fix(isaac): record rollout videos at MuJoCo parity (closes #112)

[yinsong1986]

What

Brings the Isaac LIBERO examples to parity with their MuJoCo counterparts on rollout-video output. They now write a real rollouts/<date>/<name>__<camera>.mp4 instead of an …__image.mp4.placeholder path string that was never written to disk.

Why

run_mujoco.py / run_mujoco_agent.py record a real MP4 via a start_cameras_recording / stop_cameras_recording pair. run_isaac.py / run_isaac_agent.py recorded nothing — they only emitted a placeholder path in the grep-stable videos= line, so any tooling that trusts videos= (e.g. the R15 backend matrix's rollouts/*/*--task=*.mp4 glob) got a dangling/blank entry for Isaac rows.

The deferral rationale was stale: IsaacSimulation had no recorder API (start_cameras_recording/stop_cameras_recording are MuJoCo-Simulation-specific, not on the base SimEngine), and the comments cited "render() returns blank frames". With #61/#62 merged, IsaacSimulation.render() returns real RGB frames on Isaac Sim 6.0, so videos can be produced.

Approach

This implements both suggested fixes from the issue — a backend recorder and the example-level on_frame wiring — because they compose: the recorder is synchronous (Isaac's RTX renderer + Camera.get_rgba are bound to the thread that booted SimulationApp, so a MuJoCo-style daemon recorder would deadlock), so it relies on evaluate_benchmark's on_frame= hook to capture one render() frame per control step on the eval thread.

Backend (strands_robots_sim/isaac/simulation.py)

• start_cameras_recording(cameras, output_dir, fps, name, max_frames_per_camera) — sets up one in-memory RGB buffer per camera and returns an on_frame(step, obs, action) closure in the result's json block.
• stop_cameras_recording() — flushes buffers to {name}__{camera}.mp4 via imageio (the same encoder the MuJoCo recorder uses), matching MuJoCo's filename/layout convention exactly. Idempotent.
• Errors loudly on no-world / double-start / unknown-camera / no-cameras; per-camera render failures increment an error counter rather than aborting the eval; recorder state is cleared on destroy().

Examples

• run_isaac.py — replaces the .placeholder with a real start/stop pair around evaluate_benchmark(on_frame=...). videos= now points at a file that exists.
• run_isaac_agent.py — arms the recorder before the agent runs, flushes after. The on_frame closure is stashed at module scope (it can't cross the @tool JSON-schema boundary — same constraint documented in start_cameras_recording daemon-thread artifacts under multi-threaded eval (Strands Agent tool dispatch) robots#191) and the tool wrapper threads it into evaluate_benchmark.
• Stale "blank frames" / "intentionally skipped" comments dropped; docstrings updated.

Test surface

• New TestCamerasRecording (11 cases): lifecycle, error paths (no-world / unknown camera / no cameras / double-start), idempotent stop, full capture→encode cycle asserting the MuJoCo-compatible {name}__{camera}.mp4 filename + a non-empty MP4, stale-closure no-op after stop, render-failure error counting, frame cap, and destroy() cleanup.
• The two MP4-encode cases importorskip("imageio") so the lightweight hatch run test venv (which doesn't ship imageio) skips them cleanly; they run and pass in a full env.

hatch run lint ✅ and hatch run test ✅ both pass locally (212 passed, 35 skipped). No new failures. (The pre-existing X11/OpenGL environmental skips/failures in test_rendering.py / test_policy_runner.py are unaffected — those tests aren't in strands_robots_sim/isaac/tests/.)

Acceptance criteria

• run_isaac.py and run_isaac_agent.py write a real rollouts/<date>/<name>__<camera>.mp4, like the MuJoCo scripts.
• The videos= line points at a file that exists.
• Filename/layout matches MuJoCo so the backend-matrix glob picks up Isaac rows.

Out of scope (no follow-up filed yet)

• End-to-end GPU validation of the produced MP4 on an Isaac Sim 6.0 host (this dev box has no Isaac Sim runtime); the recorder is unit-tested against a mocked RTX camera handle that returns real-shaped frames. The capture path reuses the already-merged render() frame path (feat(isaac): wire add_camera Phase 2 + new remove_camera (refs #14) #61/feat(isaac): wire render frame-path Phase 2 (refs #14) #62).
• Migrating run_isaac_agent.py to the full AgentTool enum once IsaacSimulation inherits AgentTool (tracked under R7: Add IsaacSimulation(SimEngine) backend (epic — Stage 3) #14) — the module-scope _on_frame indirection goes away then.

Closes #112. Related: #101 (6.0 migration / render works), #61/#62 (camera + render frame-path), strands-labs/robots#191 (on_frame synchronous-recorder rationale).