test(gateway): cover auto-resume full-path no-regression; clarify guard docstring

The salvaged fix's two regression tests mock adapter.handle_message, so
they only assert the pre-claimed sentinel is set/cleaned around a stub —
they never drive the real dispatch chain. Add a full-path test that
exercises _schedule_resume_pending_sessions -> _guarded_handle_message ->
adapter.handle_message -> _process_message_background -> _handle_message
and asserts the resumed session's agent runs EXACTLY ONCE: not zero (the
pre-claim must not self-bounce the resume into a queued no-op) and not
twice (the duplicate-agent bug #45456 the fix targets). Also assert no
leaked sentinel and no orphaned pending event after the drain settles.

Tighten the _guarded_handle_message docstring: on current main the real
sentinel is taken over inside _handle_message (not _process_message_background),
and note the `is _AGENT_PENDING_SENTINEL` guard only releases the slot we
ourselves placed, never one a live run owns.

gateway/run.py

@@ -4579,12 +4579,16 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
             ) -> None:
                 """Ensure the pre-claimed sentinel is always released.
 
-                In the normal flow, ``_process_message_background`` sets
-                its own sentinel and releases it in its ``finally`` block.
-                If ``handle_message`` raises *before* reaching that
-                method (e.g. during topic recovery or session key
+                In the normal flow the resume turn reaches
+                ``_handle_message``, which replaces our pre-claim with
+                its own ``_AGENT_PENDING_SENTINEL`` (and releases it in
+                its ``finally`` block) once the run begins.  If
+                ``handle_message`` raises *before* the runner takes over
+                the slot (e.g. during topic recovery or session-key
                 resolution), nobody clears our pre-claim — so we do it
-                here unconditionally.
+                here unconditionally.  The ``is _AGENT_PENDING_SENTINEL``
+                guard below only releases the slot we ourselves placed,
+                never one a live run currently owns.
                 """
                 try:
                     await _adapter.handle_message(_event)