refactor: follow-up for salvaged PR #81692

- warn (not debug) on final text-turn flush failure: a failure here
  reopens the exact #81641 data-loss window with _persist_session as
  the only remaining retry, unlike the verify siblings which retry
  in-loop; include session id for triage
- trim the flush-site comment to sibling proportion, pointing to the
  test module for the full incident narrative
- test: assert _persist_session presence before indexing, so a wiring
  change fails with a clean assertion instead of ValueError from max()

agent/conversation_loop.py

@@ -7613,25 +7613,23 @@ def run_conversation(
                     continue
 
                 messages.append(final_msg)
-                # Make the completed answer durable before leaving the loop.
-                # The text already reached the user through the streaming /
-                # interim display path, which is display-only and never writes
-                # to state.db; the next durable write would otherwise be
-                # finalize_turn's _persist_session, behind post-turn work that
-                # can include micro-compaction's aux-LLM call. A session torn
-                # down inside that window (ws_orphan_reap after a remote 1006
-                # close, container stop) lost a reply the user had already
-                # seen (#81641). Same contract the tool-call exit gets from
-                # #49045 and the verify-on-stop exits above; the
-                # _DB_PERSISTED_MARKER dedup keeps _persist_session idempotent.
-                #
-                # Unlike the tool-call exit, a failed flush must NOT abort the
-                # turn: no side effect runs after this point, the answer is
-                # already produced, and _persist_session retries the write.
+                # Make the completed answer durable before leaving the loop —
+                # a session torn down before finalize_turn's _persist_session
+                # otherwise loses a reply the user already saw (#81641). Same
+                # contract as the tool-call exit (#49045) and the verify exits
+                # above; _DB_PERSISTED_MARKER keeps _persist_session idempotent.
+                # Unlike the tool-call exit, failure must NOT abort the turn:
+                # no side effect follows and _persist_session retries the write.
+                # Full incident narrative: tests/run_agent/test_81641_*.py.
                 try:
                     agent._flush_messages_to_session_db(messages, conversation_history)
                 except Exception:
-                    logger.debug("final text-turn flush failed", exc_info=True)
+                    logger.warning(
+                        "final text-turn flush failed (session=%s) — reply is "
+                        "not yet durable; relying on finalize_turn retry",
+                        getattr(agent, "session_id", None) or "none",
+                        exc_info=True,
+                    )
 
                 _turn_exit_reason = f"text_response(finish_reason={finish_reason})"
                 if not agent.quiet_mode:

tests/run_agent/test_81641_text_turn_incremental_persistence.py

@@ -138,9 +138,14 @@ class TestCompletedTextTurnIncrementalPersistence:
             f"transcript tail; observed events: {events!r}"
         )
 
-        final_persist = max(
+        persist_indices = [
             i for i, (kind, _) in enumerate(events) if kind == "persist_session"
+        ]
+        assert persist_indices, (
+            "finalize_turn must call _persist_session after the loop exits; "
+            f"observed events: {events!r}"
         )
+        final_persist = persist_indices[-1]
         assert answer_flushes[0] < final_persist, (
             "The assistant row must be durable BEFORE post-loop finalization, "
             f"but the answer flush at index {answer_flushes[0]} did not precede "