fix(chat): 重复追问被拦截时再多问一次模型

防重复问句会把整段回复削成「这一段我记住了。」只剩一句套话时，用带纠偏说明的 system 再调一次 LLM，尽量避免用户只看到干巴巴_ack。仍只重试一次，并打日志与 meta 标记 duplicate_question_guard_llm_retry。

api/app/agents/chat/interview_agent.py

@@ -13,6 +13,7 @@ from app.agents.chat.helpers import format_history_string, get_history_with_wind
 from app.agents.chat.interview_state_hints import (
     apply_duplicate_question_guard,
     extract_recent_questions,
+    segments_are_only_duplicate_guard_fallback,
     update_recent_questions,
 )
 from app.agents.chat.interview_turn_plan import plan_interview_turn
@@ -44,6 +45,45 @@ logger = get_logger(__name__)
 # LLM 不可用或调用失败时对用户展示（不暴露异常细节、不触发 TTS）
 _FALLBACK_REPLY = "刚才网络不太稳，没接上。你可以再说一遍，或稍后再试。"
 
+# 仅在「重复问句守卫」把正文削成单句兜底时追加二次 system，只多调一次模型。
+_DUPLICATE_GUARD_LLM_RETRY_SYSTEM_APPENDIX = """## 二次生成（纠偏）
+上一版模型输出因包含与「最近已问过的问题」或「已确认事实」重复的问句，已被系统弃用。请**重新写一整条回复**：
+- 仍须遵守上文全部主规则；
+- 先贴着用户本轮原话承接半句到一两句（可有画面感）；
+- **禁止**再用与刚才同义、仅换说法的确认型问句；
+- 若要提问，须换**全新角度**，并锚在用户刚说的具体细节里；也可以本轮**完全不提问**，只并肩承接；
+- **禁止**整段只有「这一段我记住了」或同类无信息套话。"""
+
+
+def _finalize_chat_segments_after_llm(
+    response_text: str,
+    *,
+    max_segments: int,
+    max_chars: int,
+    memoir_state: MemoirStateSchema,
+    recent_questions: list[str],
+) -> tuple[list[str], bool]:
+    raw_list = segments_from_llm_response(
+        response_text,
+        max_segments=max_segments,
+    )
+    if not raw_list:
+        raw_list = [response_text.strip()]
+    out = truncate_chat_segments(
+        raw_list,
+        max_segments=max_segments,
+        max_chars_per_segment=max_chars,
+    )
+    if not out:
+        out = [response_text.strip()[:max_chars]]
+    out = nonempty_segments_or_fallback(out, fallback=_FALLBACK_REPLY)
+    out, deduped = apply_duplicate_question_guard(
+        out,
+        state=memoir_state,
+        recent_questions=recent_questions,
+    )
+    return out, deduped
+
 
 def _get_langchain_llm():
     try:
@@ -219,29 +259,70 @@ class InterviewAgent:
             log_agent_payload(
                 logger, "InterviewAgent.generate_response.raw_response", response_text
             )
-            raw_list = segments_from_llm_response(
+            rq_base = recent_questions or memoir_state.recent_questions
+            out, deduped = _finalize_chat_segments_after_llm(
                 response_text,
                 max_segments=max_segments,
+                max_chars=max_chars,
+                memoir_state=memoir_state,
+                recent_questions=rq_base,
             )
-            if not raw_list:
-                raw_list = [response_text.strip()]
-            out = truncate_chat_segments(
-                raw_list,
-                max_segments=max_segments,
-                max_chars_per_segment=max_chars,
-            )
-            if not out:
-                out = [response_text.strip()[:max_chars]]
-            out = nonempty_segments_or_fallback(out, fallback=_FALLBACK_REPLY)
-            out, deduped = apply_duplicate_question_guard(
-                out,
-                state=memoir_state,
-                recent_questions=recent_questions or memoir_state.recent_questions,
-            )
-            updated_recent_questions = update_recent_questions(
-                recent_questions or memoir_state.recent_questions,
-                out,
-            )
+            retry_used = False
+            if deduped and segments_are_only_duplicate_guard_fallback(out):
+                retry_system = (
+                    f"{system_prompt}\n\n{_DUPLICATE_GUARD_LLM_RETRY_SYSTEM_APPENDIX}"
+                )
+                retry_messages: List[Any] = [
+                    SystemMessage(content=retry_system),
+                    *hw.window,
+                    HumanMessage(content=text_for_model),
+                ]
+                log_agent_payload(
+                    logger,
+                    "InterviewAgent.generate_response.retry_prompt",
+                    format_history_string(
+                        retry_messages,
+                        omit_system_body=settings.agent_log_omit_system_message_body,
+                    ),
+                )
+                llm_t1 = time.perf_counter()
+                with agent_span(
+                    logger,
+                    "InterviewAgent.generate_response.llm_retry",
+                    conversation_id=conversation_id,
+                    stage=memoir_state.current_stage,
+                ):
+                    logger.info(
+                        "event=chat_prompt_built agent=InterviewAgent.duplicate_guard_retry "
+                        "prompt_chars={} conversation_id={}",
+                        _message_contents_char_count(retry_messages),
+                        conversation_id,
+                    )
+                    response_retry = await chat_llm.ainvoke(retry_messages)
+                logger.info(
+                    "event=chat_llm_done agent=InterviewAgent.duplicate_guard_retry "
+                    "response_latency_ms={:.2f}",
+                    (time.perf_counter() - llm_t1) * 1000,
+                )
+                response_text_retry = (
+                    response_retry.content
+                    if hasattr(response_retry, "content")
+                    else str(response_retry)
+                )
+                log_agent_payload(
+                    logger,
+                    "InterviewAgent.generate_response.raw_response_retry",
+                    response_text_retry,
+                )
+                out, deduped = _finalize_chat_segments_after_llm(
+                    response_text_retry,
+                    max_segments=max_segments,
+                    max_chars=max_chars,
+                    memoir_state=memoir_state,
+                    recent_questions=rq_base,
+                )
+                retry_used = True
+            updated_recent_questions = update_recent_questions(rq_base, out)
             log_agent_summary(
                 logger,
                 "InterviewAgent.generate_response segments={} conversation_id={} "
@@ -256,6 +337,7 @@ class InterviewAgent:
                 interview_state_meta={
                     "recent_questions": updated_recent_questions,
                     "duplicate_question_guard_triggered": deduped,
+                    "duplicate_question_guard_llm_retry": retry_used,
                 },
             )
         except Exception as e: