@furongjun1999/dsh-memory 0.4.11 → 0.5.0

package/md_cg/test_p31_insight.py

@@ -1,534 +1,534 @@
-# -*- coding: utf-8 -*-
-"""洞察条件层 + 情景重构 + 归纳聚类 + 盲区学习（P31 · P2 工程缺口验收）。
-
-对照 docs §五 P2：新增 insight op（window/record/verify/list/report/
-reconstruct/learn/outlook）与 consolidate.induce。验收覆盖：
-  ① op 契约链：ALL_OPS == 工具 schema，insight 已授权且按 action 分级
-  ② 开窗条件：C1 ≥ 0.6 ∧ 跨域 ∧ 低压力，逐闸门给出阻塞项
-  ③ 洞见记账：record 缺正文报错 / 幂等 / pending；verify 无证据不判定
-  ④ 证据裁决：V1 不足保持 pending，V3 / 显式裁决落 verified|falsified，
-     verified 后重要度保底 0.9 且置保护位
-  ⑤ CER 报告：样本 < 20 显式 insufficient 不给结论；样本足则 CER + 2×SE
-  ⑥ 结构洞察：outlook 只做结构描述与差距定位，不判有效性
-  ⑦ 情景重构：线索不足 → blindspot（不编造）；锚点齐 → 合成共同条件空间；
-     apply 落 inferred scene 节点且幂等
-  ⑧ 归纳聚类：无共同条件默认拒绝生成；apply 生成概念节点 + 对称 inferred 边；
-     幂等；放宽 require_conditions 才允许无共同条件下成团
-  ⑨ 盲区学习：未知 blindspot 不编造；unknowable 不生成路线；
-     carried/unresolved 落 gap_hint 待补线索且幂等
-  ⑩ 权限分档：output 可读不可记；写层可记账不可批量落库；guest 整 op 被拒
-
-诚实纪律：归纳 / 重构产物一律标 inferred（非事实断言）；证据不足拒绝裁决或生成。
-独立临时根，重跑 ≡ 首跑。
-
-运行：python -m md_cg.test_p31_insight
-"""
-from __future__ import annotations
-
-import os
-import shutil
-import sys
-import tempfile
-
-from . import consolidate, insight, predict, subgraph, tokens
-from .mdcos import MdCGOS
-from .mcp_server import KERNEL_TOOLS, call_tool
-from .security import AccessDenied, Principal
-
-PASS = FAIL = 0
-FAILS = []
-
-# ---- 洞察条件层夹具 -------------------------------------------------------
-STMT_MAIN = "离线环境下应优先选择本地缓存实现，避免远程调用放大延迟"
-# C1–C8 全量条件（8 项齐 → missing 为空）
-C1_OK = {"retrievability": 0.8, "outside_observer": True,
-         "cross_domain": ["存储", "网络"], "premise_questioned": True,
-         "pressure": "low", "continuity_turns": 3, "externalized": True,
-         "tone": "curious"}
-C1_BAD = {"retrievability": 0.3, "cross_domain": ["存储"], "pressure": "high"}
-
-# ---- 归纳夹具：内容相近 + 共享「生效条件」 ---------------------------------
-IND_A = ("# 功能名：离线缓存策略甲\n# 生效条件：单机离线且无外部依赖\n"
-         "# 子功能：在无网络时选择本地缓存实现\n# 执行：读取本地配置并建立缓存池\n"
-         "# 不适用条件：分布式多租户在线服务\n")
-IND_B = ("# 功能名：离线缓存策略乙\n# 生效条件：单机离线且无外部依赖\n"
-         "# 子功能：在无网络时选择本地缓存实现\n# 执行：读取本地配置并建立缓存池\n"
-         "# 不适用条件：分布式多租户在线服务\n")
-IND_C = ("# 功能名：离线缓存策略丙\n# 生效条件：单机离线且无外部依赖\n"
-         "# 子功能：在无网络时选择本地缓存实现\n# 执行：读取本地配置并建立缓存池\n"
-         "# 不适用条件：分布式多租户在线服务\n")
-# 内容同样相近，但「生效条件」各不相同 → 无共同条件，默认必须拒绝归纳
-NO_A = ("# 功能名：日志采样器甲\n# 生效条件：机房甲内的访问流量\n"
-        "# 子功能：按固定比例采样访问日志\n# 执行：随机丢弃超出比例的记录\n"
-        "# 不适用条件：计费与审计场景\n")
-NO_B = ("# 功能名：日志采样器乙\n# 生效条件：机房乙内的访问流量\n"
-        "# 子功能：按固定比例采样访问日志\n# 执行：随机丢弃超出比例的记录\n"
-        "# 不适用条件：计费与审计场景\n")
-NO_C = ("# 功能名：日志采样器丙\n# 生效条件：机房丙内的访问流量\n"
-        "# 子功能：按固定比例采样访问日志\n# 执行：随机丢弃超出比例的记录\n"
-        "# 不适用条件：计费与审计场景\n")
-
-# ---- 情景重构夹具：两条记忆共享同一生效条件 -------------------------------
-REC_COMMON = "上下文接近上限且任务尚未完成"
-REC_A = ("# 功能名：压缩前先落盘检查点甲\n# 生效条件：%s\n"
-         "# 子功能：压缩前把会话状态写入检查点文件\n# 执行：写检查点后再压缩上下文\n"
-         "# 不适用条件：一次性问答任务\n") % REC_COMMON
-REC_B = ("# 功能名：压缩前先落盘检查点乙\n# 生效条件：%s\n"
-         "# 子功能：压缩前把会话状态写入检查点文件\n# 执行：写检查点后再压缩上下文\n"
-         "# 不适用条件：一次性问答任务\n") % REC_COMMON
-
-# ---- 盲区学习夹具 ---------------------------------------------------------
-LEARN_KB = ("# 功能名：离线缓存决策知识\n# 生效条件：单机离线环境\n"
-            "# 子功能：提供离线时的缓存选型判据\n# 执行：比对候选实现并给出结论\n"
-            "# 不适用条件：需要强一致分布式场景\n")
-LEARN_ANCHOR = ("# 功能名：离线缓存锚点条目 zkv7\n# 生效条件：单机离线环境\n"
-                "# 子功能：作为离线缓存话题的检索入口 zkv7\n"
-                "# 执行：指向离线缓存决策知识\n# 不适用条件：在线服务\n")
-LEARN_CTX = ("# 功能名：日志采样待验证终点\n# 生效条件：机房日志采样链路\n"
-             "# 子功能：尚未形成可判定结论\n# 执行：待补\n# 不适用条件：离线单机\n")
-LEARN_ANCHOR2 = ("# 功能名：日志采样链路锚点条目 zxc9\n# 生效条件：机房日志采样链路\n"
-                 "# 子功能：作为日志采样话题的检索入口 zxc9\n"
-                 "# 执行：指向待验证终点\n# 不适用条件：离线单机\n")
-
-
-def check(name, cond, detail=""):
-    global PASS, FAIL
-    if cond:
-        PASS += 1
-        print(f"  [PASS] {name}" + (f"  · {detail}" if detail else ""))
-    else:
-        FAIL += 1
-        FAILS.append(name)
-        print(f"  [FAIL] {name}  · {detail}")
-
-
-def _principal(role):
-    s = tokens.role_spec(role)
-    return Principal(tenant="default", actor=role, clearance="internal",
-                     can_write=bool(s["can_write"]), can_admin=bool(s["can_admin"]),
-                     role=role, layers_allow=s["layers_allow"],
-                     ops_allow=s["ops_allow"])
-
-
-def _denied(fn):
-    try:
-        fn()
-        return False
-    except AccessDenied:
-        return True
-
-
-def main():
-    print("=" * 68)
-    print("md 认知图 P31 验收 · 洞察/重构/归纳/盲区学习（P2）")
-    print("=" * 68)
-    tmp = tempfile.mkdtemp(prefix="mdcg_p31_")
-    ROOT = os.path.join(tmp, "root")
-    from . import corpus
-    corpus.reset_root(ROOT)
-    cg = MdCGOS(ROOT, actor="p31")
-    cg.principal = Principal(tenant="default", actor="p31", clearance="internal",
-                             can_write=True, can_admin=True, role="designer")
-    try:
-        # ---------------------------------------------- ① 契约链
-        print("\n【1】op 契约链与分级授权")
-        tool = next(t for t in KERNEL_TOOLS if t["name"] == "cg")
-        declared = set(tool["inputSchema"]["properties"]["op"]["description"].split("|"))
-        check("工具 op 枚举 == ALL_OPS（防漏改）", declared == set(tokens.ALL_OPS),
-              f"diff={sorted(declared ^ set(tokens.ALL_OPS))}")
-        check("insight 已进 ALL_OPS", "insight" in tokens.ALL_OPS)
-        check("INSIGHT_ACTIONS 覆盖 P2",
-              {"window", "record", "verify", "list", "report", "reconstruct",
-               "learn", "outlook"} <= set(cg.INSIGHT_ACTIONS),
-              str(cg.INSIGHT_ACTIONS))
-        check("insight 已授权给写层/反思/校验/输出/维持",
-              all("insight" in tokens.role_spec(r)["ops_allow"]
-                  for r in ("record", "reflect", "verify", "output", "sustain")))
-        check("guest 未获 insight",
-              "insight" not in tokens.role_spec("guest")["ops_allow"])
-        cat = call_tool(cg, "cg", {"op": "insight", "action": "catalog"})
-        check("catalog 自描述条件与证据类型",
-              cat.get("ok") and {"reconstruct", "learn"} <= set(cat["actions"])
-              and len(cat["conditions"]) == 8 and set(cat["v_types"]) == {"v1", "v2", "v3"},
-              str(cat.get("actions")))
-
-        # ---------------------------------------------- ② 归纳聚类（干净池先行）
-        print("\n【2】consolidate.induce 归纳聚类")
-        for nid, body in (("ind_1", IND_A), ("ind_2", IND_B), ("ind_3", IND_C),
-                          ("no_1", NO_A), ("no_2", NO_B), ("no_3", NO_C)):
-            cg.add(nid, body, layer="contextual", importance=0.4,
-                   verification_basis="other", actor="p31")
-        dry = call_tool(cg, "cg", {"op": "consolidate", "action": "induce",
-                                   "min_jaccard": 0.5})
-        check("induce 预演只出提案、不写盘",
-              dry.get("dry_run") is True and dry.get("written") == 0
-              and bool(dry.get("concepts"))
-              and all(str(c) not in cg.index["nodes"] for c in dry["concepts"]),
-              str(dry.get("concepts")))
-        ind_prop = [p for p in (dry.get("samples") or [])
-                    if set(p.get("members") or []) >= {"ind_1", "ind_2", "ind_3"}]
-        check("共同条件的相似记忆成团",
-              bool(ind_prop), f"clusters={dry.get('clusters')}")
-        check("提炼出共同生效条件",
-              bool(ind_prop) and ind_prop[0].get("common_conditions")
-              == ["%s" % "单机离线且无外部依赖"],
-              str(ind_prop[0].get("common_conditions")) if ind_prop else "-")
-        check("无共同条件默认拒绝生成（诚实纪律）",
-              int(dry.get("skipped_no_condition") or 0) >= 1
-              and not any(set(p.get("members") or []) & {"no_1", "no_2", "no_3"}
-                          for p in (dry.get("samples") or [])),
-              f"skipped_no_condition={dry.get('skipped_no_condition')}")
-
-        app = call_tool(cg, "cg", {"op": "consolidate", "action": "induce",
-                                   "min_jaccard": 0.5, "apply": True})
-        cid = (app.get("concepts") or [None])[0]
-        check("induce 落盘生成概念节点", app.get("written", 0) >= 1 and bool(cid),
-              str(cid))
-        cnode = (cg.get(cid) or {}) if cid else {}
-        cfm = cnode.get("frontmatter") or {}
-        check("概念节点标 inferred（非事实断言）",
-              cfm.get("verification_basis") == "other"
-              and {"concept", "induced"} <= set(cfm.get("tags") or []),
-              str(cfm.get("tags")))
-        check("概念节点记 induced_from 成员清单",
-              set(cfm.get("induced_from") or []) == {"ind_1", "ind_2", "ind_3"},
-              str(cfm.get("induced_from")))
-        etargets = {e.get("target") for e in (cfm.get("edges") or [])}
-        check("概念 → 成员写 generalizes 边",
-              {"ind_1", "ind_2", "ind_3"} <= etargets, str(sorted(etargets)))
-        mfm = (cg.get("ind_1") or {}).get("frontmatter") or {}
-        check("成员 → 概念写 instance_of 边",
-              mfm.get("induced_concept") == cid
-              and cid in {e.get("target") for e in (mfm.get("edges") or [])},
-              str(mfm.get("induced_concept")))
-        # ---- 2026-09-19 阶段一：巩固留痕字段一等公民 ----
-        check("成员侧记 consolidated_into（规范名·巩固进哪一条）",
-              mfm.get("consolidated_into") == cid
-              and mfm.get("induced_concept") == cid,
-              str(mfm.get("consolidated_into")))
-        check("成员侧记 consolidated_at（何时巩固）",
-              bool(mfm.get("consolidated_at")), str(mfm.get("consolidated_at")))
-        check("概念侧记 consolidated_at（概念形成即巩固时刻）",
-              bool(cfm.get("consolidated_at"))
-              and cfm.get("induced_at") == cfm.get("consolidated_at"),
-              f"{cfm.get('consolidated_at')} / {cfm.get('induced_at')}")
-        check("前身可定位：概念侧 induced_from 与成员侧列互查一致",
-              set(cfm.get("induced_from") or []) == {"ind_1", "ind_2", "ind_3"}
-              and all(((cg.get(m) or {}).get("frontmatter") or {}
-                       ).get("consolidated_into") == cid
-                      for m in ("ind_1", "ind_2", "ind_3")))
-        # add() 是**全量重建 fm**——覆写后留痕必须仍在（回读节点继承，非索引快照）
-        cg.add("ind_1", IND_A, layer="contextual", importance=0.4,
-               verification_basis="other", actor="p31")
-        mfm2 = (cg.get("ind_1") or {}).get("frontmatter") or {}
-        check("覆写后成员侧巩固留痕继承（consolidated_into/at 不丢）",
-              mfm2.get("consolidated_into") == cid
-              and bool(mfm2.get("consolidated_at")),
-              str(mfm2.get("consolidated_into")))
-        check("induce 留痕 _maintain.jsonl",
-              any(r.get("action") == "induce"
-                  for r in consolidate._read_maintain(ROOT)))
-
-        app2 = call_tool(cg, "cg", {"op": "consolidate", "action": "induce",
-                                    "min_jaccard": 0.5, "apply": True})
-        check("induce 幂等（重复 apply 不重写）",
-              app2.get("written") == 0 and int(app2.get("skipped_existing") or 0) >= 1,
-              f"written={app2.get('written')} skip={app2.get('skipped_existing')}")
-        relax = call_tool(cg, "cg", {"op": "consolidate", "action": "induce",
-                                     "min_jaccard": 0.5, "require_conditions": False})
-        check("放宽 require_conditions 才允许无共同条件成团",
-              any(set(p.get("members") or []) >= {"no_1", "no_2", "no_3"}
-                  for p in (relax.get("samples") or [])))
-
-        # ---------------------------------------------- ③ 开窗条件
-        print("\n【3】insight.window 开窗条件")
-        w_ok = call_tool(cg, "cg", {"op": "insight", "action": "window",
-                                    "conditions": C1_OK})
-        check("C1≥0.6 ∧ 跨域 ∧ 低压力 → 开窗",
-              w_ok.get("open") is True and w_ok.get("blocked_by") == [],
-              str(w_ok.get("gates")))
-        w_bad = call_tool(cg, "cg", {"op": "insight", "action": "window",
-                                     "conditions": C1_BAD})
-        check("条件不足 → 开窗被拒并给出阻塞项",
-              w_bad.get("open") is False
-              and {"c1_retrievable", "low_pressure"} <= set(w_bad.get("blocked_by") or []),
-              str(w_bad.get("blocked_by")))
-        check("开窗不宣称有效性（写进 note）",
-              "verify" in str(w_ok.get("note") or ""))
-
-        # ---------------------------------------------- ④ 情景重构
-        print("\n【4】insight.reconstruct 情景重构")
-        cg.add("rec_a", REC_A, layer="contextual", importance=0.5,
-               verification_basis="other", actor="p31")
-        cg.add("rec_b", REC_B, layer="contextual", importance=0.5,
-               verification_basis="other", actor="p31")
-        blind = call_tool(cg, "cg", {"op": "insight", "action": "reconstruct",
-                                     "clues": ["zqxwv yykb 0009137"]})
-        check("线索不足 → blindspot（不编造）",
-              blind.get("status") == "blindspot"
-              and not blind.get("condition_space"),
-              f"status={blind.get('status')}")
-        rec = call_tool(cg, "cg", {"op": "insight", "action": "reconstruct",
-                                   "ids": ["rec_a", "rec_b"]})
-        cs = rec.get("condition_space") or {}
-        check("锚点齐 → 合成共同条件空间",
-              rec.get("status") == "reconstructed"
-              and REC_COMMON in (cs.get("common") or []),
-              str(cs.get("common")))
-        check("重构声明为近似重建（非事件回放）",
-              "非事件回放" in str(rec.get("note") or "")
-              and rec.get("confidence") is not None,
-              str(rec.get("note"))[:80])
-        rec_w = call_tool(cg, "cg", {"op": "insight", "action": "reconstruct",
-                                     "ids": ["rec_a", "rec_b"], "apply": True})
-        sid = rec_w.get("scene_id")
-        sfm = ((cg.get(sid) or {}).get("frontmatter") or {}) if sid else {}
-        check("apply 落 scene 节点（inferred，非事实层）",
-              rec_w.get("scene_written") is True and bool(sid)
-              and sfm.get("layer") == "contextual"
-              and {"scene", "reconstructed"} <= set(sfm.get("tags") or []),
-              f"sid={sid} tags={sfm.get('tags')}")
-        check("scene 记 reconstructed_from",
-              set(sfm.get("reconstructed_from") or []) == {"rec_a", "rec_b"},
-              str(sfm.get("reconstructed_from")))
-        rec_w2 = call_tool(cg, "cg", {"op": "insight", "action": "reconstruct",
-                                      "ids": ["rec_a", "rec_b"], "apply": True})
-        check("重构幂等（同一批锚点不重写）",
-              rec_w2.get("scene_written") is False and rec_w2.get("scene_id") == sid,
-              f"sid={rec_w2.get('scene_id')}")
-
-        # ---------------------------------------------- ⑤ 盲区学习
-        print("\n【5】insight.learn 盲区学习闭环")
-        cg.add("learn_kb", LEARN_KB, layer="knowledge", importance=0.7,
-               verification_basis="data", actor="p31")
-        cg.add("learn_anchor", LEARN_ANCHOR, layer="contextual", importance=0.5,
-               verification_basis="other", actor="p31",
-               edges=[{"target": "learn_kb", "relation_type": "causal",
-                       "confidence": 0.9, "evidence": "declared"}])
-        cg.add("learn_anchor2", LEARN_ANCHOR2, layer="contextual", importance=0.5,
-               verification_basis="other", actor="p31",
-               edges=[{"target": "learn_ctx", "relation_type": "causal",
-                       "confidence": 0.9, "evidence": "declared"}])
-        cg.add("learn_ctx", LEARN_CTX, layer="contextual", importance=0.4,
-               verification_basis="other", actor="p31")
-        cg.add("bs_resolved", "# 可预测性：predictable\n离线缓存锚点条目 zkv7",
-               layer="unresolved", importance=0.3, actor="p31")
-        cg.add("bs_unknowable", "# 可预测性：unknowable\n外部世界的结构性事实",
-               layer="unresolved", importance=0.3, actor="p31")
-        cg.add("bs_carried", "# 可预测性：predictable\n日志采样链路锚点条目 zxc9",
-               layer="unresolved", importance=0.3, actor="p31")
-        check("盲区描述 → 锚点（检索打分）",
-              predict.anchor_from_description(cg, "离线缓存锚点条目 zkv7") == "learn_anchor"
-              and predict.anchor_from_description(cg, "日志采样链路锚点条目 zxc9")
-              == "learn_anchor2")
-        nf = call_tool(cg, "cg", {"op": "insight", "action": "learn",
-                                  "blindspot_id": "no_such_blindspot"})
-        check("未知盲区不编造", nf.get("ok") is False and nf.get("status") == "not_found",
-              str(nf.get("reason"))[:60])
-        lr = call_tool(cg, "cg", {"op": "insight", "action": "learn"})
-        steps = {s["blindspot_id"]: s for s in (lr.get("steps") or [])}
-        check("未解问题进入学习闭环",
-              {"bs_resolved", "bs_unknowable", "bs_carried"} <= set(steps),
-              str(sorted(steps)))
-        check("终态五态可判定且不越集",
-              all(s.get("terminal") in ("unknowable", "no_anchor", "unresolved",
-                                        "carried", "resolved") for s in steps.values()))
-        check("可判定终点 → resolved",
-              steps.get("bs_resolved", {}).get("terminal") == "resolved"
-              and steps["bs_resolved"].get("terminal_node") == "learn_kb",
-              str(steps.get("bs_resolved", {}).get("terminal")))
-        check("unknowable 不生成路线（拒绝编造）",
-              steps.get("bs_unknowable", {}).get("terminal") == "unknowable"
-              and steps["bs_unknowable"].get("routes") == 0)
-        check("未达可判定终点 → carried",
-              steps.get("bs_carried", {}).get("terminal") in ("carried", "unresolved"),
-              str(steps.get("bs_carried", {}).get("terminal")))
-        lw = call_tool(cg, "cg", {"op": "insight", "action": "learn", "apply": True})
-        gap = [nid for nid in lw.get("written") or []]
-        check("carried/unresolved 落 gap_hint 待补线索",
-              len(gap) == int(lw["summary"].get("carried", 0))
-              + int(lw["summary"].get("unresolved", 0)) and len(gap) >= 1,
-              f"written={gap}")
-        gnode = (cg.get(gap[0]) or {}) if gap else {}
-        gfm = gnode.get("frontmatter") or {}
-        check("gap_hint 是线索而非事实（contextual + 显式非事实标注）",
-              gfm.get("layer") == "contextual"
-              and "gap_hint" in (gfm.get("tags") or [])
-              and "非既有事实" in (gnode.get("content") or ""),
-              str(gfm.get("tags")))
-        # 回归防线：线索节点逐字抄了盲区描述，若不隔离会顶掉真实锚点，
-        # 使「线索 → 0 路线 → 永远 unresolved」并让重复 learn 不再幂等。
-        check("线索节点不劫持锚点解析（自我污染防护）",
-              predict.anchor_from_description(
-                  cg, (predict.find_blindspot(cg, "bs_resolved") or {}).get("description") or "")
-              == "learn_anchor")
-        lw2 = call_tool(cg, "cg", {"op": "insight", "action": "learn", "apply": True})
-        check("盲区学习落库幂等（重复 apply 不新增线索）",
-              lw2.get("written") == []
-              and lw2["summary"] == lw["summary"], str(lw2.get("written")))
-
-        # ---------------------------------------------- ⑥ 洞见记账与裁决
-        print("\n【6】insight.record / verify")
-        bad = None
-        try:
-            call_tool(cg, "cg", {"op": "insight", "action": "record"})
-        except Exception as exc:                       # noqa: BLE001
-            bad = exc
-        check("record 缺正文直接报错（不臆造）", isinstance(bad, ValueError),
-              type(bad).__name__)
-        r1 = call_tool(cg, "cg", {"op": "insight", "action": "record",
-                                  "statement": STMT_MAIN, "conditions": C1_OK,
-                                  "category": "caching"})
-        nid = r1.get("node_id")
-        check("record 记为 pending 并快照条件",
-              r1.get("state") == "pending" and r1.get("duplicate") is False
-              and r1.get("window", {}).get("open") is True, str(nid))
-        rec_fm = (cg.get(nid) or {}).get("frontmatter") or {}
-        check("条件快照含 C1–C8 且无缺失",
-              set(rec_fm.get("insight_conditions") or {}) == set(insight.CONDITION_KEYS)
-              and rec_fm.get("insight_missing") == [], str(rec_fm.get("insight_missing")))
-        r1b = call_tool(cg, "cg", {"op": "insight", "action": "record",
-                                   "statement": STMT_MAIN, "conditions": C1_OK})
-        check("record 同正文幂等", r1b.get("duplicate") is True
-              and r1b.get("node_id") == nid)
-        v0 = call_tool(cg, "cg", {"op": "insight", "action": "verify", "node_id": nid})
-        check("无证据不判定（保持 pending）",
-              v0.get("decided") is False and v0.get("state") == "pending",
-              str(v0.get("reason"))[:50])
-        v1 = call_tool(cg, "cg", {"op": "insight", "action": "verify",
-                                  "node_id": nid, "v_types": ["v1"]})
-        check("V1 不足门槛不判定", v1.get("decided") is False)
-        v3 = call_tool(cg, "cg", {"op": "insight", "action": "verify",
-                                  "node_id": nid, "v_types": ["v3"]})
-        check("V3 外部确证 → verified + 重要度保底 + 保护位",
-              v3.get("state") == "verified" and v3.get("importance") >= 0.9
-              and v3.get("protected") is True, str(v3.get("reason")))
-        v3fm = (cg.get(nid) or {}).get("frontmatter") or {}
-        check("verified 后基底升为 data 且带 verified 标签",
-              v3fm.get("verification_basis") == "data"
-              and insight.TAG_VERIFIED in (v3fm.get("tags") or []))
-        bad_v = None
-        try:
-            call_tool(cg, "cg", {"op": "insight", "action": "verify",
-                                 "node_id": nid, "verdict": "maybe"})
-        except Exception as exc:                       # noqa: BLE001
-            bad_v = exc
-        check("非法 verdict 报错", isinstance(bad_v, ValueError))
-        r2 = call_tool(cg, "cg", {"op": "insight", "action": "record",
-                                  "statement": "被外部证据证伪的洞见：所有缓存都应常驻内存"})
-        vf = call_tool(cg, "cg", {"op": "insight", "action": "verify",
-                                  "node_id": r2["node_id"], "verdict": "falsified"})
-        check("显式证伪 → falsified",
-              vf.get("state") == "falsified" and vf.get("decided") is True)
-        lst = call_tool(cg, "cg", {"op": "insight", "action": "list",
-                                   "state": "verified"})
-        check("list 按状态过滤",
-              lst.get("count", 0) >= 1
-              and all(e["state"] == "verified" for e in lst.get("events") or []))
-
-        # ---------------------------------------------- ⑦ CER 报告
-        print("\n【7】insight.report CER（样本不足不判定）")
-        rep0 = call_tool(cg, "cg", {"op": "insight", "action": "report"})
-        check("样本不足 → insufficient 且不给结论",
-              rep0.get("layer_state") == "insufficient" and rep0.get("cer") is None
-              and rep0.get("decided") < rep0.get("sample_min"), str(rep0.get("decided")))
-        for i in range(20):
-            ri = call_tool(cg, "cg", {"op": "insight", "action": "record",
-                                      "statement": f"P31 样本 {i}：需在离线条件下验证的分支 {i}",
-                                      "conditions": C1_OK})
-            call_tool(cg, "cg", {"op": "insight", "action": "verify",
-                                 "node_id": ri["node_id"],
-                                 "verdict": ("verified" if i < 16 else "falsified")})
-        rep = call_tool(cg, "cg", {"op": "insight", "action": "report"})
-        import math as _m
-        exp_v, exp_f = 17, 5
-        n = exp_v + exp_f
-        p = exp_v / n
-        se = _m.sqrt(p * (1 - p) / n)
-        check("样本足 → CER = verified/(verified+falsified)",
-              rep.get("decided") == n and abs(rep.get("cer") - round(p, 4)) < 1e-4
-              and rep.get("verified") == exp_v and rep.get("falsified") == exp_f,
-              f"cer={rep.get('cer')} decided={rep.get('decided')}")
-        check("2×SE 显著性带",
-              abs(rep.get("two_se") - round(2 * se, 4)) < 1e-4
-              and rep.get("significant") is True, f"two_se={rep.get('two_se')}")
-        check("合格样本 → layer_state=reliable",
-              rep.get("layer_state") == "reliable", str(rep.get("layer_state")))
-
-        # ---------------------------------------------- ⑧ 结构洞察
-        print("\n【8】insight.outlook 结构洞察")
-        ol = call_tool(cg, "cg", {"op": "insight", "action": "outlook"})
-        check("D1 结构判断含分层/分位/保护率",
-              ol.get("d1", {}).get("total", 0) > 0
-              and "layers" in ol["d1"] and "importance" in ol["d1"]
-              and "protected_rate" in ol["d1"], str(ol.get("d1", {}).get("total")))
-        check("D2 趋势含洞见 CER 与 unresolved",
-              ol.get("d2", {}).get("insight", {}).get("recorded", 0) >= 22
-              and "unresolved_nodes" in ol["d2"], str(ol.get("d2", {}).get("insight")))
-        check("建议可执行（带证据与建议调用）",
-              isinstance(ol.get("suggestions"), list)
-              and all({"priority", "action", "evidence", "call"} <= set(s)
-                      for s in ol["suggestions"]), f"n={len(ol.get('suggestions') or [])}")
-        check("outlook 不判有效性（写进 note）",
-              "不判有效性" in str(ol.get("note") or ""))
-
-        # ---------------------------------------------- ⑨ 权限分档
-        print("\n【9】insight 按 action 分级授权")
-        cg.principal = _principal("output")
-        check("output 可读洞察",
-              not _denied(lambda: call_tool(cg, "cg", {"op": "insight",
-                                                       "action": "outlook"})))
-        check("output 不可记洞见",
-              _denied(lambda: call_tool(cg, "cg", {"op": "insight",
-                                                   "action": "record",
-                                                   "statement": "越权记录"})))
-        check("output 不可批量重构落库",
-              _denied(lambda: call_tool(cg, "cg", {"op": "insight",
-                                                   "action": "reconstruct",
-                                                   "ids": ["rec_a"], "apply": True})))
-        cg.principal = _principal("record")
-        check("写层可记可裁",
-              not _denied(lambda: call_tool(cg, "cg", {"op": "insight",
-                                                       "action": "record",
-                                                       "statement": "写层记录"}))
-              and not _denied(lambda: call_tool(cg, "cg", {"op": "insight",
-                                                           "action": "list"})))
-        check("写层不可批量学习落库",
-              _denied(lambda: call_tool(cg, "cg", {"op": "insight",
-                                                   "action": "learn", "apply": True})))
-        cg.principal = _principal("guest")
-        check("guest 整 op 被拒",
-              _denied(lambda: call_tool(cg, "cg", {"op": "insight",
-                                                   "action": "outlook"})))
-        cg.principal = Principal(tenant="default", actor="p31", clearance="internal",
-                                 can_write=True, can_admin=True, role="designer")
-
-        # ---------------------------------------------- ⑩ 未知 action
-        print("\n【10】未知 action 不静默成功")
-        for op in ("insight", "consolidate"):
-            b = None
-            try:
-                call_tool(cg, "cg", {"op": op, "action": "no_such_action"})
-            except Exception as exc:                   # noqa: BLE001
-                b = exc
-            check(f"{op} 未知 action 抛错", b is not None, type(b).__name__)
-
-    finally:
-        try:
-            cg.close()
-        except Exception:                              # noqa: BLE001
-            pass
-        shutil.rmtree(tmp, ignore_errors=True)
-
-    print("\n" + "=" * 68)
-    print(f"结果：{PASS} 通过 / {FAIL} 失败")
-    if FAILS:
-        print("失败项：")
-        for n in FAILS:
-            print(f"  - {n}")
-    print("=" * 68)
-    return 0 if FAIL == 0 else 1
-
-
-if __name__ == "__main__":
-    sys.exit(main())
+# -*- coding: utf-8 -*-
+"""洞察条件层 + 情景重构 + 归纳聚类 + 盲区学习（P31 · P2 工程缺口验收）。
+
+对照 docs §五 P2：新增 insight op（window/record/verify/list/report/
+reconstruct/learn/outlook）与 consolidate.induce。验收覆盖：
+  ① op 契约链：ALL_OPS == 工具 schema，insight 已授权且按 action 分级
+  ② 开窗条件：C1 ≥ 0.6 ∧ 跨域 ∧ 低压力，逐闸门给出阻塞项
+  ③ 洞见记账：record 缺正文报错 / 幂等 / pending；verify 无证据不判定
+  ④ 证据裁决：V1 不足保持 pending，V3 / 显式裁决落 verified|falsified，
+     verified 后重要度保底 0.9 且置保护位
+  ⑤ CER 报告：样本 < 20 显式 insufficient 不给结论；样本足则 CER + 2×SE
+  ⑥ 结构洞察：outlook 只做结构描述与差距定位，不判有效性
+  ⑦ 情景重构：线索不足 → blindspot（不编造）；锚点齐 → 合成共同条件空间；
+     apply 落 inferred scene 节点且幂等
+  ⑧ 归纳聚类：无共同条件默认拒绝生成；apply 生成概念节点 + 对称 inferred 边；
+     幂等；放宽 require_conditions 才允许无共同条件下成团
+  ⑨ 盲区学习：未知 blindspot 不编造；unknowable 不生成路线；
+     carried/unresolved 落 gap_hint 待补线索且幂等
+  ⑩ 权限分档：output 可读不可记；写层可记账不可批量落库；guest 整 op 被拒
+
+诚实纪律：归纳 / 重构产物一律标 inferred（非事实断言）；证据不足拒绝裁决或生成。
+独立临时根，重跑 ≡ 首跑。
+
+运行：python -m md_cg.test_p31_insight
+"""
+from __future__ import annotations
+
+import os
+import shutil
+import sys
+import tempfile
+
+from . import consolidate, insight, predict, subgraph, tokens
+from .mdcos import MdCGOS
+from .mcp_server import KERNEL_TOOLS, call_tool
+from .security import AccessDenied, Principal
+
+PASS = FAIL = 0
+FAILS = []
+
+# ---- 洞察条件层夹具 -------------------------------------------------------
+STMT_MAIN = "离线环境下应优先选择本地缓存实现，避免远程调用放大延迟"
+# C1–C8 全量条件（8 项齐 → missing 为空）
+C1_OK = {"retrievability": 0.8, "outside_observer": True,
+         "cross_domain": ["存储", "网络"], "premise_questioned": True,
+         "pressure": "low", "continuity_turns": 3, "externalized": True,
+         "tone": "curious"}
+C1_BAD = {"retrievability": 0.3, "cross_domain": ["存储"], "pressure": "high"}
+
+# ---- 归纳夹具：内容相近 + 共享「生效条件」 ---------------------------------
+IND_A = ("# 功能名：离线缓存策略甲\n# 生效条件：单机离线且无外部依赖\n"
+         "# 子功能：在无网络时选择本地缓存实现\n# 执行：读取本地配置并建立缓存池\n"
+         "# 不适用条件：分布式多租户在线服务\n")
+IND_B = ("# 功能名：离线缓存策略乙\n# 生效条件：单机离线且无外部依赖\n"
+         "# 子功能：在无网络时选择本地缓存实现\n# 执行：读取本地配置并建立缓存池\n"
+         "# 不适用条件：分布式多租户在线服务\n")
+IND_C = ("# 功能名：离线缓存策略丙\n# 生效条件：单机离线且无外部依赖\n"
+         "# 子功能：在无网络时选择本地缓存实现\n# 执行：读取本地配置并建立缓存池\n"
+         "# 不适用条件：分布式多租户在线服务\n")
+# 内容同样相近，但「生效条件」各不相同 → 无共同条件，默认必须拒绝归纳
+NO_A = ("# 功能名：日志采样器甲\n# 生效条件：机房甲内的访问流量\n"
+        "# 子功能：按固定比例采样访问日志\n# 执行：随机丢弃超出比例的记录\n"
+        "# 不适用条件：计费与审计场景\n")
+NO_B = ("# 功能名：日志采样器乙\n# 生效条件：机房乙内的访问流量\n"
+        "# 子功能：按固定比例采样访问日志\n# 执行：随机丢弃超出比例的记录\n"
+        "# 不适用条件：计费与审计场景\n")
+NO_C = ("# 功能名：日志采样器丙\n# 生效条件：机房丙内的访问流量\n"
+        "# 子功能：按固定比例采样访问日志\n# 执行：随机丢弃超出比例的记录\n"
+        "# 不适用条件：计费与审计场景\n")
+
+# ---- 情景重构夹具：两条记忆共享同一生效条件 -------------------------------
+REC_COMMON = "上下文接近上限且任务尚未完成"
+REC_A = ("# 功能名：压缩前先落盘检查点甲\n# 生效条件：%s\n"
+         "# 子功能：压缩前把会话状态写入检查点文件\n# 执行：写检查点后再压缩上下文\n"
+         "# 不适用条件：一次性问答任务\n") % REC_COMMON
+REC_B = ("# 功能名：压缩前先落盘检查点乙\n# 生效条件：%s\n"
+         "# 子功能：压缩前把会话状态写入检查点文件\n# 执行：写检查点后再压缩上下文\n"
+         "# 不适用条件：一次性问答任务\n") % REC_COMMON
+
+# ---- 盲区学习夹具 ---------------------------------------------------------
+LEARN_KB = ("# 功能名：离线缓存决策知识\n# 生效条件：单机离线环境\n"
+            "# 子功能：提供离线时的缓存选型判据\n# 执行：比对候选实现并给出结论\n"
+            "# 不适用条件：需要强一致分布式场景\n")
+LEARN_ANCHOR = ("# 功能名：离线缓存锚点条目 zkv7\n# 生效条件：单机离线环境\n"
+                "# 子功能：作为离线缓存话题的检索入口 zkv7\n"
+                "# 执行：指向离线缓存决策知识\n# 不适用条件：在线服务\n")
+LEARN_CTX = ("# 功能名：日志采样待验证终点\n# 生效条件：机房日志采样链路\n"
+             "# 子功能：尚未形成可判定结论\n# 执行：待补\n# 不适用条件：离线单机\n")
+LEARN_ANCHOR2 = ("# 功能名：日志采样链路锚点条目 zxc9\n# 生效条件：机房日志采样链路\n"
+                 "# 子功能：作为日志采样话题的检索入口 zxc9\n"
+                 "# 执行：指向待验证终点\n# 不适用条件：离线单机\n")
+
+
+def check(name, cond, detail=""):
+    global PASS, FAIL
+    if cond:
+        PASS += 1
+        print(f"  [PASS] {name}" + (f"  · {detail}" if detail else ""))
+    else:
+        FAIL += 1
+        FAILS.append(name)
+        print(f"  [FAIL] {name}  · {detail}")
+
+
+def _principal(role):
+    s = tokens.role_spec(role)
+    return Principal(tenant="default", actor=role, clearance="internal",
+                     can_write=bool(s["can_write"]), can_admin=bool(s["can_admin"]),
+                     role=role, layers_allow=s["layers_allow"],
+                     ops_allow=s["ops_allow"])
+
+
+def _denied(fn):
+    try:
+        fn()
+        return False
+    except AccessDenied:
+        return True
+
+
+def main():
+    print("=" * 68)
+    print("md 认知图 P31 验收 · 洞察/重构/归纳/盲区学习（P2）")
+    print("=" * 68)
+    tmp = tempfile.mkdtemp(prefix="mdcg_p31_")
+    ROOT = os.path.join(tmp, "root")
+    from . import corpus
+    corpus.reset_root(ROOT)
+    cg = MdCGOS(ROOT, actor="p31")
+    cg.principal = Principal(tenant="default", actor="p31", clearance="internal",
+                             can_write=True, can_admin=True, role="designer")
+    try:
+        # ---------------------------------------------- ① 契约链
+        print("\n【1】op 契约链与分级授权")
+        tool = next(t for t in KERNEL_TOOLS if t["name"] == "cg")
+        declared = set(tool["inputSchema"]["properties"]["op"]["description"].split("|"))
+        check("工具 op 枚举 == ALL_OPS（防漏改）", declared == set(tokens.ALL_OPS),
+              f"diff={sorted(declared ^ set(tokens.ALL_OPS))}")
+        check("insight 已进 ALL_OPS", "insight" in tokens.ALL_OPS)
+        check("INSIGHT_ACTIONS 覆盖 P2",
+              {"window", "record", "verify", "list", "report", "reconstruct",
+               "learn", "outlook"} <= set(cg.INSIGHT_ACTIONS),
+              str(cg.INSIGHT_ACTIONS))
+        check("insight 已授权给写层/反思/校验/输出/维持",
+              all("insight" in tokens.role_spec(r)["ops_allow"]
+                  for r in ("record", "reflect", "verify", "output", "sustain")))
+        check("guest 未获 insight",
+              "insight" not in tokens.role_spec("guest")["ops_allow"])
+        cat = call_tool(cg, "cg", {"op": "insight", "action": "catalog"})
+        check("catalog 自描述条件与证据类型",
+              cat.get("ok") and {"reconstruct", "learn"} <= set(cat["actions"])
+              and len(cat["conditions"]) == 8 and set(cat["v_types"]) == {"v1", "v2", "v3"},
+              str(cat.get("actions")))
+
+        # ---------------------------------------------- ② 归纳聚类（干净池先行）
+        print("\n【2】consolidate.induce 归纳聚类")
+        for nid, body in (("ind_1", IND_A), ("ind_2", IND_B), ("ind_3", IND_C),
+                          ("no_1", NO_A), ("no_2", NO_B), ("no_3", NO_C)):
+            cg.add(nid, body, layer="contextual", importance=0.4,
+                   verification_basis="other", actor="p31")
+        dry = call_tool(cg, "cg", {"op": "consolidate", "action": "induce",
+                                   "min_jaccard": 0.5})
+        check("induce 预演只出提案、不写盘",
+              dry.get("dry_run") is True and dry.get("written") == 0
+              and bool(dry.get("concepts"))
+              and all(str(c) not in cg.index["nodes"] for c in dry["concepts"]),
+              str(dry.get("concepts")))
+        ind_prop = [p for p in (dry.get("samples") or [])
+                    if set(p.get("members") or []) >= {"ind_1", "ind_2", "ind_3"}]
+        check("共同条件的相似记忆成团",
+              bool(ind_prop), f"clusters={dry.get('clusters')}")
+        check("提炼出共同生效条件",
+              bool(ind_prop) and ind_prop[0].get("common_conditions")
+              == ["%s" % "单机离线且无外部依赖"],
+              str(ind_prop[0].get("common_conditions")) if ind_prop else "-")
+        check("无共同条件默认拒绝生成（诚实纪律）",
+              int(dry.get("skipped_no_condition") or 0) >= 1
+              and not any(set(p.get("members") or []) & {"no_1", "no_2", "no_3"}
+                          for p in (dry.get("samples") or [])),
+              f"skipped_no_condition={dry.get('skipped_no_condition')}")
+
+        app = call_tool(cg, "cg", {"op": "consolidate", "action": "induce",
+                                   "min_jaccard": 0.5, "apply": True})
+        cid = (app.get("concepts") or [None])[0]
+        check("induce 落盘生成概念节点", app.get("written", 0) >= 1 and bool(cid),
+              str(cid))
+        cnode = (cg.get(cid) or {}) if cid else {}
+        cfm = cnode.get("frontmatter") or {}
+        check("概念节点标 inferred（非事实断言）",
+              cfm.get("verification_basis") == "other"
+              and {"concept", "induced"} <= set(cfm.get("tags") or []),
+              str(cfm.get("tags")))
+        check("概念节点记 induced_from 成员清单",
+              set(cfm.get("induced_from") or []) == {"ind_1", "ind_2", "ind_3"},
+              str(cfm.get("induced_from")))
+        etargets = {e.get("target") for e in (cfm.get("edges") or [])}
+        check("概念 → 成员写 generalizes 边",
+              {"ind_1", "ind_2", "ind_3"} <= etargets, str(sorted(etargets)))
+        mfm = (cg.get("ind_1") or {}).get("frontmatter") or {}
+        check("成员 → 概念写 instance_of 边",
+              mfm.get("induced_concept") == cid
+              and cid in {e.get("target") for e in (mfm.get("edges") or [])},
+              str(mfm.get("induced_concept")))
+        # ---- 2026-09-19 阶段一：巩固留痕字段一等公民 ----
+        check("成员侧记 consolidated_into（规范名·巩固进哪一条）",
+              mfm.get("consolidated_into") == cid
+              and mfm.get("induced_concept") == cid,
+              str(mfm.get("consolidated_into")))
+        check("成员侧记 consolidated_at（何时巩固）",
+              bool(mfm.get("consolidated_at")), str(mfm.get("consolidated_at")))
+        check("概念侧记 consolidated_at（概念形成即巩固时刻）",
+              bool(cfm.get("consolidated_at"))
+              and cfm.get("induced_at") == cfm.get("consolidated_at"),
+              f"{cfm.get('consolidated_at')} / {cfm.get('induced_at')}")
+        check("前身可定位：概念侧 induced_from 与成员侧列互查一致",
+              set(cfm.get("induced_from") or []) == {"ind_1", "ind_2", "ind_3"}
+              and all(((cg.get(m) or {}).get("frontmatter") or {}
+                       ).get("consolidated_into") == cid
+                      for m in ("ind_1", "ind_2", "ind_3")))
+        # add() 是**全量重建 fm**——覆写后留痕必须仍在（回读节点继承，非索引快照）
+        cg.add("ind_1", IND_A, layer="contextual", importance=0.4,
+               verification_basis="other", actor="p31")
+        mfm2 = (cg.get("ind_1") or {}).get("frontmatter") or {}
+        check("覆写后成员侧巩固留痕继承（consolidated_into/at 不丢）",
+              mfm2.get("consolidated_into") == cid
+              and bool(mfm2.get("consolidated_at")),
+              str(mfm2.get("consolidated_into")))
+        check("induce 留痕 _maintain.jsonl",
+              any(r.get("action") == "induce"
+                  for r in consolidate._read_maintain(ROOT)))
+
+        app2 = call_tool(cg, "cg", {"op": "consolidate", "action": "induce",
+                                    "min_jaccard": 0.5, "apply": True})
+        check("induce 幂等（重复 apply 不重写）",
+              app2.get("written") == 0 and int(app2.get("skipped_existing") or 0) >= 1,
+              f"written={app2.get('written')} skip={app2.get('skipped_existing')}")
+        relax = call_tool(cg, "cg", {"op": "consolidate", "action": "induce",
+                                     "min_jaccard": 0.5, "require_conditions": False})
+        check("放宽 require_conditions 才允许无共同条件成团",
+              any(set(p.get("members") or []) >= {"no_1", "no_2", "no_3"}
+                  for p in (relax.get("samples") or [])))
+
+        # ---------------------------------------------- ③ 开窗条件
+        print("\n【3】insight.window 开窗条件")
+        w_ok = call_tool(cg, "cg", {"op": "insight", "action": "window",
+                                    "conditions": C1_OK})
+        check("C1≥0.6 ∧ 跨域 ∧ 低压力 → 开窗",
+              w_ok.get("open") is True and w_ok.get("blocked_by") == [],
+              str(w_ok.get("gates")))
+        w_bad = call_tool(cg, "cg", {"op": "insight", "action": "window",
+                                     "conditions": C1_BAD})
+        check("条件不足 → 开窗被拒并给出阻塞项",
+              w_bad.get("open") is False
+              and {"c1_retrievable", "low_pressure"} <= set(w_bad.get("blocked_by") or []),
+              str(w_bad.get("blocked_by")))
+        check("开窗不宣称有效性（写进 note）",
+              "verify" in str(w_ok.get("note") or ""))
+
+        # ---------------------------------------------- ④ 情景重构
+        print("\n【4】insight.reconstruct 情景重构")
+        cg.add("rec_a", REC_A, layer="contextual", importance=0.5,
+               verification_basis="other", actor="p31")
+        cg.add("rec_b", REC_B, layer="contextual", importance=0.5,
+               verification_basis="other", actor="p31")
+        blind = call_tool(cg, "cg", {"op": "insight", "action": "reconstruct",
+                                     "clues": ["zqxwv yykb 0009137"]})
+        check("线索不足 → blindspot（不编造）",
+              blind.get("status") == "blindspot"
+              and not blind.get("condition_space"),
+              f"status={blind.get('status')}")
+        rec = call_tool(cg, "cg", {"op": "insight", "action": "reconstruct",
+                                   "ids": ["rec_a", "rec_b"]})
+        cs = rec.get("condition_space") or {}
+        check("锚点齐 → 合成共同条件空间",
+              rec.get("status") == "reconstructed"
+              and REC_COMMON in (cs.get("common") or []),
+              str(cs.get("common")))
+        check("重构声明为近似重建（非事件回放）",
+              "非事件回放" in str(rec.get("note") or "")
+              and rec.get("confidence") is not None,
+              str(rec.get("note"))[:80])
+        rec_w = call_tool(cg, "cg", {"op": "insight", "action": "reconstruct",
+                                     "ids": ["rec_a", "rec_b"], "apply": True})
+        sid = rec_w.get("scene_id")
+        sfm = ((cg.get(sid) or {}).get("frontmatter") or {}) if sid else {}
+        check("apply 落 scene 节点（inferred，非事实层）",
+              rec_w.get("scene_written") is True and bool(sid)
+              and sfm.get("layer") == "contextual"
+              and {"scene", "reconstructed"} <= set(sfm.get("tags") or []),
+              f"sid={sid} tags={sfm.get('tags')}")
+        check("scene 记 reconstructed_from",
+              set(sfm.get("reconstructed_from") or []) == {"rec_a", "rec_b"},
+              str(sfm.get("reconstructed_from")))
+        rec_w2 = call_tool(cg, "cg", {"op": "insight", "action": "reconstruct",
+                                      "ids": ["rec_a", "rec_b"], "apply": True})
+        check("重构幂等（同一批锚点不重写）",
+              rec_w2.get("scene_written") is False and rec_w2.get("scene_id") == sid,
+              f"sid={rec_w2.get('scene_id')}")
+
+        # ---------------------------------------------- ⑤ 盲区学习
+        print("\n【5】insight.learn 盲区学习闭环")
+        cg.add("learn_kb", LEARN_KB, layer="knowledge", importance=0.7,
+               verification_basis="data", actor="p31")
+        cg.add("learn_anchor", LEARN_ANCHOR, layer="contextual", importance=0.5,
+               verification_basis="other", actor="p31",
+               edges=[{"target": "learn_kb", "relation_type": "causal",
+                       "confidence": 0.9, "evidence": "declared"}])
+        cg.add("learn_anchor2", LEARN_ANCHOR2, layer="contextual", importance=0.5,
+               verification_basis="other", actor="p31",
+               edges=[{"target": "learn_ctx", "relation_type": "causal",
+                       "confidence": 0.9, "evidence": "declared"}])
+        cg.add("learn_ctx", LEARN_CTX, layer="contextual", importance=0.4,
+               verification_basis="other", actor="p31")
+        cg.add("bs_resolved", "# 可预测性：predictable\n离线缓存锚点条目 zkv7",
+               layer="unresolved", importance=0.3, actor="p31")
+        cg.add("bs_unknowable", "# 可预测性：unknowable\n外部世界的结构性事实",
+               layer="unresolved", importance=0.3, actor="p31")
+        cg.add("bs_carried", "# 可预测性：predictable\n日志采样链路锚点条目 zxc9",
+               layer="unresolved", importance=0.3, actor="p31")
+        check("盲区描述 → 锚点（检索打分）",
+              predict.anchor_from_description(cg, "离线缓存锚点条目 zkv7") == "learn_anchor"
+              and predict.anchor_from_description(cg, "日志采样链路锚点条目 zxc9")
+              == "learn_anchor2")
+        nf = call_tool(cg, "cg", {"op": "insight", "action": "learn",
+                                  "blindspot_id": "no_such_blindspot"})
+        check("未知盲区不编造", nf.get("ok") is False and nf.get("status") == "not_found",
+              str(nf.get("reason"))[:60])
+        lr = call_tool(cg, "cg", {"op": "insight", "action": "learn"})
+        steps = {s["blindspot_id"]: s for s in (lr.get("steps") or [])}
+        check("未解问题进入学习闭环",
+              {"bs_resolved", "bs_unknowable", "bs_carried"} <= set(steps),
+              str(sorted(steps)))
+        check("终态五态可判定且不越集",
+              all(s.get("terminal") in ("unknowable", "no_anchor", "unresolved",
+                                        "carried", "resolved") for s in steps.values()))
+        check("可判定终点 → resolved",
+              steps.get("bs_resolved", {}).get("terminal") == "resolved"
+              and steps["bs_resolved"].get("terminal_node") == "learn_kb",
+              str(steps.get("bs_resolved", {}).get("terminal")))
+        check("unknowable 不生成路线（拒绝编造）",
+              steps.get("bs_unknowable", {}).get("terminal") == "unknowable"
+              and steps["bs_unknowable"].get("routes") == 0)
+        check("未达可判定终点 → carried",
+              steps.get("bs_carried", {}).get("terminal") in ("carried", "unresolved"),
+              str(steps.get("bs_carried", {}).get("terminal")))
+        lw = call_tool(cg, "cg", {"op": "insight", "action": "learn", "apply": True})
+        gap = [nid for nid in lw.get("written") or []]
+        check("carried/unresolved 落 gap_hint 待补线索",
+              len(gap) == int(lw["summary"].get("carried", 0))
+              + int(lw["summary"].get("unresolved", 0)) and len(gap) >= 1,
+              f"written={gap}")
+        gnode = (cg.get(gap[0]) or {}) if gap else {}
+        gfm = gnode.get("frontmatter") or {}
+        check("gap_hint 是线索而非事实（contextual + 显式非事实标注）",
+              gfm.get("layer") == "contextual"
+              and "gap_hint" in (gfm.get("tags") or [])
+              and "非既有事实" in (gnode.get("content") or ""),
+              str(gfm.get("tags")))
+        # 回归防线：线索节点逐字抄了盲区描述，若不隔离会顶掉真实锚点，
+        # 使「线索 → 0 路线 → 永远 unresolved」并让重复 learn 不再幂等。
+        check("线索节点不劫持锚点解析（自我污染防护）",
+              predict.anchor_from_description(
+                  cg, (predict.find_blindspot(cg, "bs_resolved") or {}).get("description") or "")
+              == "learn_anchor")
+        lw2 = call_tool(cg, "cg", {"op": "insight", "action": "learn", "apply": True})
+        check("盲区学习落库幂等（重复 apply 不新增线索）",
+              lw2.get("written") == []
+              and lw2["summary"] == lw["summary"], str(lw2.get("written")))
+
+        # ---------------------------------------------- ⑥ 洞见记账与裁决
+        print("\n【6】insight.record / verify")
+        bad = None
+        try:
+            call_tool(cg, "cg", {"op": "insight", "action": "record"})
+        except Exception as exc:                       # noqa: BLE001
+            bad = exc
+        check("record 缺正文直接报错（不臆造）", isinstance(bad, ValueError),
+              type(bad).__name__)
+        r1 = call_tool(cg, "cg", {"op": "insight", "action": "record",
+                                  "statement": STMT_MAIN, "conditions": C1_OK,
+                                  "category": "caching"})
+        nid = r1.get("node_id")
+        check("record 记为 pending 并快照条件",
+              r1.get("state") == "pending" and r1.get("duplicate") is False
+              and r1.get("window", {}).get("open") is True, str(nid))
+        rec_fm = (cg.get(nid) or {}).get("frontmatter") or {}
+        check("条件快照含 C1–C8 且无缺失",
+              set(rec_fm.get("insight_conditions") or {}) == set(insight.CONDITION_KEYS)
+              and rec_fm.get("insight_missing") == [], str(rec_fm.get("insight_missing")))
+        r1b = call_tool(cg, "cg", {"op": "insight", "action": "record",
+                                   "statement": STMT_MAIN, "conditions": C1_OK})
+        check("record 同正文幂等", r1b.get("duplicate") is True
+              and r1b.get("node_id") == nid)
+        v0 = call_tool(cg, "cg", {"op": "insight", "action": "verify", "node_id": nid})
+        check("无证据不判定（保持 pending）",
+              v0.get("decided") is False and v0.get("state") == "pending",
+              str(v0.get("reason"))[:50])
+        v1 = call_tool(cg, "cg", {"op": "insight", "action": "verify",
+                                  "node_id": nid, "v_types": ["v1"]})
+        check("V1 不足门槛不判定", v1.get("decided") is False)
+        v3 = call_tool(cg, "cg", {"op": "insight", "action": "verify",
+                                  "node_id": nid, "v_types": ["v3"]})
+        check("V3 外部确证 → verified + 重要度保底 + 保护位",
+              v3.get("state") == "verified" and v3.get("importance") >= 0.9
+              and v3.get("protected") is True, str(v3.get("reason")))
+        v3fm = (cg.get(nid) or {}).get("frontmatter") or {}
+        check("verified 后基底升为 data 且带 verified 标签",
+              v3fm.get("verification_basis") == "data"
+              and insight.TAG_VERIFIED in (v3fm.get("tags") or []))
+        bad_v = None
+        try:
+            call_tool(cg, "cg", {"op": "insight", "action": "verify",
+                                 "node_id": nid, "verdict": "maybe"})
+        except Exception as exc:                       # noqa: BLE001
+            bad_v = exc
+        check("非法 verdict 报错", isinstance(bad_v, ValueError))
+        r2 = call_tool(cg, "cg", {"op": "insight", "action": "record",
+                                  "statement": "被外部证据证伪的洞见：所有缓存都应常驻内存"})
+        vf = call_tool(cg, "cg", {"op": "insight", "action": "verify",
+                                  "node_id": r2["node_id"], "verdict": "falsified"})
+        check("显式证伪 → falsified",
+              vf.get("state") == "falsified" and vf.get("decided") is True)
+        lst = call_tool(cg, "cg", {"op": "insight", "action": "list",
+                                   "state": "verified"})
+        check("list 按状态过滤",
+              lst.get("count", 0) >= 1
+              and all(e["state"] == "verified" for e in lst.get("events") or []))
+
+        # ---------------------------------------------- ⑦ CER 报告
+        print("\n【7】insight.report CER（样本不足不判定）")
+        rep0 = call_tool(cg, "cg", {"op": "insight", "action": "report"})
+        check("样本不足 → insufficient 且不给结论",
+              rep0.get("layer_state") == "insufficient" and rep0.get("cer") is None
+              and rep0.get("decided") < rep0.get("sample_min"), str(rep0.get("decided")))
+        for i in range(20):
+            ri = call_tool(cg, "cg", {"op": "insight", "action": "record",
+                                      "statement": f"P31 样本 {i}：需在离线条件下验证的分支 {i}",
+                                      "conditions": C1_OK})
+            call_tool(cg, "cg", {"op": "insight", "action": "verify",
+                                 "node_id": ri["node_id"],
+                                 "verdict": ("verified" if i < 16 else "falsified")})
+        rep = call_tool(cg, "cg", {"op": "insight", "action": "report"})
+        import math as _m
+        exp_v, exp_f = 17, 5
+        n = exp_v + exp_f
+        p = exp_v / n
+        se = _m.sqrt(p * (1 - p) / n)
+        check("样本足 → CER = verified/(verified+falsified)",
+              rep.get("decided") == n and abs(rep.get("cer") - round(p, 4)) < 1e-4
+              and rep.get("verified") == exp_v and rep.get("falsified") == exp_f,
+              f"cer={rep.get('cer')} decided={rep.get('decided')}")
+        check("2×SE 显著性带",
+              abs(rep.get("two_se") - round(2 * se, 4)) < 1e-4
+              and rep.get("significant") is True, f"two_se={rep.get('two_se')}")
+        check("合格样本 → layer_state=reliable",
+              rep.get("layer_state") == "reliable", str(rep.get("layer_state")))
+
+        # ---------------------------------------------- ⑧ 结构洞察
+        print("\n【8】insight.outlook 结构洞察")
+        ol = call_tool(cg, "cg", {"op": "insight", "action": "outlook"})
+        check("D1 结构判断含分层/分位/保护率",
+              ol.get("d1", {}).get("total", 0) > 0
+              and "layers" in ol["d1"] and "importance" in ol["d1"]
+              and "protected_rate" in ol["d1"], str(ol.get("d1", {}).get("total")))
+        check("D2 趋势含洞见 CER 与 unresolved",
+              ol.get("d2", {}).get("insight", {}).get("recorded", 0) >= 22
+              and "unresolved_nodes" in ol["d2"], str(ol.get("d2", {}).get("insight")))
+        check("建议可执行（带证据与建议调用）",
+              isinstance(ol.get("suggestions"), list)
+              and all({"priority", "action", "evidence", "call"} <= set(s)
+                      for s in ol["suggestions"]), f"n={len(ol.get('suggestions') or [])}")
+        check("outlook 不判有效性（写进 note）",
+              "不判有效性" in str(ol.get("note") or ""))
+
+        # ---------------------------------------------- ⑨ 权限分档
+        print("\n【9】insight 按 action 分级授权")
+        cg.principal = _principal("output")
+        check("output 可读洞察",
+              not _denied(lambda: call_tool(cg, "cg", {"op": "insight",
+                                                       "action": "outlook"})))
+        check("output 不可记洞见",
+              _denied(lambda: call_tool(cg, "cg", {"op": "insight",
+                                                   "action": "record",
+                                                   "statement": "越权记录"})))
+        check("output 不可批量重构落库",
+              _denied(lambda: call_tool(cg, "cg", {"op": "insight",
+                                                   "action": "reconstruct",
+                                                   "ids": ["rec_a"], "apply": True})))
+        cg.principal = _principal("record")
+        check("写层可记可裁",
+              not _denied(lambda: call_tool(cg, "cg", {"op": "insight",
+                                                       "action": "record",
+                                                       "statement": "写层记录"}))
+              and not _denied(lambda: call_tool(cg, "cg", {"op": "insight",
+                                                           "action": "list"})))
+        check("写层不可批量学习落库",
+              _denied(lambda: call_tool(cg, "cg", {"op": "insight",
+                                                   "action": "learn", "apply": True})))
+        cg.principal = _principal("guest")
+        check("guest 整 op 被拒",
+              _denied(lambda: call_tool(cg, "cg", {"op": "insight",
+                                                   "action": "outlook"})))
+        cg.principal = Principal(tenant="default", actor="p31", clearance="internal",
+                                 can_write=True, can_admin=True, role="designer")
+
+        # ---------------------------------------------- ⑩ 未知 action
+        print("\n【10】未知 action 不静默成功")
+        for op in ("insight", "consolidate"):
+            b = None
+            try:
+                call_tool(cg, "cg", {"op": op, "action": "no_such_action"})
+            except Exception as exc:                   # noqa: BLE001
+                b = exc
+            check(f"{op} 未知 action 抛错", b is not None, type(b).__name__)
+
+    finally:
+        try:
+            cg.close()
+        except Exception:                              # noqa: BLE001
+            pass
+        shutil.rmtree(tmp, ignore_errors=True)
+
+    print("\n" + "=" * 68)
+    print(f"结果：{PASS} 通过 / {FAIL} 失败")
+    if FAILS:
+        print("失败项：")
+        for n in FAILS:
+            print(f"  - {n}")
+    print("=" * 68)
+    return 0 if FAIL == 0 else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())