docs: update for pipecat PR #4938

api-reference/server/utilities/turn-management/filter-incomplete-turns.mdx

@@ -31,6 +31,12 @@ The system automatically:
 4. Starts a timeout based on the incomplete type
 5. Re-prompts the LLM when the timeout expires
 
+**Safeguards:**
+
+- **Stale completion protection**: A completion (`✓`) resolves with some latency, so the user may have resumed speaking by the time it arrives. Completions that arrive while the user is speaking are dropped to prevent talking over them.
+- **Duplicate completion prevention**: The acoustic detector can trigger several inferences within one user turn, each producing a `✓`. Only the first completion is voiced; later duplicates are dropped until the user starts a new turn or resumes speaking.
+- **Smart timeout cancellation**: Re-prompt timeouts are cancelled when the user resumes speaking or when a new inference begins, preventing redundant inferences or prompts that interrupt the user.
+
 <Note>
   The public
   [`on_user_turn_stopped`](/api-reference/server/utilities/turn-management/turn-events#on_user_turn_stopped)

api-reference/server/utilities/turn-management/user-turn-strategies.mdx

@@ -469,6 +469,8 @@ stop = [
 ]
 ```
 
+A completion resolves with some latency, so the user may have resumed speaking by the time it arrives. The controller drops a finalization that arrives while the user is speaking, so a stale completion does not end the turn (and talk over the user); the turn stays open for the next inference to re-evaluate.
+
 <Note>
   If the producer never emits `UserTurnInferenceCompletedFrame`, the
   `user_turn_stop_timeout` watchdog finalizes the turn after no activity. Tune