fix(isaac): pump app after timeline.stop() so the #159 velocity crash actually clears

[yinsong1986]

Summary

examples/libero/run_isaac.py still crashed end-to-end after both prior
fixes for #159 (PR #161 probe-gated, PR #163 unconditional
timeline.stop()), with the same bare exception during
DynamicCuboid construction in load_scene:

Exception: Failed to get rigid body velocities from backend

This PR makes the deferral actually take effect.

Root cause

timeline.stop() is asynchronous. It posts a GLOBAL_EVENT_STOP
that tears down the physics-tensor view only on a later app tick.
Isaac's RigidPrim.__init__ gates its eager _on_physics_ready
velocity query on:

if SimulationManager.get_physics_sim_view() is not None:   # rigid_prim.py:200
    ... RigidPrim._on_physics_ready(self, None)             # -> get_velocities() -> raises

and that handle stays non-None until the stop event is dispatched.
So #163's "stop the timeline, then immediately construct the Dynamic*
prim" still saw a live view and hit the eager query.

A diagnostic on a real Isaac Sim 6.0 run confirms it:

[probe:after reset+step]              get_physics_sim_view() is None == False
[probe:immediately after stop()]      get_physics_sim_view() is None == False   <-- #163 builds the prim here
[probe:after app.update() x3]         get_physics_sim_view() is None == True

Fix

After timeline.stop(), pump SimulationApp.update() until
SimulationManager.get_physics_sim_view() returns None, bounded to 8
ticks (so a host that never clears fails loud at prim construction rather
than spinning). SimulationApp.update() is render-only — it does not
advance physics — so it doesn't drift the already-placed prims. Converted
_stop_timeline_for_deferred_physics from a @staticmethod to an
instance method so it can reach self._app.

Verification

• End-to-end (Isaac Sim 6.0, L4 GPU, headless):
  OMNI_KIT_ACCEPT_EULA=YES python examples/libero/run_isaac.py --policy mock --n-episodes 3 --seed 42
  • Before: exit 1, on_episode_start failed ... Failed to get rigid body velocities, 720-frame warm-up-only MP4.
  • After: completes all 3 episodes, 0 frame errors, records a
    real 2160-frame (72 s) rollout MP4. No traceback, clean shutdown.
• Unit tests: strands_robots_sim/isaac/tests/test_unit.py — 189
  passed (1 deselected: a pre-existing env flake unrelated to this
  change — test_is_available_probes_omni_isaac_kit_specifically fails
  only because the dev box has real omni installed; it passes on the
  CPU CI runner). Added:
  • test_stop_timeline_pumps_app_until_physics_view_cleared
  • test_stop_timeline_pump_is_bounded
  • updated test_stop_timeline_is_best_effort_without_omni for the
    staticmethod→instance change.
• Lint: black --check, isort --check-only, flake8 all clean
  (line-length 120, per pyproject.toml).

Notes

• This only fixes the crash that aborted the eval. The LIBERO→USD
  scene translation itself (table/objects/robot placement fidelity) is
  separate and out of scope here.
• Refs Isaac LIBERO: on_episode_start crashes in load_scene — DynamicCuboid queries velocities before physics-tensor backend is ready ("Failed to get rigid body velocities") #159 (reopened after fix(isaac): defer dynamic-prim physics init during load_scene (closes #159) #161/fix(isaac): stop timeline unconditionally before dynamic prim construction (#159) #163 proved necessary-but-insufficient).

[cagataycali]

Approve. The root-cause analysis is correct and the fix is surgical.

The key insight - timeline.stop() posts an async GLOBAL_EVENT_STOP that only tears down the physics-tensor view on a later tick, so the prior unconditional stop was necessary-but-insufficient - matches the documented Isaac 6.0 diagnostic. Pumping SimulationApp.update() (render-only, no physics advance) until get_physics_sim_view() clears is the right deferral, and bounding it to 8 ticks (fail-loud at prim construction rather than spin) is a sound safety valve.

Defensive details I checked and like:

• view is checked at the top of each iteration, so an already-cleared view returns immediately with zero pumps;
• graceful degradation when SimulationManager or self._app.update is unavailable (returns without pumping);
• the staticmethod to instance-method change is the minimal way to reach self._app.

Regression coverage pins both behaviours precisely: the happy-path test asserts exactly 2 update calls for a [obj, obj, None] view sequence, and the bound test asserts exactly 8 when the view never clears. CI green, scope tight, docstring carries the why. Good to merge.