@furongjun1999/dsh-memory 0.4.4 → 0.4.5

package/md_cg/test_p36_kp_align.py

@@ -0,0 +1,225 @@
+# -*- coding: utf-8 -*-
+"""kp_ 对齐端到端（P36 · 设计态验收）。
+
+对照交付计划「决策一：三段硬依赖，不可并行倒置」与「条目 4：豁免分批解除与验收对比」：
+
+  ① 基线：`ccg_exempt=true` 的节点四态恒 DEFER（可检索、不可判定）
+  ② 反证：**证据未补就摘豁免** → 由 DEFER 退化为 BLINDSPOT（必须比现状更差）
+  ③ 顺序：证据（crosscheck）→ 5 要素齐备 → 才允许摘豁免
+  ④ 分批：`exempt_plan` 只列就绪节点，`exempt_apply` 逐批推进、可续跑、天然幂等
+  ⑤ 打破：摘豁免后有证据节点转 ACCEPT，DEFER 恒定被打破
+  ⑥ 诚实：占位空壳与缺证据节点**不摘**，保持 DEFER（不以「已完成」充数）
+  ⑦ 对比：产出改造前后四态分布
+  ⑧ 回滚：`exempt_rollback` 还原豁免标记，四态回退
+  ⑨ 分派：`backfill.run(action="exempt*")` 与 `MdCGOS.MAINTAIN_ACTIONS` 已登记
+
+独立临时根，重跑 ≡ 首跑。
+运行：python -m md_cg.test_p36_kp_align
+"""
+from __future__ import annotations
+
+import os
+import shutil
+import tempfile
+
+from . import backfill, crosscheck as cc, nodefile
+from .mdcg import STATE_ACCEPT, STATE_BLINDSPOT, STATE_DEFER
+from .mdcos import MdCGOS
+
+PASS = FAIL = 0
+FAILS = []
+
+
+def ok(cond, label):
+    global PASS, FAIL
+    if cond:
+        PASS += 1
+    else:
+        FAIL += 1
+        FAILS.append(label)
+        print(f"  FAIL {label}")
+
+
+SRC_BOOK = ["人教版《中外历史纲要》上册 第 3 单元"]
+PH = "[历史·高中] 骨架锚点，内容待填充"
+
+
+def _doc(fn, cond, sub, exe, neg="问完全无关主题"):
+    return (f"# 功能名：{fn}\n# 生效条件：{cond}\n# 子功能：{sub}\n"
+            f"# 执行：{exe}\n# 不适用条件：{neg}\n")
+
+
+def _seed(root):
+    """模拟「接线已完成、等待补证据」的迁移期现场。"""
+    cg = MdCGOS(root)
+    for nid, name in (("kp_a", "历史·秦朝郡县制"), ("kp_b", "历史·科举变迁"),
+                      ("kp_c", "历史·市舶司")):
+        cg.add(nid, _doc(name.split("·")[1], f"问{nid[-1]}", f"说明{nid}",
+                         f"讲{nid}"),
+               layer="knowledge", ccg_exempt=True,
+               state_attributes={"name": name})
+    # 已齐备（有验证方式行 + 合法基底），仅差摘豁免
+    cg.add("kp_present",
+           _doc("浮力计算", "问浮力", "按阿基米德原理算", "代入公式")
+           .replace("# 执行：", "# 验证方式：编译器/静态检查通过\n# 执行："),
+           layer="knowledge", ccg_exempt=True, verification_basis="test",
+           state_attributes={"name": "数学·浮力计算"})
+    # 占位空壳：不可摘（不冒充齐全）
+    cg.add("kp_ph", "# 功能名：骨架节点\n\n迁移期空壳。\n", layer="knowledge",
+           ccg_exempt=True,
+           state_attributes={"comment": {"生效条件": [PH], "子功能": PH}})
+    return cg
+
+
+def _state(cg, nid, q="问主题"):
+    e = cg.index["nodes"][nid]
+    fm, c = cg._read(e)
+    return MdCGOS.judge_qualification({"frontmatter": fm, "content": c}, q,
+                                      {"query": q})["state"]
+
+
+def _dist(cg, ids_):
+    out = {}
+    for nid in ids_:
+        st = _state(cg, nid)
+        out[st] = out.get(st, 0) + 1
+    return out
+
+
+ALL = ["kp_a", "kp_b", "kp_c", "kp_present", "kp_ph"]
+TARGETS = ["kp_a", "kp_b", "kp_c"]
+
+
+def main():
+    global PASS, FAIL
+    tmp = tempfile.mkdtemp(prefix="mdcg_p36_")
+    try:
+        root = os.path.join(tmp, "root")
+        os.makedirs(root, exist_ok=True)
+        cg = _seed(root)
+
+        # 前置：豁免标记确实落到 frontmatter
+        fm, _ = cg._read(cg.index["nodes"]["kp_a"])
+        ok(fm.get("ccg_exempt") is True, "①豁免标记已写入 frontmatter")
+
+        # ---------- ① 基线 ----------
+        before = _dist(cg, ALL)
+        ok(before.get(STATE_DEFER) == 5 and before.get(STATE_ACCEPT, 0) == 0,
+           f"①基线：5 个豁免节点四态恒 DEFER（实测 {before}）")
+
+        # ---------- ② 反证：顺序倒置更差 ----------
+        bad = backfill.exempt_apply(cg, ids=["kp_a"], require_ready=False,
+                                    batch="p36bad", actor="tester")
+        ok(bad["written"] == 1, "②强制（跳过就绪校验）摘除成功，可复现风险路径")
+        ok(_state(cg, "kp_a") == STATE_BLINDSPOT,
+           "②证据未补就摘豁免 → 由 DEFER 退化为 BLINDSPOT（证明顺序不可倒置）")
+        rb_bad = backfill.exempt_rollback(cg, batch="p36bad", actor="tester")
+        ok(rb_bad["reverted"] == 1 and _state(cg, "kp_a") == STATE_DEFER,
+           "②还原后回到 DEFER（风险路径可撤销）")
+
+        # ---------- ③ 就绪闸门：证据没补，一个也摘不了 ----------
+        p0 = backfill.exempt_plan(cg, require_ready=True)
+        ok(p0["targeted"] == 1 and p0["planned_ids"] == ["kp_present"],
+           f"③只列就绪节点（实测 {p0['planned_ids']}）")
+        ok(p0["skipped_unready"] == 4 and p0["unready_reasons"].get("verification_basis") == 4,
+           f"③未就绪 4 个，归因 verification_basis（实测 {p0['unready_reasons']}）")
+        ok("ccg5" in p0["unready_reasons"], "③5 要素未齐备同样计入归因")
+
+        # ---------- ③ 补证据（crosscheck） ----------
+        verdicts = []
+        for nid in TARGETS:
+            verdicts += [
+                {"id": nid, "unit": "reflect", "field": "验证方式",
+                 "value": "据教材单元小结核对", "basis": "textbook",
+                 "source": SRC_BOOK, "verdict": "accept"},
+                {"id": nid, "unit": "verify", "field": "验证方式",
+                 "value": "据教材单元小结核对", "verdict": "accept"},
+            ]
+        rep = cc.crosscheck(cg, ids=TARGETS, verdicts=verdicts, apply=True,
+                            batch="p36ev", actor="tester")
+        ok(rep["written"] == 3, f"③证据落库 3 个（实测 {rep['written']}）")
+        for nid in TARGETS:
+            fm, c = cg._read(cg.index["nodes"][nid])
+            ok(nodefile.ccg_completeness(c).get("complete"),
+               f"③{nid} 5 要素齐备（含验证方式行）")
+
+        # ---------- ④ 分批摘豁免（可续跑、幂等） ----------
+        p1 = backfill.exempt_plan(cg, require_ready=True, sample=2)
+        ok(p1["targeted"] == 4, f"④就绪 4 个（实测 {p1['targeted']}）")
+        ok(len(p1["sample"]) == 2 and all(s in p1["planned_ids"] for s in p1["sample"]),
+           "④抽样接口可用（确定性样本，用于 50 节点验收）")
+        b1 = backfill.exempt_apply(cg, require_ready=True, limit=2, batch="p36a",
+                                   actor="tester")
+        ok(b1["written"] == 2, f"④第一批摘 2 个（实测 {b1['written']}）")
+        b2 = backfill.exempt_apply(cg, require_ready=True, limit=2, batch="p36a",
+                                   actor="tester")
+        ok(b2["written"] == 2, f"④第二批自动续（实测 {b2['written']}）")
+        b3 = backfill.exempt_apply(cg, require_ready=True, limit=2, batch="p36a",
+                                   actor="tester")
+        ok(b3["written"] == 0 and b3["skipped_drift"] == 0,
+           "④第三批无事可做（幂等，不重复摘）")
+
+        # kp_ph 始终不可摘
+        ok(_state(cg, "kp_ph") == STATE_DEFER, "⑥占位空壳始终不摘，保持 DEFER")
+        fm, _ = cg._read(cg.index["nodes"]["kp_ph"])
+        ok(fm.get("ccg_exempt") is True, "⑥空壳豁免标记未被误摘")
+
+        # ---------- ⑤ 打破 DEFER 恒定 ----------
+        for nid in TARGETS:
+            ok(_state(cg, nid) == STATE_ACCEPT,
+               f"⑤{nid} 摘豁免后转 ACCEPT（DEFER 恒定被打破）")
+        ok(_state(cg, "kp_present") == STATE_ACCEPT, "⑤已齐备节点转 ACCEPT")
+        after = _dist(cg, ALL)
+        ok(after.get(STATE_ACCEPT) == 4 and after.get(STATE_DEFER) == 1,
+           f"⑤术后分布 ACCEPT=4 / DEFER=1（实测 {after}）")
+
+        # ---------- ⑦ 前后对比 ----------
+        report = {"before": before, "after": after,
+                  "accept_gain": after.get(STATE_ACCEPT, 0) - before.get(STATE_ACCEPT, 0),
+                  "still_defer": [n for n in ALL if _state(cg, n) == STATE_DEFER]}
+        ok(report["accept_gain"] == 4 and report["still_defer"] == ["kp_ph"],
+           f"⑦前后对比：ACCEPT +4，剩余 DEFER 仅为不可闭合项（实测 {report}）")
+
+        # ---------- ⑧ 留痕与回滚 ----------
+        h = backfill.history(cg, action="exempt", limit=100)
+        ok(h["total"] >= 4 and all("exempt_before" in r for r in h["records"]),
+           "⑧豁免留痕含原始标记（可反向应用）")
+        # 显式点名一个未就绪节点：必须留痕，不得静默丢弃
+        skip = backfill.exempt_apply(cg, ids=["kp_ph"], require_ready=True,
+                                     batch="p36skip", actor="tester")
+        ok(skip["written"] == 0 and skip["skipped_unready"] == 1,
+           f"⑧点名未就绪节点 → 拒绝执行（实测 {skip['skipped_unready']}）")
+        ok("exempt_skip" in {r.get("action") for r in
+                             (backfill.history(cg, limit=200)["records"])},
+           "⑧未就绪节点留 exempt_skip 记录（不静默跳过）")
+        rb = backfill.exempt_rollback(cg, batch="p36a", actor="tester")
+        ok(rb["reverted"] == 4, f"⑧回滚 4 条（实测 {rb['reverted']}）")
+        back = _dist(cg, ALL)
+        ok(back.get(STATE_DEFER) == 5 and back.get(STATE_ACCEPT, 0) == 0,
+           f"⑧回滚后四态回到基线（实测 {back}）")
+
+        # ---------- ⑨ 分派登记 ----------
+        ok("exempt" in backfill.ACTIONS and "exempt" in MdCGOS.MAINTAIN_ACTIONS,
+           "⑨exempt 三动作已在 ACTIONS 与 MAINTAIN_ACTIONS 登记")
+        ok(all(a in MdCGOS.MAINTAIN_ACTIONS for a in backfill.ACTIONS),
+           "⑨契约：backfill.ACTIONS ⊆ MAINTAIN_ACTIONS")
+        r_plan = backfill.run(cg, "exempt", layer="knowledge")
+        ok(r_plan["action"] == "exempt" and r_plan["dry_run"] is True,
+           "⑨run(action='exempt') 默认 dry-run")
+        r_hist = backfill.run(cg, "exempt_history", limit=5)
+        ok(r_hist["total"] >= 4, "⑨run(action='exempt_history') 可查留痕")
+
+    finally:
+        shutil.rmtree(tmp, ignore_errors=True)
+
+    print(f"\nPASS={PASS} FAIL={FAIL}")
+    if FAILS:
+        print("FAILED:")
+        for f in FAILS:
+            print("  -", f)
+        raise SystemExit(1)
+    print("全部通过：P36 kp_ 端到端（顺序/就绪闸门/分批/四态对比/回滚）")
+
+
+if __name__ == "__main__":
+    main()

package/md_cg/test_p37_condition_space.py

@@ -0,0 +1,248 @@
+# -*- coding: utf-8 -*-
+"""条件空间合成与生效条件口径（P37 · 设计态验收）。
+
+口径：**观测位置 ≠ 生效条件**。生效条件是**整条条件空间声明的合成**——
+「什么智能载体、在什么时间、使用什么方法、在什么位置、得出什么结论」这一完整
+坐标；结论即「功能名」，不进生效条件。旧版取 `condition_space.observation_position`
+单槽加前缀「观测位置：」冒充生效条件（真实库 109 条），本轮废止。
+
+验收覆盖：
+  ① 合成纯函数 `nodefile.condition_space_text`：槽序固定、缺失槽不写、
+     四槽不全不构成声明、多值用「、」不污染槽分隔、全时窗可读化、
+     具体时间窗渲染 UTC、占位槽不冒充
+  ② `backfill.derive_fields` **无单槽回退**：四槽不齐 → 不生成生效条件
+  ③ 生效条件进入必填口径：缺它 → `ccg_completeness.complete=False`
+     → `judge_qualification` 落 BLINDSPOT（而非静默 ACCEPT）
+  ④ 待补台账 `conditions_pending`：缺声明且推不出的节点可见、缺失槽可数
+  ⑤ 存量清洗 `fix_conditions_*`：rewrite / drop 分流、留痕、幂等、真逆回滚
+  ⑥ 复算 `verify_conditions`：弱等价残留（legacy_position）归零
+
+运行：python -m md_cg.test_p37_condition_space
+"""
+from __future__ import annotations
+
+import os
+import shutil
+import tempfile
+
+from . import backfill, nodefile
+from .mdcos import MdCGOS
+
+PASS = FAIL = 0
+FAILS = []
+
+LEGACY_TEXT = "观测位置：某个外部观测位"
+
+
+def ok(cond, label):
+    global PASS, FAIL
+    if cond:
+        PASS += 1
+    else:
+        FAIL += 1
+        FAILS.append(label)
+        print(f"  FAIL {label}")
+
+
+def _full_cs():
+    return {
+        "observation_position": "本地文件系统（摄取器内部观测位）",
+        "observation_tool": "扩展名映射表 + Ingestor",
+        "time_window": [0.0, 9999999999.0],
+        "existence_constraint": "仅对本地可读文件生效",
+    }
+
+
+def _seed(root):
+    """造：四槽齐备 / 单槽 / 旧口径四槽 / 旧口径单槽，四类对照。"""
+    cg = MdCGOS(root)
+    cg.add("cx_full",
+           "# 功能名：文件摄取分派\n\n按扩展名把文件路由到对应摄取器。\n",
+           layer="knowledge", condition_space=_full_cs(),
+           verification_basis="test", non_applicable_conditions=["二进制文件"],
+           state_attributes={"comment": {
+               "子功能": "按扩展名路由", "执行": "读映射表→调用 Ingestor"}})
+    cg.add("cx_posonly",
+           "# 功能名：只有观测位置的历史记忆\n\n条件空间只声明了一维。\n",
+           layer="knowledge",
+           condition_space={"observation_position": "某个外部观测位"},
+           state_attributes={"comment": {
+               "子功能": "待补：条件空间其余三维", "执行": "待补"}})
+    # 旧口径落地形态：正文行与 comment 里都是「观测位置：…」单槽冒充
+    cg.add("cx_legacy",
+           f"# 功能名：旧口径四槽记忆\n# 生效条件：{LEGACY_TEXT}\n\n历史迁移。\n",
+           layer="knowledge", condition_space=_full_cs(),
+           verification_basis="test", non_applicable_conditions=["无"],
+           state_attributes={"comment": {
+               "生效条件": LEGACY_TEXT, "子功能": "s", "执行": "e"}})
+    cg.add("cx_legacy_drop",
+           f"# 功能名：旧口径单槽记忆\n# 生效条件：{LEGACY_TEXT}\n\n历史迁移。\n",
+           layer="knowledge",
+           condition_space={"observation_position": "某个外部观测位"})
+    cg.flush()
+    cg.rebuild_index()
+    return cg
+
+
+def _inject_legacy_log(cg, nodes):
+    """把「我们自己写下的旧口径行」补进留痕——清洗的真源是留痕，不是猜测。"""
+    for nid in nodes:
+        backfill.append_jsonl(backfill._log_path(cg), {
+            "action": "backfill", "ts": 1.0, "batch": "legacy",
+            "entry_id": "legacy|" + nid, "write_id": "w|" + nid,
+            "node": nid, "layer": "knowledge",
+            "fields": {"生效条件": {
+                "before": None, "after": LEGACY_TEXT,
+                "basis": backfill.LEGACY_CONDITION_BASIS}}})
+
+
+def main():
+    global PASS, FAIL
+    tmp = tempfile.mkdtemp(prefix="mdcg_p37_")
+    try:
+        root = os.path.join(tmp, "root")
+        os.makedirs(root, exist_ok=True)
+        cg = _seed(root)
+
+        # ---------- ① 合成纯函数 ----------
+        cs = {"observation_position": "观测位P",
+              "observation_tool": ["工具A", "工具B"],
+              "time_window": [1787914173.0, 1787917773.0],
+              "existence_constraint": "仅限X"}
+        t = nodefile.condition_space_text(cs)
+        ok(t.startswith("载体/位置：观测位P；时间："),
+           "①以「载体/位置」起头（载体与位置同源合并）")
+        ok("；方法：工具A、工具B；约束：仅限X" in t,
+           "①槽序固定；多值用「、」，不污染槽分隔符")
+        ok(t.count("；") == 3, "①四槽 → 三段分隔（无重复段）")
+        ok(nodefile.condition_space_text({"observation_position": "P"}) == "",
+           "①四槽不全 → 空串（部分槽不构成条件空间声明）")
+        ok(nodefile.condition_space_text({"observation_position": "P"},
+                                        require_full=False) == "载体/位置：P",
+           "①require_full=False 仅渲染已声明槽（展示用，非生效条件来源）")
+        ok("全时窗（任意时刻成立）" in nodefile.condition_space_text({
+            "observation_position": "P", "observation_tool": "T",
+            "time_window": [0.0, 9999999999.0],
+            "existence_constraint": "E"}),
+           "①[0, 1e10] 渲染为「全时窗」（合法声明，非空占位）")
+        ok("2026-" in nodefile.time_window_text([1787914173.0, 1787917773.0])
+           and "（UTC）" in nodefile.time_window_text([1787914173.0, 1787917773.0]),
+           "①具体时间窗渲染为 UTC 时刻（不落裸数字进召回键）")
+        ok(nodefile.condition_space_text({
+            "observation_position": "骨架锚点", "observation_tool": "T",
+            "time_window": [0.0, 1.0], "existence_constraint": "E"}) == "",
+           "①占位槽不冒充已声明（整体 fail-closed 为不可合成）")
+        ok(nodefile.is_legacy_position_condition(LEGACY_TEXT)
+           and not nodefile.is_legacy_position_condition(t),
+           "①旧口径识别只认前缀形态（不做语义猜测）")
+
+        # ---------- ② derive 无单槽回退 ----------
+        fm_po, c_po = cg._read(cg.index["nodes"]["cx_posonly"])
+        d_po = backfill.derive_fields(fm_po, c_po)
+        ok("生效条件" not in d_po, "②单槽 observation_position 不生成生效条件")
+        ok(bool(d_po.get("子功能")) and bool(d_po.get("执行")),
+           "②其余可推字段仍照补（一项缺证据不停摆整批）")
+        fm_f, c_f = cg._read(cg.index["nodes"]["cx_full"])
+        d_f = backfill.derive_fields(fm_f, c_f)
+        ok(d_f.get("生效条件", ("", ""))[1] == backfill.BASIS_CONDITION_SYNTH,
+           "②四槽齐备 → 来源标记为「条件空间四槽合成」")
+        ok(d_f["生效条件"][0] == nodefile.condition_space_text(
+               fm_f.get("condition_space")),
+           "②合成值 == 唯一入口 condition_space_text 的输出")
+
+        # ---------- ③ 必填口径：缺生效条件 → BLINDSPOT ----------
+        ok(nodefile.CCG_REQUIRED == nodefile.CCG_MARKS,
+           "③门槛即全部 MARKS（生效条件不再可隐含）")
+        ok(nodefile.ccg_completeness(c_po)["complete"] is False,
+           "③缺生效条件 → complete 为 False")
+        ok("生效条件" in (set(nodefile.CCG_REQUIRED)
+                         - set(nodefile.ccg_completeness(c_po)["required_present"])),
+           "③缺失项被精确归因到「生效条件」")
+        j_po = cg.judge_qualification(
+            {"frontmatter": fm_po, "content": c_po}, query="")
+        ok(j_po["state"] == "BLINDSPOT",
+           "③缺生效条件 → BLINDSPOT（不静默 ACCEPT）")
+        ok("生效条件" in j_po.get("reason", ""), "③判定理由点名缺失项")
+        # 齐全节点的正文须含全部 6 行 MARKS（ccg_completeness 只读正文）
+        c_full = ("# 功能名：文件摄取分派\n"
+                  f"# 生效条件：{nodefile.condition_space_text(fm_f.get('condition_space'))}\n"
+                  "# 子功能：按扩展名路由\n"
+                  "# 执行：读映射表→调用 Ingestor\n"
+                  "# 验证方式：test（单测覆盖扩展名分派）\n"
+                  "# 不适用条件：二进制文件\n\n按扩展名把文件路由到对应摄取器。\n")
+        j_f = cg.judge_qualification(
+            {"frontmatter": fm_f, "content": c_full}, query="")
+        ok(j_f["state"] != "BLINDSPOT", "③四槽齐备且齐全 → 不误判 BLINDSPOT")
+
+        # ---------- ④ 待补台账 ----------
+        pend = backfill.conditions_pending(cg, prefix="cx_")
+        ok(pend["dry_run"] is True, "④台账标记 dry_run（只读）")
+        ids = [i["id"] for i in pend["items"]]
+        ok("cx_posonly" in ids, "④单槽节点进台账")
+        ok("cx_full" not in ids, "④四槽齐备节点不进台账")
+        it_po = next(i for i in pend["items"] if i["id"] == "cx_posonly")
+        # mdcos.add 在调用方未给 time_window 时以写入时刻锚定 1 小时观测窗
+        # （mdcg.OBSERVATION_WINDOW_SEC），故此处缺的是方法 + 约束两维
+        ok(it_po["missing_slots"] == ["observation_tool", "existence_constraint"],
+           "④缺失槽逐条列出（供后续补充，不靠猜）")
+        ok("BLINDSPOT" in it_po["fallback"], "④台账写明兜底裁决口径")
+
+        # ---------- ⑤ 存量清洗：rewrite / drop ----------
+        _inject_legacy_log(cg, ["cx_legacy", "cx_legacy_drop"])
+        p = backfill.fix_conditions_plan(cg)
+        ok(p["dry_run"] is True and p["legacy"] >= 2,
+           "⑤预演识别留痕里的旧口径行")
+        ok(p["rewrite"] >= 1 and p["drop"] >= 1,
+           "⑤分流：四槽可合成→rewrite / 仍不全→drop")
+        ok(LEGACY_TEXT in cg._read(cg.index["nodes"]["cx_legacy"])[1],
+           "⑤预演不改盘")
+
+        a = backfill.fix_conditions_apply(cg, batch="p37fix", actor="tester")
+        ok(a["rewritten"] >= 1 and a["dropped"] >= 1, "⑤清洗写入非零且有留痕")
+        fm_l, c_l = cg._read(cg.index["nodes"]["cx_legacy"])
+        synth_l = nodefile.condition_space_text(fm_l.get("condition_space"))
+        ok(f"# 生效条件：{synth_l}" in c_l,
+           "⑤rewrite：冒充行改写为四槽合成声明")
+        ok(fm_l["state_attributes"]["comment"]["生效条件"] == synth_l,
+           "⑤rewrite：frontmatter.comment 同步到同一口径")
+        fm_d, c_d = cg._read(cg.index["nodes"]["cx_legacy_drop"])
+        ok("# 生效条件" not in c_d, "⑤drop：补不全的冒充行被删除")
+        ok("生效条件" not in (fm_d.get("state_attributes", {}).get("comment") or {}),
+           "⑤drop：comment 里的同一污染副本一并还原")
+        ok(nodefile.ccg_completeness(c_d)["complete"] is False,
+           "⑤drop 后节点不再「假装」齐全")
+        a2 = backfill.fix_conditions_apply(cg, batch="p37fix", actor="tester")
+        ok(a2["rewritten"] == 0 and a2["dropped"] == 0,
+           "⑤幂等：清洗后不再有候选行")
+        pend2 = backfill.conditions_pending(cg, prefix="cx_")
+        ok(pend2["by_missing"].get("缺方法+约束", 0) >= 1,
+           "⑤drop 转入待补台账（缺哪几维可见）")
+        ok(any(r.get("action") == "condition_pending"
+               for r in backfill.history(cg)["records"]),
+           "⑤待补登记落留痕（可追溯，不丢证据）")
+
+        # ---------- ⑥ 复算 + 回滚 ----------
+        v = backfill.verify_conditions(cg, prefix="cx_")
+        ok(v["legacy_position"] == 0, "⑥复算：弱等价残留归零（验收指标）")
+        ok(v["from_synthesis"] >= 1, "⑥复算：至少一条来自四槽合成")
+        r = backfill.fix_conditions_rollback(cg, batch="p37fix", actor="tester")
+        ok(r["reverted"] >= 2, "⑥清洗可回滚（真逆操作，非暴力删除）")
+        ok(LEGACY_TEXT in cg._read(cg.index["nodes"]["cx_legacy"])[1],
+           "⑥回滚恢复被改写的旧值")
+        ok(LEGACY_TEXT in cg._read(cg.index["nodes"]["cx_legacy_drop"])[1],
+           "⑥回滚重建被删除的行")
+        v2 = backfill.verify_conditions(cg, prefix="cx_")
+        ok(v2["legacy_position"] >= 2, "⑥回滚后残留计数如实回升（复算可信）")
+
+        print(f"\nP37 结果：{PASS} 通过 / {FAIL} 失败")
+        if FAILS:
+            print("失败项：")
+            for f in FAILS:
+                print(f"  - {f}")
+    finally:
+        shutil.rmtree(tmp, ignore_errors=True)
+    return 1 if FAIL else 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())