feat(api): 访谈路径轻量门控、Memoir Phase1 批处理与叙事/记忆管线加固

- 新增 utterance_substance：短时/应答/元话语可跳过记忆检索、阶段 LLM 与资料抽取 LLM；可配置
- 输入归一化：LLM 模式默认仅语音/ASR；配置项写入 .env.example
- Memoir Phase1：可选 batch LLM 一次性抽取+分类（失败回退逐段）；Extraction 空槽位时阶段与 current_stage 对齐，prompt 约束收紧
- 叙事与忠实度：narrative_safety、证据重叠/场合锚点、标题 slots 与履历短语 grounded；fidelity 解析失败 fail-open 可配置
- 章节管线：锁 TTL 上调、锁竞争 Celery 重试、Phase2 immediate singleflight 等；story_pipeline_sync / chapter_compose / memoir_tasks 联动
- Memory：compaction / repo / summarizer / evidence 小修；事实 FTS 未命中是否回退最近事实可配置
- 新增 memoir_pipeline_trace；补充 memoir_reliability 文档与多项回归/门控测试

api/app/agents/chat/orchestrator.py

@@ -19,9 +19,13 @@ from app.agents.chat.stage_detection import (
     detect_primary_life_stage,
     life_stage_display_name,
 )
+from app.agents.chat.utterance_substance import should_run_chat_stage_memory_heavy_work
 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.conversation.input_normalize import (
+    apply_conversation_input_rules,
+    normalize_chat_input_for_agent,
+)
 from app.features.memoir.state_service import get_or_create_state, switch_stage
 
 
@@ -58,6 +62,11 @@ async def _fetch_interview_memory_evidence(
     msg = (user_message or "").strip()
     if not msg:
         return ""
+    if (
+        settings.chat_memory_retrieval_require_substantive
+        and not should_run_chat_stage_memory_heavy_work(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)
@@ -122,9 +131,19 @@ class ChatOrchestrator:
                         missing,
                         len(user_message or ""),
                     )
-                    extracted = await self.profile_agent.extract_profile_from_message(
-                        user_message, missing, conversation_id=conversation_id
-                    )
+                    run_extract = True
+                    if settings.chat_profile_extract_require_substantive:
+                        rules_only = apply_conversation_input_rules(user_message or "")
+                        run_extract = should_run_chat_stage_memory_heavy_work(
+                            rules_only
+                        )
+                    extracted = None
+                    if run_extract:
+                        extracted = (
+                            await self.profile_agent.extract_profile_from_message(
+                                user_message, missing, conversation_id=conversation_id
+                            )
+                        )
                     if extracted:
                         await apply_extracted_profile_fn(user, extracted, db)
 
@@ -184,12 +203,17 @@ class ChatOrchestrator:
         normalized_user_message = normalize_chat_input_for_agent(
             user_message or "",
             llm=llm_n,
+            is_from_voice=is_from_voice,
         )
         state = await get_or_create_state(user_id, db)
+        substantive_turn = should_run_chat_stage_memory_heavy_work(
+            normalized_user_message
+        )
         detected = await detect_primary_life_stage(
             normalized_user_message,
             state.current_stage,
             self.interview_agent.llm,
+            skip_llm=not substantive_turn,
         )
         if detected != state.current_stage:
             state = await switch_stage(user_id, detected, db)

api/app/agents/chat/stage_detection.py

@@ -55,15 +55,22 @@ async def detect_primary_life_stage(
     user_message: str,
     current_stage: str,
     llm: Any,
+    *,
+    skip_llm: bool = False,
 ) -> str:
     """
     返回合法的人生阶段 key；失败时回退为 current_stage。
+    skip_llm=True 时仅用关键词（短时/元话语等路径，不调阶段 LLM）。
     """
     fb = normalize_chat_stage(current_stage, "childhood")
     if not settings.chat_stage_detection_enabled:
         k = keyword_fallback_primary_stage(user_message)
         return normalize_chat_stage(k, fb) if k else fb
 
+    if skip_llm and settings.chat_stage_detection_skip_llm_on_insufficient_signal:
+        k = keyword_fallback_primary_stage(user_message)
+        return normalize_chat_stage(k, fb) if k else fb
+
     if not llm:
         k = keyword_fallback_primary_stage(user_message)
         return normalize_chat_stage(k, fb) if k else fb

api/app/agents/chat/utterance_substance.py

@@ -0,0 +1,73 @@
+"""
+启发式判断访谈「本轮」是否值得跑阶段 LLM / 记忆检索等高成本步骤。
+
+短答、应答词、元话语（谈整理回忆本身而非人生经历）为 False；长文本或中等长度非常用词为 True。
+与配置 `chat_substantive_*` 配合；关闭启发式时恒为 True。
+"""
+
+from __future__ import annotations
+
+import re
+from typing import Final
+
+from app.core.config import settings
+
+# 极短应答（整句精确匹配）
+_SHORT_ACK_EXACT: Final[frozenset[str]] = frozenset(
+    {
+        "嗯",
+        "对",
+        "好",
+        "是",
+        "行的",
+        "是的",
+        "没有",
+        "行",
+        "噢",
+        "哦",
+        "好吧",
+        "嗯嗯",
+        "对对",
+        "好嘞",
+        "对的",
+        "没了",
+        "可以",
+        "就这样",
+        "还行",
+        "还好",
+    }
+)
+
+# 元话语：谈回忆过程/访谈本身，不足以切换人生阶段或拉记忆证据
+_META_PROCESS: Final[re.Pattern[str]] = re.compile(
+    r"(回忆|想起).{0,20}(细节|收获|快忘|忘的|很多东西)"
+    r"|(整理|聊聊|谈到).{0,8}(回忆|访谈|记录)"
+    r"|最大的收获",
+    re.UNICODE,
+)
+
+
+def should_run_chat_stage_memory_heavy_work(text: str) -> bool:
+    """
+    True：值得调用阶段检测 LLM、记忆检索（向量等）。
+    False：仅用关键词阶段回退、跳过记忆检索。
+    """
+    if not settings.chat_substantive_heuristic_enabled:
+        return True
+    s = (text or "").strip()
+    if not s:
+        return False
+    # 元话语可略长，须在「达到 min_chars」分支之前判断
+    if _META_PROCESS.search(s):
+        return False
+    min_chars = int(settings.chat_substantive_min_chars)
+    if len(s) >= min_chars:
+        return True
+    if s in _SHORT_ACK_EXACT:
+        return False
+    if len(s) <= 4:
+        # 极短：多为语气/应答
+        if all(ch in "嗯哦噢对对好好的没行是的不没一下的了呗嘛呀啊" for ch in s):
+            return False
+    # 偏短但未命中噪音规则：默认走完整路径；5 字常见为有信息短句（旧逻辑用 >=6 会误杀）
+    return len(s) >= 5