Fix PhysicsManager.clear_callbacks ID-recycle bug

clear_callbacks() reset _callback_id back to 0, so subsequent
register_callback() calls handed out IDs that collided with old
CallbackHandle.id values still held by not-yet-garbage-collected
sensors / assets.  When the old object's __del__ eventually ran it
deregistered the NEW callback by ID — leaving the new sensor's
_initialize_callback wired up but never fired.

The symptom under kitless OVPhysX was a sensor that registered its
callback, then PHYSICS_READY dispatched, then sensor.reset()
crashed with 'has no attribute _ALL_ENV_MASK' (init never ran).
The PhysX backend hid this because its callback registry is the
Carbonite event bus rather than this Python dict.

Keep _callback_id monotonic across clear_callbacks invocations to
preserve uniqueness for the lifetime of the process.

Author: AntoineRichard

source/isaaclab/isaaclab/physics/physics_manager.py

@@ -156,11 +156,23 @@ def dispatch_event(cls, event: PhysicsEvent, payload: Any = None) -> None:
 
     @classmethod
     def clear_callbacks(cls) -> None:
-        """Remove all registered callbacks."""
+        """Remove all registered callbacks.
+
+        Do NOT reset ``_callback_id`` — handle IDs must remain monotonically
+        unique across the lifetime of the process.  Resetting the counter
+        would let a future :meth:`register_callback` hand out an ID that an
+        old, still-alive :class:`CallbackHandle` (e.g. on a sensor that has
+        not been garbage-collected yet) holds, so when the old object
+        eventually finalizes its ``__del__`` would deregister the new
+        callback.  This bit ovphysx's kitless multi-context tests where two
+        ``InteractiveScene``s are created in sequence: the first scene's
+        sensor would post-GC deregister the second scene's
+        ``_initialize_callback`` by ID collision, leaving the second sensor
+        forever uninitialized.
+        """
         for cid in list(cls._callbacks.keys()):
             cls.deregister_callback(cid)
         cls._callbacks.clear()
-        cls._callback_id = 0
 
     @classmethod
     def _wrap_weak_ref(cls, callback: Callable) -> Callable:

source/isaaclab_ovphysx/isaaclab_ovphysx/sensors/contact_sensor/contact_sensor.py

@@ -229,13 +229,20 @@ def _initialize_impl(self) -> None:
             )
 
         # Optional: pose tracking via a RIGID_BODY_POSE tensor binding.
-        # ovphysx uses fnmatch and does not brace-expand, so we widen to a single
-        # "*" leaf pattern under the base glob. This relies on the prim_path
-        # already isolating the sensor bodies (e.g. ".*_FOOT" matches all four
-        # feet and no siblings). The post-bind count check below catches a
-        # mismatch.
+        # ovphysx fnmatch does not brace-expand, so we cannot match multiple
+        # body names with a single glob.  Single-body sensors (the common case
+        # — one prim path per sensor) use a tight per-body pattern.  Multi-body
+        # sensors are rejected here; they need per-body bindings + an
+        # interleaved-read kernel that does not exist yet.
         if self.cfg.track_pose:
-            single_pose_pattern = f"{base_glob}/*"
+            if self._num_sensors != 1:
+                raise NotImplementedError(
+                    "ovphysx ContactSensor.track_pose is not yet supported for sensors that "
+                    f"resolve to more than one body per env (got {self._num_sensors} bodies "
+                    f"under '{self.cfg.prim_path}').  Workaround: create one ContactSensor "
+                    "per body."
+                )
+            single_pose_pattern = f"{base_glob}/{body_names[0]}"
             self._pose_binding = physx_instance.create_tensor_binding(
                 pattern=single_pose_pattern,
                 tensor_type=TT.RIGID_BODY_POSE,