diff --git a/docs/design/fullduplex.md b/docs/design/fullduplex.md
index 324cd5bdc..df698880e 100644
--- a/docs/design/fullduplex.md
+++ b/docs/design/fullduplex.md
@@ -286,8 +286,11 @@ output   DuplexOrchestrator._intercept_stage_output -> runner.on_stage_output ->
 detach   DuplexOmni.detach_session -> touch(DETACH): engine-owned disconnect grace; expiry -> SessionExpired
 resume   DuplexOmni.resume_session(expected_lease_generation) -> lease CAS; the existing handle is re-entered
 close    DuplexOmni.close_session -> close RPC; the manager tears the runner down, then the stage cleanup
-         (abort submitted requests, release reserved ids), then SessionClosed, then the RPC result: seeing
-         the event means the admission slot is free (it is held until the cleanup succeeded)
+         (abort submitted requests, release reserved ids), then SessionClosed, then the RPC result.
+         SessionClosed is emitted after the cleanup attempt (in a finally), so it is also sent when the
+         stage cleanup failed; in that case the admission slot is intentionally retained and the cleanup
+         is retried by the reaper, and opening a replacement session can still be refused with
+         resource_exhausted. Seeing the event therefore does not by itself guarantee a free slot.
 reap     DuplexSessionManager.reaper_loop: idle TTL / disconnect grace expiry, cleanup retries
 ```
 
diff --git a/vllm_omni/engine/duplex/session/manager.py b/vllm_omni/engine/duplex/session/manager.py
index 15e038806..605ed8eba 100644
--- a/vllm_omni/engine/duplex/session/manager.py
+++ b/vllm_omni/engine/duplex/session/manager.py
@@ -604,9 +604,12 @@ class DuplexSessionManager:
         The session keeps its admission slot (``_closing``) until stage cleanup
         succeeded; a failed cleanup is retried by the reaper. For an explicit
         close the ``session.closed`` event is emitted only after the cleanup
-        attempt, so a client that sees it can open a replacement session at
-        once (the runner used to emit it before the stage requests were
-        aborted, which let a prompt reopen hit ``resource_exhausted``).
+        attempt, in a ``finally``, including when the cleanup failed (the
+        runner used to emit it before the stage requests were aborted). A
+        client that sees the event can normally open a replacement session
+        right away, but after a failed cleanup the slot is still held until
+        the reaper succeeds, so a prompt reopen can be refused with
+        ``resource_exhausted``.
         """
         session = runner.session
         submitted = tuple(session.resource_request_ids(submitted=True))