrate robots-sim examples into this repo #1269 children. Until they land, libero_backend_matrix.py
  prints unavailable (no <driver>.py) for the mujoco/isaac-4096 rows
  (graceful degradation by design) and evaluates the isaac-1 row.
• Update robots-sim's stale examples/MIGRATION.md (tracked in [EPIC] Migrate robots-sim examples into this repo #1269).

Parent: #1269. Board: this issue should move to In review on
Strands Labs - Robots.

[cagataycali]

Review — LIBERO Isaac examples + shared harness

Reviewed against a local checkout. Independent of the other examples; lands last in the roll-up ordering.

Verified

• No strands_robots_sim refs, no per-example dep files, all three scripts (run_isaac.py, run_isaac_agent.py, libero_backend_matrix.py) indexed in examples/README.md, no host paths, no emojis in runtime strings.

CodeQL triage

• libero_backend_matrix.py:129 — Unused global _RE_BENCHMARK: real dead code. The comment says it guards against print-format drift, but _RE_RESULT is the one actually parsed. Either wire _RE_BENCHMARK into the parse (so a missing benchmark_name= line is detected) or drop it. One-liner.
• run_isaac.py:888 and run_isaac_agent.py:871 — Except handles BaseException: defensible, keep as-is. Both are top-level if __name__ == "__main__" crash handlers that re-raise SystemExit first and os._exit(1) to bypass Isaac Sim's atexit hook (which otherwise swallows non-zero status into a misleading exit 0). They carry # noqa: BLE001 and an explanatory comment — this is the correct pattern for the SimulationApp fast-shutdown problem.
• run_isaac_agent.py:849 — Empty except: this is except Exception: pass around _sim.destroy() in a finally teardown. Recommend narrowing to the concrete destroy error or adding a one-line logger.debug(...) + comment so a real teardown failure isn't silently swallowed (project convention: clean up on partial failure, but don't bury the reason).

No blockers. The shared-harness + backend-matrix approach (one LIBERO task across every installed backend with a side-by-side success_rate / wall_time table) is a genuinely useful addition. Thanks Yin.

[yinsong1986]

Thanks for the CodeQL triage @cagataycali. Addressed the two actionable items in db4480f:

• libero_backend_matrix.py:129 (unused _RE_BENCHMARK) — wired it into _parse_driver_output. The parser now checks the benchmark_name= announce line: if a driver exits 0 with neither the metrics line nor the announce line, that's genuine output-format drift and the row status reads ok (no benchmark_name= line (output format drift?)) instead of silently looking like an early exit. A driver that announced its benchmark but exited before the metrics line still gets the benign -- cells. Updated the comment block to match the new semantics.
• run_isaac_agent.py:849 (empty except) — narrowed except Exception: pass around _sim.destroy() in the finally teardown to the concrete (RuntimeError, OSError, AttributeError) (matching what IsaacSimulation.destroy()'s own internal cleanup catches) and added a logger.debug(...) + comment so a real teardown failure is logged rather than buried, while keeping teardown best-effort so it never masks the eval result.

Left the two top-level BaseException crash handlers (run_isaac.py:888 / run_isaac_agent.py:871) as-is per your note.

hatch run lint is clean. hatch run test hits a pre-existing native-extension segfault on this box that reproduces on the unmodified branch too (the changes only touch examples/ scripts, which the test suite does not import).