feat(api): 访谈人格/回复长度策略、口述归一、背景语气与输入净稿全链路

Chat 访谈
- 新增 persona 系统（default / warm_listener / curious_guide）与 background_voice 语气层
- 回复长度由 compute_reply_plan 统一决策（brief / standard / expanded），融合信息密度启发式
- 输入净稿（input_normalize）：编排层可选 rules/llm 归一用户口语后再喂模型与记忆检索
- 记忆证据注入：按用户话检索 memory evidence 并注入 prompt

Memoir 回忆录
- 口述归一（oral_normalize）：segment 原文保留，story 管线取派生净稿作叙事输入
- segment 入队批次门闸：累计字数 + 最长等待秒数，减少零碎提交
- fidelity_check / prompts / narrative_agent 微调
- Alembic 0005：清理跨章节 story 外键

Infra
- Dockerfile 加入 ffmpeg
- pyproject.toml 新增依赖并同步 uv.lock
- .env.example / .env.production 补全新配置项

Tests
- 新增 test_background_voice、test_chat_input_normalize、test_experience_regressions
- 扩展 test_interview_prompts、test_interview_reply_length、test_story_route_oral_invariant

Made-with: Cursor

api/app/agents/chat/orchestrator.py

@@ -19,8 +19,20 @@ from app.agents.chat.stage_detection import (
     detect_primary_life_stage,
     life_stage_display_name,
 )
+from app.core.config import settings
+from app.core.dependencies import get_llm_provider
+from app.features.conversation.input_normalize import normalize_chat_input_for_agent
 from app.features.memoir.state_service import get_or_create_state, switch_stage
 
+
+def _llm_for_chat_input_normalize():
+    try:
+        p = get_llm_provider()
+        return getattr(p, "langchain_llm", None)
+    except Exception:
+        return None
+
+
 if TYPE_CHECKING:
     from app.features.user.models import User
 
@@ -31,6 +43,41 @@ _UNAUTH_TURN = AgentChatTurn(
 )
 
 
+async def _fetch_interview_memory_evidence(
+    db: AsyncSession,
+    user_id: str,
+    user_message: str,
+) -> str:
+    """按本轮用户话检索记忆，格式化为短文本；失败或未启用时返回空串。"""
+    from app.core.dependencies import get_embedding_provider
+    from app.features.memory.evidence_format import format_evidence_chunks_for_prompt
+    from app.features.memory.service import MemoryService
+
+    if not settings.chat_memory_retrieval_enabled:
+        return ""
+    msg = (user_message or "").strip()
+    if not msg:
+        return ""
+    try:
+        ms = MemoryService(db, embedding_provider=get_embedding_provider())
+        bundle = await ms.retrieve(user_id, msg, top_k=settings.chat_memory_top_k)
+        text = format_evidence_chunks_for_prompt(bundle.model_dump())
+        t = (text or "").strip()
+        if not t:
+            return ""
+        max_c = settings.chat_memory_evidence_max_chars
+        if len(t) > max_c:
+            return t[: max_c - 3] + "..."
+        return t
+    except Exception as e:
+        try:
+            await db.rollback()
+        except Exception as rollback_error:
+            logger.warning("访谈记忆检索失败后回滚也失败: {}", rollback_error)
+        logger.warning("访谈记忆检索失败: {}", e)
+        return ""
+
+
 class ChatOrchestrator:
     """
     聊天编排器：根据用户资料完成度路由到 ProfileAgent 或 InterviewAgent。
@@ -106,6 +153,10 @@ class ChatOrchestrator:
                     return AgentChatTurn(messages=responses, skip_tts=False)
                 except Exception as e:
                     logger.error(f"资料收集处理失败: {e}", exc_info=True)
+                    return AgentChatTurn(
+                        messages=["不好意思刚才没接住，你再说一遍好吗？"],
+                        skip_tts=False,
+                    )
 
         # --- 正式访谈模式 ---
         user_id = user.id if user else None
@@ -125,9 +176,18 @@ class ChatOrchestrator:
             conversation_id,
             len(user_message or ""),
         )
+        llm_n = None
+        if settings.chat_input_normalize_enabled and (
+            (settings.chat_input_normalize_mode or "").strip().lower() == "llm"
+        ):
+            llm_n = _llm_for_chat_input_normalize()
+        normalized_user_message = normalize_chat_input_for_agent(
+            user_message or "",
+            llm=llm_n,
+        )
         state = await get_or_create_state(user_id, db)
         detected = await detect_primary_life_stage(
-            user_message,
+            normalized_user_message,
             state.current_stage,
             self.interview_agent.llm,
         )
@@ -138,9 +198,11 @@ class ChatOrchestrator:
             conversation.conversation_stage = state.current_stage
             await db.commit()
 
+        from app.agents.chat.background_voice import infer_background_voice
         from app.agents.chat.prompts_profile import format_user_profile_context
 
         user_profile_context = ""
+        background_voice = "default"
         if user:
             user_profile_context = format_user_profile_context(
                 birth_year=user.birth_year,
@@ -148,6 +210,11 @@ class ChatOrchestrator:
                 grew_up_place=user.grew_up_place,
                 occupation=user.occupation,
             )
+            background_voice = infer_background_voice(user.occupation)
+
+        memory_evidence_text = await _fetch_interview_memory_evidence(
+            db, user_id, normalized_user_message
+        )
 
         turn = await self.interview_agent.generate_response_with_state(
             conversation_id=conversation_id,
@@ -155,6 +222,9 @@ class ChatOrchestrator:
             memoir_state=state,
             user_profile_context=user_profile_context,
             detected_user_stage=detected,
+            memory_evidence_text=memory_evidence_text,
+            background_voice=background_voice,
+            normalized_user_message=normalized_user_message,
         )
         if agent_summary_enabled():
             logger.info(
@@ -224,6 +294,9 @@ class ChatOrchestrator:
         user_message_timestamp: datetime | None = None,
         audio_duration_seconds: int | None = None,
         detected_user_stage: str | None = None,
+        memory_evidence_text: str = "",
+        background_voice: str = "default",
+        normalized_user_message: str | None = None,
     ) -> AgentChatTurn:
         """委托 InterviewAgent 生成访谈回复（持久化由调用方负责）。"""
         return await self.interview_agent.generate_response_with_state(
@@ -232,6 +305,9 @@ class ChatOrchestrator:
             memoir_state=memoir_state,
             user_profile_context=user_profile_context,
             detected_user_stage=detected_user_stage,
+            memory_evidence_text=memory_evidence_text,
+            background_voice=background_voice,
+            normalized_user_message=normalized_user_message,
         )
 
     def detect_user_stage(self, user_message: str) -> str:
@@ -243,6 +319,7 @@ class ChatOrchestrator:
         conversation_id: str,
         memoir_state: MemoirStateSchema,
         user_profile_context: str = "",
+        background_voice: str = "default",
     ) -> List[str]:
         """
         委托 InterviewAgent 生成访谈开场白（持久化由调用方 ConversationHistoryStore 负责）。
@@ -251,4 +328,5 @@ class ChatOrchestrator:
             conversation_id=conversation_id,
             memoir_state=memoir_state,
             user_profile_context=user_profile_context,
+            background_voice=background_voice,
         )