@furongjun1999/dsh-memory 0.4.11 → 0.5.0

package/md_cg/test_p20_evolution.py

@@ -1,315 +1,315 @@
-# -*- coding: utf-8 -*-
-"""md_cg · 第 20 篇：演化账本（md 载体 · 记录规律与状态 · 可回滚）
-
-设计立场：演化记录的价值不在机器回放，而在**认知规律可读**。
-  · 载体是 md（`_evolution/ledger.md`），人类可读，单一真相源
-  · 每一次修改 = 对一条缺失条件的补充（pattern 规律 + missing 缺失条件）
-  · 记录的是认知规律与状态，不是实现细节（实现属于 git）
-  · 回滚把状态撤回 before，且撤销本身也记一条条目（不可静默）
-
-覆盖：
-  A 载体与格式：md 生成 · 中文标签人类可读 · 状态快照内嵌可解析 · append-only
-  B 规律统计：按缺失条件维度聚类 · top_patterns · evolution/rollback 分账
-  C 回滚：状态撤回 before · 删除改动后才出现的字段 · 撤销留痕 · dry_run · 拒绝条件
-  D 与 consolidate 联动：固化 = 补缺失条件 → 自动落账 + 可回滚
-  E 自描述与 MCP：catalog / summary / health / mdcg_evolution / cg op / 管理权
-  F 幂等与并发：多次追加不丢
-
-运行：python -m md_cg.test_p20_evolution
-"""
-from __future__ import annotations
-
-import json
-import os
-import shutil
-import tempfile
-
-from . import consolidate, evolution
-from .mdcos import MdCGOS, MdCGSecure
-from .security import AccessDenied, Principal
-
-PASS = FAIL = 0
-FAILS = []
-
-
-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}")
-
-
-# ---------------- consolidate 的假双模型（同 P6 口径） ----------------
-BODY = ("# 功能名：微分方程应用\n"
-        "微分方程应用：建模与求解，使用数值方法迭代求解，需给定边界条件\n")
-CAND = {"生效条件": ["微分方程", "数值方法"], "子功能": ["建模", "求解"],
-        "执行": "使用数值方法迭代求解", "不适用条件": ["边界条件缺失"]}
-
-
-def fake_reflect(prompt):
-    return "```json\n" + json.dumps(CAND, ensure_ascii=False) + "\n```"
-
-
-def fake_verify(prompt):
-    out = {f: {"keep": v if isinstance(v, list) else [v], "drop": [],
-               "reason": "逐条核验"} for f, v in CAND.items()}
-    return json.dumps(out, ensure_ascii=False)
-
-
-def main():
-    root = tempfile.mkdtemp(prefix="mdcg_p20_")
-    cg = MdCGOS(root)
-
-    # ================= A. 载体与格式 =================
-    print("\n[A] 载体与格式（md 人类可读 · 状态内嵌可解析 · append-only）")
-
-    cg.add("n1", "# 功能名：示例节点\n用于演化账本测试\n", layer="knowledge",
-           importance=0.5, confidence=0.6)
-
-    raised = False
-    try:
-        cg.evolution_record(node_id="n1", pattern="", missing="执行")
-    except ValueError:
-        raised = True
-    check("A1 缺「规律」被拒（没有规律就没有可复用认知）", raised)
-
-    raised = False
-    try:
-        cg.evolution_record(node_id="n1", pattern="规律X", kind="不存在的类型")
-    except ValueError:
-        raised = True
-    check("A2 未知演化类型被拒", raised)
-
-    e1 = cg.evolution_record(
-        node_id="n1", pattern="缺时间窗口的条件在跨季度查询里被误召回",
-        missing="conditions.time_window", change="补充生效条件的时间窗口",
-        evidence="replay通过", source="manual",
-        before={"confidence": 0.6, "importance": 0.5},
-        after={"confidence": 0.55, "importance": 0.5})
-
-    led = os.path.join(root, evolution.EVOLUTION_DIR, evolution.LEDGER_NAME)
-    check("A3 账本落在 _evolution/ledger.md", os.path.exists(led), led)
-
-    text = evolution.read_ledger(cg)
-    check("A4 正文是人类可读中文标签",
-          all(f"**{k}**" in text for k in ("规律", "缺失条件", "动作", "状态", "证据")),
-          text[:0] or "")
-    check("A5 规律原文入账", "跨季度查询里被误召回" in text)
-    check("A6 状态差异人可读", "confidence 0.6→0.55" in text,
-          [l for l in text.split("\n") if "状态" in l][:1])
-
-    blocks = evolution._STATE_RE.findall(text)
-    parsed_state = json.loads(blocks[0]) if blocks else {}
-    check("A7 状态快照内嵌 json 块且可解析",
-          parsed_state.get("before", {}).get("confidence") == 0.6
-          and parsed_state.get("kind") == "condition_gap",
-          str(parsed_state)[:80])
-
-    e2 = cg.evolution_record(node_id="n1", pattern="缺执行维的节点无法判定适用性",
-                             missing="execution", change="补执行", source="manual")
-    text2 = evolution.read_ledger(cg)
-    check("A8 append-only：旧条目不被覆盖",
-          e1["entry_id"] in text2 and e2["entry_id"] in text2)
-    check("A9 账本头部说明只出现一次",
-          text2.count("# 演化账本 · md_cg") == 1)
-
-    ents = cg.evolution_entries(limit=10)["entries"]
-    check("A10 entries 倒序（最新在前）", ents and ents[0]["entry_id"] == e2["entry_id"])
-    check("A11 entries limit 生效",
-          len(cg.evolution_entries(limit=1)["entries"]) == 1)
-    check("A12 show 取单条", cg.evolution_show(e1["entry_id"])["entry"]["node_id"] == "n1")
-
-    # ================= B. 规律统计 =================
-    print("\n[B] 规律统计（认知功能记录规律：看清反复缺什么）")
-
-    cg.evolution_record(node_id="n1", pattern="缺时间窗口的条件在跨季度查询里被误召回",
-                        missing="conditions.time_window", source="consolidate")
-    pat = cg.evolution_patterns()
-    check("B1 按缺失条件维度聚类", pat["by_missing"].get("conditions.time_window") == 2,
-          str(pat["by_missing"]))
-    check("B2 top_patterns 计数重复规律",
-          pat["top_patterns"] and pat["top_patterns"][0]["count"] == 2,
-          str(pat["top_patterns"][:1]))
-    check("B3 来源统计", pat["by_source"].get("manual") == 2
-          and pat["by_source"].get("consolidate") == 1, str(pat["by_source"]))
-    check("B4 evolution 与 rollback 分账（此时无回滚）",
-          pat["evolutions"] == 3 and pat["rollbacks"] == 0, str(pat["total"]))
-
-    # ================= C. 回滚 =================
-    print("\n[C] 回滚（撤回补充 · 撤销留痕 · 拒绝条件）")
-
-    fm0 = cg.get("n1")["frontmatter"]
-    fm0["non_applicable_conditions"] = ["边界条件缺失"]
-    fm0["confidence"] = 0.55
-    cg._write_node("n1", os.path.join(root, cg.get("n1")["path"]), fm0,
-                   cg.get("n1")["content"])
-
-    before = {"confidence": 0.6, "importance": 0.5}
-    after = {"confidence": 0.55, "importance": 0.5,
-             "non_applicable_conditions": ["边界条件缺失"]}
-    e3 = cg.evolution_record(node_id="n1", pattern="补负条件以防误召回",
-                             missing="non_applicable_conditions",
-                             change="补充不适用条件", source="manual",
-                             before=before, after=after)
-
-    dry = cg.evolution_rollback(e3["entry_id"], dry_run=True)
-    check("C1 dry_run 预演列出将删除的字段",
-          dry["ok"] and dry["would_remove"] == ["non_applicable_conditions"],
-          str(dry.get("would_remove")))
-    check("C2 dry_run 不落盘", cg.get("n1")["frontmatter"].get("confidence") == 0.55)
-    n_before = len(cg.evolution_entries(limit=0)["entries"])
-    check("C3 dry_run 不写条目",
-          len(cg.evolution_entries(limit=0)["entries"]) == n_before)
-
-    rb = cg.evolution_rollback(e3["entry_id"])
-    fm1 = cg.get("n1")["frontmatter"]
-    check("C4 回滚把状态撤回 before", fm1.get("confidence") == 0.6, str(fm1.get("confidence")))
-    check("C5 回滚删除「改动后才出现」的字段（真的撤回补充）",
-          "non_applicable_conditions" not in fm1, str(fm1.get("non_applicable_conditions")))
-    check("C6 回滚本身记一条条目且不可静默",
-          rb["rollback_entry"]["kind"] == evolution.KIND_ROLLBACK
-          and rb["rollback_entry"]["rollback_of"] == e3["entry_id"],
-          str(rb["rollback_entry"].get("rollback_of")))
-    pat2 = cg.evolution_patterns()
-    check("C7 回滚计入 rollbacks 分账", pat2["rollbacks"] == 1, str(pat2))
-
-    again = cg.evolution_rollback(rb["rollback_entry"]["entry_id"])
-    check("C8 不能回滚一条回滚记录", again["ok"] is False, str(again.get("error")))
-
-    e4 = cg.evolution_record(node_id="n1", pattern="无状态快照的条目")
-    no_state = cg.evolution_rollback(e4["entry_id"])
-    check("C9 无 before 的条目拒绝回滚", no_state["ok"] is False, str(no_state.get("error")))
-
-    missing = cg.evolution_rollback("evo-19700101-000000-dead")
-    check("C10 不存在的条目拒绝回滚", missing["ok"] is False, str(missing.get("error")))
-
-    # 层迁移回滚
-    cg._move_layer("n1", "contextual", reason="p20 test")
-    e5 = cg.evolution_record(node_id="n1", pattern="上下文相关，降级到 contextual",
-                             missing="conditions", change="层间迁移",
-                             source="manual",
-                             before={"layer": "knowledge"},
-                             after={"layer": "contextual"})
-    rb5 = cg.evolution_rollback(e5["entry_id"])
-    check("C11 回滚层迁移（contextual → knowledge）",
-          "layer" in rb5["applied"] and cg.get("n1")["frontmatter"].get("layer") == "knowledge",
-          str(rb5["applied"]))
-
-    # ================= D. 与 consolidate 联动 =================
-    print("\n[D] 与 consolidate 联动（每一次固化 = 补一条缺失条件）")
-
-    root2 = tempfile.mkdtemp(prefix="mdcg_p20c_")
-    try:
-        cg2 = MdCGOS(root2)
-        cg2.add("n1", BODY, layer="knowledge")
-        n0 = len(cg2.evolution_entries(limit=0)["entries"])
-        rep = consolidate.consolidate(root2, apply=True, reflect_fn=fake_reflect,
-                                      verify_fn=fake_verify, verification_basis="测试依据",
-                                      verbose=False)
-        ents = cg2.evolution_entries(limit=0)["entries"]
-        check("D1 固化前账本为空", n0 == 0, str(n0))
-        check("D2 固化后自动落账 1 条", len(ents) == 1 and rep["written"] == 1,
-              f"entries={len(ents)} written={rep['written']}")
-        ce = ents[0]
-        check("D3 来源标记 consolidate", ce["source"] == "consolidate", str(ce.get("source")))
-        check("D4 缺失条件记全（含执行维）",
-              all(f in ce["missing"] for f in ("执行", "生效条件", "不适用条件")),
-              str(ce.get("missing")))
-        check("D5 状态快照记下补充后的负条件",
-              (ce["state"].get("after") or {}).get("non_applicable_conditions") == ["边界条件缺失"],
-              str(ce["state"].get("after")))
-        check("D6 规律可读且非空", len(ce["pattern"]) > 8, ce["pattern"])
-
-        rb = cg2.evolution_rollback(ce["entry_id"])
-        fm = cg2.get("n1")["frontmatter"]
-        check("D7 consolidate 的补充可回滚（负条件被撤回）",
-              rb["ok"] and not fm.get("non_applicable_conditions"),
-              str(fm.get("non_applicable_conditions")))
-        cg2.close()
-    finally:
-        shutil.rmtree(root2, ignore_errors=True)
-
-    # ================= E. 自描述与 MCP =================
-    print("\n[E] 自描述与 MCP（可发现 · 可路由 · 管理权）")
-
-    cat = cg.evolution_catalog()
-    check("E1 原则含「每一次修改 = 对一条缺失条件的补充」",
-          any("缺失条件" in p for p in cat["principles"]), str(cat["principles"][:1]))
-    check("E2 自描述写清回滚范围与未覆盖项",
-          "scope" in cat["rollback"] and cat["rollback"].get("not_covered"),
-          str(cat["rollback"].get("not_covered")))
-    check("E3 自描述列出状态字段与条件维度",
-          "confidence" in cat["state_fields"] and "execution" in cat["condition_dims"])
-
-    s = cg.evolution_summary()
-    check("E4 summary 一句话 + 最近条目", "演化" in s["text"] and s["recent"],
-          s["text"])
-    h = cg.health_os()
-    check("E5 health_os 并入 evolution 面", "evolution" in h["os"],
-          str(list(h["os"].keys()))[:60])
-
-    from . import mcp_server as ms
-    names = {t["name"] for t in ms.TOOLS}
-    check("E6 MCP 注册 mdcg_evolution", "mdcg_evolution" in names)
-    r = ms._cg_call(cg, {"op": "evolution", "action": "catalog"})
-    check("E7 cg op=evolution 路由到账本", r.get("module") == "evolution",
-          str(r.get("module")))
-    r2 = ms._evolution_call(cg, {"action": "patterns", "limit": 3})
-    check("E8 细粒度入口与 cg op 同源", "by_missing" in r2, str(list(r2.keys())))
-
-    root3 = tempfile.mkdtemp(prefix="mdcg_p20s_")
-    try:
-        cg_s = MdCGSecure(root3, principal=Principal(
-            tenant="private", actor="op", clearance="private",
-            can_write=True, can_admin=False))
-        cg_s.add("n1", "# 受管节点\n", layer="knowledge")
-        es = cg_s.evolution_record(node_id="n1", pattern="管理权测试",
-                                   before={"confidence": 0.6},
-                                   after={"confidence": 0.5})
-        denied = False
-        try:
-            cg_s.evolution_rollback(es["entry_id"])
-        except AccessDenied:
-            denied = True
-        check("E9 回滚需 can_admin（直接 API 亦拦截）", denied)
-        cg_s.close()
-    finally:
-        shutil.rmtree(root3, ignore_errors=True)
-
-    # ================= F. 幂等与并发 =================
-    print("\n[F] 幂等与并发（append-only 不丢）")
-
-    root4 = tempfile.mkdtemp(prefix="mdcg_p20f_")
-    try:
-        cg4 = MdCGOS(root4)
-        cg4.add("n1", "# 节点\n", layer="knowledge")
-        ids = [cg4.evolution_record(node_id="n1", pattern=f"规律{i}",
-                                    missing="execution")["entry_id"]
-               for i in range(10)]
-        got = {e["entry_id"] for e in cg4.evolution_entries(limit=0)["entries"]}
-        check("F1 连续 10 次追加全部保留", set(ids) <= got and len(got) == 10,
-              f"got={len(got)}")
-        cg4.evolution_record(node_id="other", pattern="别的节点", missing="conditions")
-        hh = cg4.evolution_history("n1")
-        check("F2 history 按节点过滤", len(hh["entries"]) == 10, str(len(hh["entries"])))
-        check("F3 账本可被重新解析（幂等读取）",
-              len(cg4.evolution_entries(limit=0)["entries"]) == 11)
-        cg4.close()
-    finally:
-        shutil.rmtree(root4, ignore_errors=True)
-
-    cg.close()
-    shutil.rmtree(root, ignore_errors=True)
-
-    print(f"\n结果：PASS={PASS}  FAIL={FAIL}")
-    if FAILS:
-        print("失败项：" + "、".join(FAILS))
-    return 0 if FAIL == 0 else 1
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())
+# -*- coding: utf-8 -*-
+"""md_cg · 第 20 篇：演化账本（md 载体 · 记录规律与状态 · 可回滚）
+
+设计立场：演化记录的价值不在机器回放，而在**认知规律可读**。
+  · 载体是 md（`_evolution/ledger.md`），人类可读，单一真相源
+  · 每一次修改 = 对一条缺失条件的补充（pattern 规律 + missing 缺失条件）
+  · 记录的是认知规律与状态，不是实现细节（实现属于 git）
+  · 回滚把状态撤回 before，且撤销本身也记一条条目（不可静默）
+
+覆盖：
+  A 载体与格式：md 生成 · 中文标签人类可读 · 状态快照内嵌可解析 · append-only
+  B 规律统计：按缺失条件维度聚类 · top_patterns · evolution/rollback 分账
+  C 回滚：状态撤回 before · 删除改动后才出现的字段 · 撤销留痕 · dry_run · 拒绝条件
+  D 与 consolidate 联动：固化 = 补缺失条件 → 自动落账 + 可回滚
+  E 自描述与 MCP：catalog / summary / health / mdcg_evolution / cg op / 管理权
+  F 幂等与并发：多次追加不丢
+
+运行：python -m md_cg.test_p20_evolution
+"""
+from __future__ import annotations
+
+import json
+import os
+import shutil
+import tempfile
+
+from . import consolidate, evolution
+from .mdcos import MdCGOS, MdCGSecure
+from .security import AccessDenied, Principal
+
+PASS = FAIL = 0
+FAILS = []
+
+
+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}")
+
+
+# ---------------- consolidate 的假双模型（同 P6 口径） ----------------
+BODY = ("# 功能名：微分方程应用\n"
+        "微分方程应用：建模与求解，使用数值方法迭代求解，需给定边界条件\n")
+CAND = {"生效条件": ["微分方程", "数值方法"], "子功能": ["建模", "求解"],
+        "执行": "使用数值方法迭代求解", "不适用条件": ["边界条件缺失"]}
+
+
+def fake_reflect(prompt):
+    return "```json\n" + json.dumps(CAND, ensure_ascii=False) + "\n```"
+
+
+def fake_verify(prompt):
+    out = {f: {"keep": v if isinstance(v, list) else [v], "drop": [],
+               "reason": "逐条核验"} for f, v in CAND.items()}
+    return json.dumps(out, ensure_ascii=False)
+
+
+def main():
+    root = tempfile.mkdtemp(prefix="mdcg_p20_")
+    cg = MdCGOS(root)
+
+    # ================= A. 载体与格式 =================
+    print("\n[A] 载体与格式（md 人类可读 · 状态内嵌可解析 · append-only）")
+
+    cg.add("n1", "# 功能名：示例节点\n用于演化账本测试\n", layer="knowledge",
+           importance=0.5, confidence=0.6)
+
+    raised = False
+    try:
+        cg.evolution_record(node_id="n1", pattern="", missing="执行")
+    except ValueError:
+        raised = True
+    check("A1 缺「规律」被拒（没有规律就没有可复用认知）", raised)
+
+    raised = False
+    try:
+        cg.evolution_record(node_id="n1", pattern="规律X", kind="不存在的类型")
+    except ValueError:
+        raised = True
+    check("A2 未知演化类型被拒", raised)
+
+    e1 = cg.evolution_record(
+        node_id="n1", pattern="缺时间窗口的条件在跨季度查询里被误召回",
+        missing="conditions.time_window", change="补充生效条件的时间窗口",
+        evidence="replay通过", source="manual",
+        before={"confidence": 0.6, "importance": 0.5},
+        after={"confidence": 0.55, "importance": 0.5})
+
+    led = os.path.join(root, evolution.EVOLUTION_DIR, evolution.LEDGER_NAME)
+    check("A3 账本落在 _evolution/ledger.md", os.path.exists(led), led)
+
+    text = evolution.read_ledger(cg)
+    check("A4 正文是人类可读中文标签",
+          all(f"**{k}**" in text for k in ("规律", "缺失条件", "动作", "状态", "证据")),
+          text[:0] or "")
+    check("A5 规律原文入账", "跨季度查询里被误召回" in text)
+    check("A6 状态差异人可读", "confidence 0.6→0.55" in text,
+          [l for l in text.split("\n") if "状态" in l][:1])
+
+    blocks = evolution._STATE_RE.findall(text)
+    parsed_state = json.loads(blocks[0]) if blocks else {}
+    check("A7 状态快照内嵌 json 块且可解析",
+          parsed_state.get("before", {}).get("confidence") == 0.6
+          and parsed_state.get("kind") == "condition_gap",
+          str(parsed_state)[:80])
+
+    e2 = cg.evolution_record(node_id="n1", pattern="缺执行维的节点无法判定适用性",
+                             missing="execution", change="补执行", source="manual")
+    text2 = evolution.read_ledger(cg)
+    check("A8 append-only：旧条目不被覆盖",
+          e1["entry_id"] in text2 and e2["entry_id"] in text2)
+    check("A9 账本头部说明只出现一次",
+          text2.count("# 演化账本 · md_cg") == 1)
+
+    ents = cg.evolution_entries(limit=10)["entries"]
+    check("A10 entries 倒序（最新在前）", ents and ents[0]["entry_id"] == e2["entry_id"])
+    check("A11 entries limit 生效",
+          len(cg.evolution_entries(limit=1)["entries"]) == 1)
+    check("A12 show 取单条", cg.evolution_show(e1["entry_id"])["entry"]["node_id"] == "n1")
+
+    # ================= B. 规律统计 =================
+    print("\n[B] 规律统计（认知功能记录规律：看清反复缺什么）")
+
+    cg.evolution_record(node_id="n1", pattern="缺时间窗口的条件在跨季度查询里被误召回",
+                        missing="conditions.time_window", source="consolidate")
+    pat = cg.evolution_patterns()
+    check("B1 按缺失条件维度聚类", pat["by_missing"].get("conditions.time_window") == 2,
+          str(pat["by_missing"]))
+    check("B2 top_patterns 计数重复规律",
+          pat["top_patterns"] and pat["top_patterns"][0]["count"] == 2,
+          str(pat["top_patterns"][:1]))
+    check("B3 来源统计", pat["by_source"].get("manual") == 2
+          and pat["by_source"].get("consolidate") == 1, str(pat["by_source"]))
+    check("B4 evolution 与 rollback 分账（此时无回滚）",
+          pat["evolutions"] == 3 and pat["rollbacks"] == 0, str(pat["total"]))
+
+    # ================= C. 回滚 =================
+    print("\n[C] 回滚（撤回补充 · 撤销留痕 · 拒绝条件）")
+
+    fm0 = cg.get("n1")["frontmatter"]
+    fm0["non_applicable_conditions"] = ["边界条件缺失"]
+    fm0["confidence"] = 0.55
+    cg._write_node("n1", os.path.join(root, cg.get("n1")["path"]), fm0,
+                   cg.get("n1")["content"])
+
+    before = {"confidence": 0.6, "importance": 0.5}
+    after = {"confidence": 0.55, "importance": 0.5,
+             "non_applicable_conditions": ["边界条件缺失"]}
+    e3 = cg.evolution_record(node_id="n1", pattern="补负条件以防误召回",
+                             missing="non_applicable_conditions",
+                             change="补充不适用条件", source="manual",
+                             before=before, after=after)
+
+    dry = cg.evolution_rollback(e3["entry_id"], dry_run=True)
+    check("C1 dry_run 预演列出将删除的字段",
+          dry["ok"] and dry["would_remove"] == ["non_applicable_conditions"],
+          str(dry.get("would_remove")))
+    check("C2 dry_run 不落盘", cg.get("n1")["frontmatter"].get("confidence") == 0.55)
+    n_before = len(cg.evolution_entries(limit=0)["entries"])
+    check("C3 dry_run 不写条目",
+          len(cg.evolution_entries(limit=0)["entries"]) == n_before)
+
+    rb = cg.evolution_rollback(e3["entry_id"])
+    fm1 = cg.get("n1")["frontmatter"]
+    check("C4 回滚把状态撤回 before", fm1.get("confidence") == 0.6, str(fm1.get("confidence")))
+    check("C5 回滚删除「改动后才出现」的字段（真的撤回补充）",
+          "non_applicable_conditions" not in fm1, str(fm1.get("non_applicable_conditions")))
+    check("C6 回滚本身记一条条目且不可静默",
+          rb["rollback_entry"]["kind"] == evolution.KIND_ROLLBACK
+          and rb["rollback_entry"]["rollback_of"] == e3["entry_id"],
+          str(rb["rollback_entry"].get("rollback_of")))
+    pat2 = cg.evolution_patterns()
+    check("C7 回滚计入 rollbacks 分账", pat2["rollbacks"] == 1, str(pat2))
+
+    again = cg.evolution_rollback(rb["rollback_entry"]["entry_id"])
+    check("C8 不能回滚一条回滚记录", again["ok"] is False, str(again.get("error")))
+
+    e4 = cg.evolution_record(node_id="n1", pattern="无状态快照的条目")
+    no_state = cg.evolution_rollback(e4["entry_id"])
+    check("C9 无 before 的条目拒绝回滚", no_state["ok"] is False, str(no_state.get("error")))
+
+    missing = cg.evolution_rollback("evo-19700101-000000-dead")
+    check("C10 不存在的条目拒绝回滚", missing["ok"] is False, str(missing.get("error")))
+
+    # 层迁移回滚
+    cg._move_layer("n1", "contextual", reason="p20 test")
+    e5 = cg.evolution_record(node_id="n1", pattern="上下文相关，降级到 contextual",
+                             missing="conditions", change="层间迁移",
+                             source="manual",
+                             before={"layer": "knowledge"},
+                             after={"layer": "contextual"})
+    rb5 = cg.evolution_rollback(e5["entry_id"])
+    check("C11 回滚层迁移（contextual → knowledge）",
+          "layer" in rb5["applied"] and cg.get("n1")["frontmatter"].get("layer") == "knowledge",
+          str(rb5["applied"]))
+
+    # ================= D. 与 consolidate 联动 =================
+    print("\n[D] 与 consolidate 联动（每一次固化 = 补一条缺失条件）")
+
+    root2 = tempfile.mkdtemp(prefix="mdcg_p20c_")
+    try:
+        cg2 = MdCGOS(root2)
+        cg2.add("n1", BODY, layer="knowledge")
+        n0 = len(cg2.evolution_entries(limit=0)["entries"])
+        rep = consolidate.consolidate(root2, apply=True, reflect_fn=fake_reflect,
+                                      verify_fn=fake_verify, verification_basis="测试依据",
+                                      verbose=False)
+        ents = cg2.evolution_entries(limit=0)["entries"]
+        check("D1 固化前账本为空", n0 == 0, str(n0))
+        check("D2 固化后自动落账 1 条", len(ents) == 1 and rep["written"] == 1,
+              f"entries={len(ents)} written={rep['written']}")
+        ce = ents[0]
+        check("D3 来源标记 consolidate", ce["source"] == "consolidate", str(ce.get("source")))
+        check("D4 缺失条件记全（含执行维）",
+              all(f in ce["missing"] for f in ("执行", "生效条件", "不适用条件")),
+              str(ce.get("missing")))
+        check("D5 状态快照记下补充后的负条件",
+              (ce["state"].get("after") or {}).get("non_applicable_conditions") == ["边界条件缺失"],
+              str(ce["state"].get("after")))
+        check("D6 规律可读且非空", len(ce["pattern"]) > 8, ce["pattern"])
+
+        rb = cg2.evolution_rollback(ce["entry_id"])
+        fm = cg2.get("n1")["frontmatter"]
+        check("D7 consolidate 的补充可回滚（负条件被撤回）",
+              rb["ok"] and not fm.get("non_applicable_conditions"),
+              str(fm.get("non_applicable_conditions")))
+        cg2.close()
+    finally:
+        shutil.rmtree(root2, ignore_errors=True)
+
+    # ================= E. 自描述与 MCP =================
+    print("\n[E] 自描述与 MCP（可发现 · 可路由 · 管理权）")
+
+    cat = cg.evolution_catalog()
+    check("E1 原则含「每一次修改 = 对一条缺失条件的补充」",
+          any("缺失条件" in p for p in cat["principles"]), str(cat["principles"][:1]))
+    check("E2 自描述写清回滚范围与未覆盖项",
+          "scope" in cat["rollback"] and cat["rollback"].get("not_covered"),
+          str(cat["rollback"].get("not_covered")))
+    check("E3 自描述列出状态字段与条件维度",
+          "confidence" in cat["state_fields"] and "execution" in cat["condition_dims"])
+
+    s = cg.evolution_summary()
+    check("E4 summary 一句话 + 最近条目", "演化" in s["text"] and s["recent"],
+          s["text"])
+    h = cg.health_os()
+    check("E5 health_os 并入 evolution 面", "evolution" in h["os"],
+          str(list(h["os"].keys()))[:60])
+
+    from . import mcp_server as ms
+    names = {t["name"] for t in ms.TOOLS}
+    check("E6 MCP 注册 mdcg_evolution", "mdcg_evolution" in names)
+    r = ms._cg_call(cg, {"op": "evolution", "action": "catalog"})
+    check("E7 cg op=evolution 路由到账本", r.get("module") == "evolution",
+          str(r.get("module")))
+    r2 = ms._evolution_call(cg, {"action": "patterns", "limit": 3})
+    check("E8 细粒度入口与 cg op 同源", "by_missing" in r2, str(list(r2.keys())))
+
+    root3 = tempfile.mkdtemp(prefix="mdcg_p20s_")
+    try:
+        cg_s = MdCGSecure(root3, principal=Principal(
+            tenant="private", actor="op", clearance="private",
+            can_write=True, can_admin=False))
+        cg_s.add("n1", "# 受管节点\n", layer="knowledge")
+        es = cg_s.evolution_record(node_id="n1", pattern="管理权测试",
+                                   before={"confidence": 0.6},
+                                   after={"confidence": 0.5})
+        denied = False
+        try:
+            cg_s.evolution_rollback(es["entry_id"])
+        except AccessDenied:
+            denied = True
+        check("E9 回滚需 can_admin（直接 API 亦拦截）", denied)
+        cg_s.close()
+    finally:
+        shutil.rmtree(root3, ignore_errors=True)
+
+    # ================= F. 幂等与并发 =================
+    print("\n[F] 幂等与并发（append-only 不丢）")
+
+    root4 = tempfile.mkdtemp(prefix="mdcg_p20f_")
+    try:
+        cg4 = MdCGOS(root4)
+        cg4.add("n1", "# 节点\n", layer="knowledge")
+        ids = [cg4.evolution_record(node_id="n1", pattern=f"规律{i}",
+                                    missing="execution")["entry_id"]
+               for i in range(10)]
+        got = {e["entry_id"] for e in cg4.evolution_entries(limit=0)["entries"]}
+        check("F1 连续 10 次追加全部保留", set(ids) <= got and len(got) == 10,
+              f"got={len(got)}")
+        cg4.evolution_record(node_id="other", pattern="别的节点", missing="conditions")
+        hh = cg4.evolution_history("n1")
+        check("F2 history 按节点过滤", len(hh["entries"]) == 10, str(len(hh["entries"])))
+        check("F3 账本可被重新解析（幂等读取）",
+              len(cg4.evolution_entries(limit=0)["entries"]) == 11)
+        cg4.close()
+    finally:
+        shutil.rmtree(root4, ignore_errors=True)
+
+    cg.close()
+    shutil.rmtree(root, ignore_errors=True)
+
+    print(f"\n结果：PASS={PASS}  FAIL={FAIL}")
+    if FAILS:
+        print("失败项：" + "、".join(FAILS))
+    return 0 if FAIL == 0 else 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())