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/features/memoir/narrative_safety.py

@@ -0,0 +1,165 @@
+"""叙事落库前的确定性安检：防止 prompt 分区标记或摘录块泄漏进正文。"""
+
+from __future__ import annotations
+
+# 与 app.agents.memoir.prompts.format_narrative_user_content 保持一致
+ORAL_SECTION_MARKER = "【本段用户口述】"
+EVIDENCE_SECTION_MARKER = "【仅供参考的相关记忆摘录"
+
+# 摘录引导语中的固定短语（用于粗检）
+EVIDENCE_SECTION_TAIL = "不得把其中具体事实写成本轮亲历经历"
+
+
+def body_contains_prompt_artifact(markdown_body: str) -> bool:
+    s = (markdown_body or "").strip()
+    if not s:
+        return False
+    if ORAL_SECTION_MARKER in s:
+        return True
+    if EVIDENCE_SECTION_MARKER in s:
+        return True
+    if EVIDENCE_SECTION_TAIL in s:
+        return True
+    return False
+
+
+def longest_common_substring_len(a: str, b: str, min_len: int = 14) -> int:
+    """O(n*m) DP；仅用于短 evidence / body，防止过大。"""
+    a = a or ""
+    b = b or ""
+    if len(a) > 8000 or len(b) > 8000:
+        return 0
+    if not a or not b:
+        return 0
+    best = 0
+    prev = [0] * (len(b) + 1)
+    for i in range(1, len(a) + 1):
+        cur = [0] * (len(b) + 1)
+        for j in range(1, len(b) + 1):
+            if a[i - 1] == b[j - 1]:
+                cur[j] = prev[j - 1] + 1
+                if cur[j] > best:
+                    best = cur[j]
+            else:
+                cur[j] = 0
+        prev = cur
+    return best if best >= min_len else 0
+
+
+def evidence_substring_leak_score(
+    body: str, evidence_plain: str, min_len: int = 14
+) -> int:
+    """
+    若正文与 evidence 存在较长公共子串，且该子串不在 oral/existing 中，
+    则视为摘录渗漏风险（返回子串长度），否则 0。
+    """
+    body = (body or "").strip()
+    ev = (evidence_plain or "").strip()
+    if not body or not ev or len(ev) < min_len:
+        return 0
+    return longest_common_substring_len(body, ev, min_len=min_len)
+
+
+def longest_common_substring(a: str, b: str) -> str:
+    """返回 a、b 的最长公共子串（长度上限防 DP 爆内存）。"""
+    a = a or ""
+    b = b or ""
+    if len(a) > 8000 or len(b) > 8000:
+        return ""
+    best_i, best_len = 0, 0
+    prev = [0] * (len(b) + 1)
+    for i in range(1, len(a) + 1):
+        cur = [0] * (len(b) + 1)
+        for j in range(1, len(b) + 1):
+            if a[i - 1] == b[j - 1]:
+                cur[j] = prev[j - 1] + 1
+                if cur[j] > best_len:
+                    best_len = cur[j]
+                    best_i = i
+            else:
+                cur[j] = 0
+        prev = cur
+    if best_len <= 0:
+        return ""
+    start = best_i - best_len
+    return a[start:best_i]
+
+
+# 具体场合描写：易由「相关摘录」渗入正文但长 LCS 抓不住（词短）。
+EVIDENCE_SCENE_ANCHOR_TOKENS: tuple[str, ...] = (
+    "聚餐",
+    "酒席",
+    "酒桌",
+    "宴会",
+    "宴席",
+    "当晚",
+    "那晚",
+    "昨夜",
+    "前一晚",
+    "前一天晚上",
+)
+
+
+def evidence_scene_anchor_leak(
+    body: str,
+    evidence_plain: str,
+    oral: str,
+    existing: str,
+) -> bool:
+    """
+    True：正文出现了与「摘录」共享的具体场合锚点词，且口述与旧正文均未出现，
+    视为摘录场景渗漏（短词不走 LCS 阈值）。
+    """
+    body = (body or "").strip()
+    ev = (evidence_plain or "").strip()
+    o = (oral or "").strip()
+    ex = (existing or "").strip()
+    if not body or not ev:
+        return False
+    base = f"{o}\n{ex}"
+    for tok in EVIDENCE_SCENE_ANCHOR_TOKENS:
+        if tok not in body:
+            continue
+        if tok in base:
+            continue
+        if tok in ev:
+            return True
+    return False
+
+
+def evidence_leakage_heuristic(
+    body: str,
+    evidence_plain: str,
+    oral: str,
+    existing: str,
+    min_len: int,
+) -> bool:
+    """
+    True：正文与 evidence 的最长公共子串足够长，且该子串未出现在口述或已有正文中，
+    视为摘录渗漏，应回退安全正文。
+    """
+    body = (body or "").strip()
+    ev = (evidence_plain or "").strip()
+    if not body or not ev:
+        return False
+    lcs = longest_common_substring(body, ev)
+    if len(lcs) < min_len:
+        return False
+    o = oral or ""
+    ex = existing or ""
+    if lcs in o or lcs in ex:
+        return False
+    return True
+
+
+def strip_evidence_for_overlap_check(evidence_text: str) -> str:
+    """去掉 chunk 标记行等噪声，仅保留内容用于 overlap。"""
+    lines: list[str] = []
+    for line in (evidence_text or "").splitlines():
+        t = line.strip()
+        if t.startswith("[chunk_id="):
+            continue
+        if t.startswith("[摘要:"):
+            continue
+        lines.append(line)
+    return "\n".join(lines).strip()