Explain the stream-scoped slot reuse rule in plain words

Author: pv-nvidia

source/isaaclab_ov/isaaclab_ov/renderers/ovrtx_renderer_strategies.py

@@ -244,10 +244,18 @@ class _AsyncRenderStrategy(_RenderStrategy):
     Rendering then overlaps the next step's simulation and Python work. Camera outputs are one
     step stale.
 
-    Transform staging always uses two slots. A finished write op is only a fence in the CUDA
-    stream the write named. Work on other streams is not ordered after OVRTX's read. Slot refills
-    are safe because they run on the same stream that every write names. A refill from any other
-    stream would race the read.
+    Transform staging always uses two slots, so one slot can be refilled while the other still
+    backs the frame in flight.
+
+    Refilling a slot buffer is safe under one condition, and it is narrower than it looks. OVRTX
+    can still be reading a buffer on the GPU after its write op completes. A completed write op
+    guarantees one thing: OVRTX has planted a wait in the CUDA stream that the write named. GPU
+    work that enters that stream afterwards starts only after OVRTX has finished reading. GPU work
+    on any other stream is not held back and could overwrite the buffer during the read.
+
+    The refill kernels run on the device's current Warp stream, and every write names that same
+    stream, so the wait covers them. Enqueue any work that touches a slot buffer on that stream
+    and no other.
     """
 
     # See :meth:`_create_slots` for why two is always enough.