框架: parse试拆三轮病理修复——ingest字段指纹改型+0卡不软删(防重跑自噬损耗)+parse_llm判重名录排除本章

.claude/skills/parse-book/scripts/parse_ingest.py

@@ -171,12 +171,15 @@ def patterns(work_id, chapter_order, file_):
         ch_id, src = chapter_of(conn, work_id, chapter_order)
         contracts = {t: field_contract(conn, t) for t in PATTERN_TYPES}
         book = conn.execute("SELECT title FROM muse_content_work WHERE id=%s", (work_id,)).fetchone()[0]
-        # 幂等：重跑本章 = 软删本章旧 draft
-        conn.execute(
-            """UPDATE muse_knowledge_draft SET deleted=TRUE, updater=%s
-               WHERE tenant_id=%s AND source_type='parse_book' AND source_id=%s
-                 AND draft_payload->>'章序'=%s AND deleted=FALSE""",
-            (ACTOR, TENANT, work_id, str(chapter_order)))
+        # 幂等：重跑本章 = 软删本章旧 draft。
+        # 但新轮 0 卡时保留旧卡——LLM 判卡有随机性，重跑「0 卡」不应清掉上轮已验证的产出
+        #（实测组合病：判重名录+随机性+软删 → 每次重跑都损耗卡）
+        if cards:
+            conn.execute(
+                """UPDATE muse_knowledge_draft SET deleted=TRUE, updater=%s
+                   WHERE tenant_id=%s AND source_type='parse_book' AND source_id=%s
+                     AND draft_payload->>'章序'=%s AND deleted=FALSE""",
+                (ACTOR, TENANT, work_id, str(chapter_order)))
         ok, rejected = 0, []
         for i, c in enumerate(cards):
             t = c.get("型")
@@ -189,10 +192,19 @@ def patterns(work_id, chapter_order, file_):
             if not (src_ref.get("书名") and src_ref.get("回目")):
                 reasons.append("出处不完整（需书名+回目）")
             fields = c.get("字段") or {}
-            if t in contracts:
+            retyped = None
+            if t in contracts and fields:
                 illegal = set(fields) - contracts[t]
                 if illegal:
-                    reasons.append(f"字段越合同:{sorted(illegal)}")
+                    # 字段指纹改型：LLM 的 type 标注不可靠（实测 M3 系统性全标 craft），
+                    # 五型字段 key 互不重叠——字段集合全命中唯一他型合同时，以字段为准改型
+                    fits = [t2 for t2 in PATTERN_TYPES
+                            if t2 != t and fields and set(fields) <= contracts[t2]]
+                    if len(fits) == 1:
+                        retyped = f"{t}→{fits[0]}(字段指纹改型)"
+                        c["型"] = t = fits[0]
+                    else:
+                        reasons.append(f"字段越合同:{sorted(illegal)}")
             texts = [c.get("名称"), c.get("一句话摘要"), *fields.values()]
             leak = leak_check([x for x in texts if x], src)
             if leak:
@@ -203,6 +215,8 @@ def patterns(work_id, chapter_order, file_):
             payload = {"型": t, "名称": c["名称"], "一句话摘要": c.get("一句话摘要", ""),
                        "字段": fields, "出处": src_ref, "目标库": "公共范式库",
                        "章序": chapter_order, "来源": f"拆书@{book}", "状态": "草稿"}
+            if retyped:
+                payload["改型"] = retyped  # 审计：机械改型可追溯
             cid = f"parse-{work_id}-{chapter_order}-{i}-" + hashlib.sha256(
                 json.dumps(payload, ensure_ascii=False, sort_keys=True).encode()).hexdigest()[:8]
             conn.execute(

.claude/skills/parse-book/scripts/parse_llm.py

@@ -167,10 +167,13 @@ def main(work_id, from_, to, model):
                    JOIN muse_content_chapter c ON c.id=s.chapter_id
                    WHERE s.tenant_id=%s AND s.work_id=%s AND c.order_no<%s AND s.deleted=FALSE""",
                 (TENANT, work_id, ch)).fetchall() for e in ents]
+            # 判重名录排除本章自己的卡：否则重跑时 M3 对着自己首轮的卡判重不出卡，
+            # ingest 幂等软删旧卡后插 0 张——卡被自己删了（重跑自噬,实测病）
             cards = conn.execute(
                 """SELECT draft_payload->>'型', draft_payload->>'名称' FROM muse_knowledge_draft
-                   WHERE tenant_id=%s AND source_type='parse_book' AND deleted=FALSE""",
-                (TENANT,)).fetchall()
+                   WHERE tenant_id=%s AND source_type='parse_book' AND deleted=FALSE
+                     AND NOT (source_id=%s AND (draft_payload->>'章序')::int=%s)""",
+                (TENANT, work_id, ch)).fetchall()
             outline = None
             if s_st == "done":
                 r = conn.execute(