@furongjun1999/dsh-memory 0.4.11 → 0.5.1

package/md_cg/test_wisdom_md_store.py

@@ -1,292 +1,292 @@
-# -*- coding: utf-8 -*-
-"""智慧之书 md 真源存储测试：MdStore 对拍 + ConditionDex md 模式读写闭环。
-
-口径（使用者裁定 2026-09-14）：智慧之书知识源=md 文档，sqlite 黑箱退役为
-可重建派生物。 MdStore 不得悄悄分叉（字段/集合/边域/检索打分）
-- 组B 写闭环：空语料根 → ConditionDex(md_root) → seed_base → md 落盘 →
-  重开读回（读写同源）+ seed 幂等
-- 组C 往返与留痕：知识卡+边+观测留痕 双模往返一致
-- 组D 兼容回归：sqlite 形态（tmp 库）全链路不炸；read_conn duck
-
-运行：python -m md_cg.test_wisdom_md_store
-"""
-import json
-import os
-import shutil
-import sys
-import tempfile
-
-_HERE = os.path.dirname(os.path.abspath(__file__))
-_ROOT = os.path.dirname(_HERE)
-for _p in (_ROOT, _HERE, os.path.join(_HERE, "whitebox_kb"),
-           os.path.join(_HERE, "whitebox_kb", "wisdom")):
-    if _p not in sys.path:
-        sys.path.insert(0, _p)
-
-from wisdom_book import (ConditionDex, DEFAULT_DB, DEFAULT_MD_ROOT,  # noqa: E402
-                         BASE_ENTRIES, BASE_RELATIONS)
-from md_access import MdConn, read_conn  # noqa: E402
-from md_whitebox import corpus_ready  # noqa: E402
-from md_store import MdStore  # noqa: E402
-from aeis_core import ConditionSpace, MemoryLayer, STNode  # noqa: E402
-
-_N = 0
-
-
-def ok(cond, label):
-    global _N
-    _N += 1
-    assert cond, f"FAIL #{_N}: {label}"
-    print(f"  ok#{_N} {label}")
-
-
-# ---------------------------------------------------------------------------
-# 组A · 生产数据对拍（sqlite 派生库 ↔ md 语料根，两者都在时执行）
-# ---------------------------------------------------------------------------
-
-def group_a(tmp):
-    # 语料就绪判据 = 「目录存在**且含 .md**」（2026-09-20 v14 缺陷 F）：空壳
-    # 目录（兄弟测试经 MdCGOS(root) 的 makedirs 留下）不算就绪，否则组A 会
-    # 以「md 语料行数 0 > 4000」的失败示人，而真实原因是真源根本不在位。
-    if not (os.path.exists(DEFAULT_DB) and corpus_ready(DEFAULT_MD_ROOT)):
-        print("组A SKIP：生产对拍数据不全（派生库缺失或语料根缺失/空壳）")
-        return
-    print("组A · 生产数据对拍（全量）")
-    # sqlite 侧用副本隔离：search_content 的 increment_access 会落库，
-    # 直连生产库等于测试污染运行库（access_count 漂移实证）——backup 到
-    # 副本后权威实现在副本上跑，源库全程只读
-    import sqlite3
-    copy_db = os.path.join(tmp, "parity_copy.db")
-    src = sqlite3.connect(DEFAULT_DB)
-    dst = sqlite3.connect(copy_db)
-    src.backup(dst)
-    dst.close()
-    src.close()
-    sq = ConditionDex(db_path=copy_db)
-    md = ConditionDex(md_root=DEFAULT_MD_ROOT)
-    try:
-        # sqlite 侧容错读：生产库存在 cs 含未知键的脏行（ConditionSpace.from_json
-        # 严格构造会炸——黑箱脆弱性实证）；与 MdStore._ensure 同口径跳过，
-        # 保证对拍集合=双侧干净集
-        s_by_id = {}
-        n_dirty = 0
-        for row in sqlite3.connect(copy_db).execute("SELECT * FROM nodes"):
-            try:
-                n = STNode.from_row(tuple(row))
-            except (TypeError, ValueError, KeyError):
-                n_dirty += 1
-                continue
-            if n.layer.value == "knowledge":
-                s_by_id[n.id] = n
-        s_nodes = list(s_by_id.values())
-        print(f"  （sqlite 侧脏行跳过 {n_dirty}）")
-        m_conn = MdConn(DEFAULT_MD_ROOT)
-        m_rows = m_conn._all_rows()
-        ok(len(m_rows) > 4000, f"md 语料行数 {len(m_rows)} > 4000（导出快照在位）")
-        m_ids = {r[0] for r in m_rows}
-        s_ids = set(s_by_id)
-        # md 是 sqlite 的导出快照：交集逐字段等价（sqlite 可能另有运行时新增）
-        inter = m_ids & s_ids
-        ok(len(inter) >= min(1000, int(len(m_ids) * 0.9)),
-           f"md<->sqlite 知识层交集 {len(inter)}（md {len(m_ids)} / sq {len(s_ids)}）")
-
-        # 字段级对拍（抽 40）——语义不变字段逐字节；豁免四个非知识字段：
-        # spatial（空值 '{}'↔'[]'）/ temporal（转写进 cs.time_window）/
-        # condition_space（时间窗归一注入）/ access_count（运行时统计，
-        # sqlite 版尽力而为落库、md 版进程内统计，两侧都不承诺一致）
-        m_store = md.store
-        m_store._ensure()
-        bad = 0
-        for nid in sorted(inter)[:40]:
-            sn, mn = s_by_id[nid], m_store.get_node(nid)
-            if (sn.content != mn.content or sn.modality != mn.modality
-                    or list(sn.tags) != list(mn.tags)
-                    or sn.state_attributes != mn.state_attributes
-                    or sn.semantic_coordinates != mn.semantic_coordinates
-                    or sn.entity_id != mn.entity_id
-                    or abs(sn.importance - mn.importance) > 1e-9
-                    or abs(sn.confidence - mn.confidence) > 1e-9
-                    or abs(sn.created_at - mn.created_at) > 1e-6):
-                bad += 1
-        ok(bad == 0, f"字段级对拍 40 节点零分叉·语义不变字段（bad={bad}）")
-
-        # 全量集合对拍：两种存储物理序不同（sqlite 插入序 vs md 路径字典序），
-        # 首屏 limit 截断集合必不同——对拍对象是全集（差≤1 = 双侧容错粒度差：
-        # 脏行在 sqlite 侧被跳过、在 md 侧以归一形态装载成功）
-        m_all = {n.id for n in m_store.query_nodes(
-            layer=MemoryLayer.KNOWLEDGE, limit=1000000)}
-        ok(set(s_by_id) <= m_all and len(m_all - set(s_by_id)) <= 1,
-           f"query_nodes 全量集合（md {len(m_all)} >= sq {len(s_by_id)}，差<=1）")
-
-        # 双向边对拍（抽 10 节点）
-        bad_e = 0
-        for nid in sorted(inter)[:10]:
-            s_out = {(e.target_id, e.relation_type.value)
-                     for e in sq.store.get_outgoing_edges(nid)}
-            m_out = {(e.target_id, e.relation_type.value)
-                     for e in m_store.get_outgoing_edges(nid)}
-            if s_out != m_out:
-                bad_e += 1
-        ok(bad_e == 0, f"出边域对拍 10 节点零分叉（bad={bad_e}）")
-
-        # 检索对拍：同查询 top5 序列一致（同义词扩展+bigram 打分同构）
-        for q in ("存在论", "条件空间的适用边界", "反馈回路如何稳定系统"):
-            s_top = [n.id for n, _ in sq.store.search_content(
-                q, layers=[MemoryLayer.KNOWLEDGE], limit=5)]
-            m_top = [n.id for n, _ in m_store.search_content(
-                q, layers=[MemoryLayer.KNOWLEDGE], limit=5)]
-            ok(s_top == m_top, f"search_content top5 一致：{q!r}")
-    finally:
-        sq.close()
-        md.close()
-
-
-# ---------------------------------------------------------------------------
-# 组B · md 模式写闭环（tmp 语料根）
-# ---------------------------------------------------------------------------
-
-def group_b(tmp):
-    print("组B · md 模式写闭环")
-    root = os.path.join(tmp, "wb")
-    os.makedirs(root, exist_ok=True)
-    dex = ConditionDex(md_root=root)
-    ok(dex._md_mode and dex.engine is None, "md 模式构造（engine=None）")
-    r = dex.seed_base()
-    ok(r["status"] == "seeded", f"seed_base 首播 {r['status']}")
-
-    kdir = os.path.join(root, "knowledge")
-    import hashlib
-    ids_on_disk = set()
-    if os.path.isdir(kdir):
-        for dp, _dn, fs in os.walk(kdir):
-            for f in fs:
-                if f.endswith(".md"):
-                    ids_on_disk.add(f[:-3])
-    card_ids = {"wis_" + hashlib.sha1(n["name"].encode("utf-8")).hexdigest()[:12]
-                for n in BASE_ENTRIES}
-    missing = card_ids - ids_on_disk
-    ok(not missing,
-       f"26 张知识卡确定性 id 全部落盘（knowledge 层共 "
-       f"{len(ids_on_disk)} 文件=卡+seed 留痕，缺 {len(missing)}）")
-
-    # 重开读回：读写同源（新实例从 md 装载）
-    dex.close()
-    dex2 = ConditionDex(md_root=root)
-    ok(dex2.seed_base()["status"] == "already_seeded", "seed 幂等（重开判重）")
-    seeds = [n for n in dex2.store.query_nodes(
-        layer=MemoryLayer.KNOWLEDGE, limit=300)
-        if "domain:存在论" in (n.tags or [])]
-    ok(len(seeds) >= 1, f"知识层从 md 读回（{len(seeds)} 张 domain:存在论 卡）")
-
-    # add_relation → fm.edges 落盘（schema 与导出语料一致）
-    eid = dex2.add_relation("存在论", "条件论", "causal",
-                            note="条件论以存在论为前提", confidence=0.85)
-    ok(eid.startswith("mde_"), "add_relation 返回确定性边 id")
-    nid_a = dex2._by_name["存在论"]
-    fm_txt = ""
-    for dp, _dn, fs in os.walk(kdir):
-        for f in fs:
-            if not f.endswith(".md"):
-                continue
-            with open(os.path.join(dp, f), encoding="utf-8") as fh:
-                txt = fh.read()
-            if nid_a in txt:
-                fm_txt = txt.split("---")[1]
-                break
-        if fm_txt:
-            break
-    ok(fm_txt and '"relation_type": "causal"' in fm_txt
-       and '"verified": 1' in fm_txt,
-       "fm.edges 落盘含 causal 边（verified=1）")
-
-    # 读写同源：写入即可读（MdStore.reload 生效）
-    edges = dex2.store.get_outgoing_edges(nid_a)
-    ok(any(e.relation_type.value == "causal" and e.verified
-           for e in edges), "写入后 MdStore 即时可读（verified 边）")
-    return dex2
-
-
-# ---------------------------------------------------------------------------
-# 组C · 往返与留痕
-# ---------------------------------------------------------------------------
-
-def group_c(dex2):
-    print("组C · 往返与留痕")
-    cs = ConditionSpace(observation_position="测试台",
-                        observation_tool="test_wisdom_md_store",
-                        time_window=(0.0, 9999999999.0),
-                        existence_constraint="测试用")
-    nid = dex2.add_entry("测试学科X", "存在论", "测试断言：往返一致", cs,
-                         level=2, status="verified")
-    ok(nid.startswith("wis_"), "add_entry md 确定性 id")
-    node = dex2.store.get_node(nid)
-    ok(node is not None and node.state_attributes.get("name") == "测试学科X"
-       and node.layer.value == "knowledge", "写入即可读（state_attributes 往返）")
-
-    rec = dex2._perceive(json.dumps({"type": "测试留痕"}, ensure_ascii=False),
-                         condition_space=cs, importance=0.6,
-                         tags=["观测层", "测试"])
-    ok(len(rec.id) >= 12, "留痕节点返回 id")
-    found = False
-    for dp, _dn, fs in os.walk(os.path.join(dex2.md_root, "knowledge")):
-        if rec.id + ".md" in fs:
-            found = True
-            break
-    ok(found, "留痕落 knowledge 层（对齐 sqlite add_perception 自动进知识层）")
-    # 重开可读（留痕读回）
-    dex2.close()
-    dex3 = ConditionDex(md_root=dex2.md_root)
-    got = dex3.store.get_node(rec.id)
-    ok(got is not None and got.layer.value == "knowledge", "留痕重开可读")
-    return dex3
-
-
-# ---------------------------------------------------------------------------
-# 组D · 兼容回归（sqlite 形态 + duck conn）
-# ---------------------------------------------------------------------------
-
-def group_d(tmp):
-    print("组D · 兼容回归")
-    db = os.path.join(tmp, "legacy.db")
-    dex = ConditionDex(db_path=db)          # 显式 db_path → sqlite 形态
-    ok(not dex._md_mode and dex.engine is not None, "显式 db_path → sqlite 形态")
-    ok(dex.seed_base()["status"] == "seeded", "sqlite 形态 seed 正常")
-    nid = dex.add_entry("旧路学科Y", "条件论", "旧路断言", None, level=2,
-                        status="verified")
-    ok(bool(nid), "sqlite 形态 add_entry 正常")
-    ok(bool(dex.add_relation("旧路学科Y", "条件论", "similar")),
-       "sqlite 形态建边正常")
-    dex.close()
-
-    dex2 = ConditionDex(fresh=True, db_path=os.path.join(tmp, "fresh.db"))
-    ok(not dex2._md_mode, "fresh=True → sqlite 形态（兼容 api.py 兜底）")
-    dex2.close()
-
-    if corpus_ready(DEFAULT_MD_ROOT):
-        dex3 = ConditionDex(md_root=DEFAULT_MD_ROOT)
-        try:
-            conn = read_conn(dex3)           # duck：MdStore.conn → MdConn
-            ok(isinstance(conn, MdConn), "read_conn(dex) md 模式返回 MdConn")
-            rows = conn._all_rows()
-            ok(len(rows) > 0, f"duck conn 可查询（{len(rows)} 行）")
-        finally:
-            dex3.close()
-
-
-def main():
-    ok(len(BASE_ENTRIES) >= 20, f"BASE_ENTRIES {len(BASE_ENTRIES)} 在位")
-    ok(len(BASE_RELATIONS) >= 20, f"BASE_RELATIONS {len(BASE_RELATIONS)} 在位")
-    tmp = tempfile.mkdtemp(prefix="wb_md_test_")
-    try:
-        group_a(tmp)
-        dex2 = group_b(tmp)
-        dex3 = group_c(dex2)
-        dex3.close()
-        group_d(tmp)
-    finally:
-        shutil.rmtree(tmp, ignore_errors=True)
-    print(f"\nALL PASS · {_N} assertions")
-
-
-if __name__ == "__main__":
-    main()
+# -*- coding: utf-8 -*-
+"""智慧之书 md 真源存储测试：MdStore 对拍 + ConditionDex md 模式读写闭环。
+
+口径（使用者裁定 2026-09-14）：智慧之书知识源=md 文档，sqlite 黑箱退役为
+可重建派生物。 MdStore 不得悄悄分叉（字段/集合/边域/检索打分）
+- 组B 写闭环：空语料根 → ConditionDex(md_root) → seed_base → md 落盘 →
+  重开读回（读写同源）+ seed 幂等
+- 组C 往返与留痕：知识卡+边+观测留痕 双模往返一致
+- 组D 兼容回归：sqlite 形态（tmp 库）全链路不炸；read_conn duck
+
+运行：python -m md_cg.test_wisdom_md_store
+"""
+import json
+import os
+import shutil
+import sys
+import tempfile
+
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_ROOT = os.path.dirname(_HERE)
+for _p in (_ROOT, _HERE, os.path.join(_HERE, "whitebox_kb"),
+           os.path.join(_HERE, "whitebox_kb", "wisdom")):
+    if _p not in sys.path:
+        sys.path.insert(0, _p)
+
+from wisdom_book import (ConditionDex, DEFAULT_DB, DEFAULT_MD_ROOT,  # noqa: E402
+                         BASE_ENTRIES, BASE_RELATIONS)
+from md_access import MdConn, read_conn  # noqa: E402
+from md_whitebox import corpus_ready  # noqa: E402
+from md_store import MdStore  # noqa: E402
+from aeis_core import ConditionSpace, MemoryLayer, STNode  # noqa: E402
+
+_N = 0
+
+
+def ok(cond, label):
+    global _N
+    _N += 1
+    assert cond, f"FAIL #{_N}: {label}"
+    print(f"  ok#{_N} {label}")
+
+
+# ---------------------------------------------------------------------------
+# 组A · 生产数据对拍（sqlite 派生库 ↔ md 语料根，两者都在时执行）
+# ---------------------------------------------------------------------------
+
+def group_a(tmp):
+    # 语料就绪判据 = 「目录存在**且含 .md**」（2026-09-20 v14 缺陷 F）：空壳
+    # 目录（兄弟测试经 MdCGOS(root) 的 makedirs 留下）不算就绪，否则组A 会
+    # 以「md 语料行数 0 > 4000」的失败示人，而真实原因是真源根本不在位。
+    if not (os.path.exists(DEFAULT_DB) and corpus_ready(DEFAULT_MD_ROOT)):
+        print("组A SKIP：生产对拍数据不全（派生库缺失或语料根缺失/空壳）")
+        return
+    print("组A · 生产数据对拍（全量）")
+    # sqlite 侧用副本隔离：search_content 的 increment_access 会落库，
+    # 直连生产库等于测试污染运行库（access_count 漂移实证）——backup 到
+    # 副本后权威实现在副本上跑，源库全程只读
+    import sqlite3
+    copy_db = os.path.join(tmp, "parity_copy.db")
+    src = sqlite3.connect(DEFAULT_DB)
+    dst = sqlite3.connect(copy_db)
+    src.backup(dst)
+    dst.close()
+    src.close()
+    sq = ConditionDex(db_path=copy_db)
+    md = ConditionDex(md_root=DEFAULT_MD_ROOT)
+    try:
+        # sqlite 侧容错读：生产库存在 cs 含未知键的脏行（ConditionSpace.from_json
+        # 严格构造会炸——黑箱脆弱性实证）；与 MdStore._ensure 同口径跳过，
+        # 保证对拍集合=双侧干净集
+        s_by_id = {}
+        n_dirty = 0
+        for row in sqlite3.connect(copy_db).execute("SELECT * FROM nodes"):
+            try:
+                n = STNode.from_row(tuple(row))
+            except (TypeError, ValueError, KeyError):
+                n_dirty += 1
+                continue
+            if n.layer.value == "knowledge":
+                s_by_id[n.id] = n
+        s_nodes = list(s_by_id.values())
+        print(f"  （sqlite 侧脏行跳过 {n_dirty}）")
+        m_conn = MdConn(DEFAULT_MD_ROOT)
+        m_rows = m_conn._all_rows()
+        ok(len(m_rows) > 4000, f"md 语料行数 {len(m_rows)} > 4000（导出快照在位）")
+        m_ids = {r[0] for r in m_rows}
+        s_ids = set(s_by_id)
+        # md 是 sqlite 的导出快照：交集逐字段等价（sqlite 可能另有运行时新增）
+        inter = m_ids & s_ids
+        ok(len(inter) >= min(1000, int(len(m_ids) * 0.9)),
+           f"md<->sqlite 知识层交集 {len(inter)}（md {len(m_ids)} / sq {len(s_ids)}）")
+
+        # 字段级对拍（抽 40）——语义不变字段逐字节；豁免四个非知识字段：
+        # spatial（空值 '{}'↔'[]'）/ temporal（转写进 cs.time_window）/
+        # condition_space（时间窗归一注入）/ access_count（运行时统计，
+        # sqlite 版尽力而为落库、md 版进程内统计，两侧都不承诺一致）
+        m_store = md.store
+        m_store._ensure()
+        bad = 0
+        for nid in sorted(inter)[:40]:
+            sn, mn = s_by_id[nid], m_store.get_node(nid)
+            if (sn.content != mn.content or sn.modality != mn.modality
+                    or list(sn.tags) != list(mn.tags)
+                    or sn.state_attributes != mn.state_attributes
+                    or sn.semantic_coordinates != mn.semantic_coordinates
+                    or sn.entity_id != mn.entity_id
+                    or abs(sn.importance - mn.importance) > 1e-9
+                    or abs(sn.confidence - mn.confidence) > 1e-9
+                    or abs(sn.created_at - mn.created_at) > 1e-6):
+                bad += 1
+        ok(bad == 0, f"字段级对拍 40 节点零分叉·语义不变字段（bad={bad}）")
+
+        # 全量集合对拍：两种存储物理序不同（sqlite 插入序 vs md 路径字典序），
+        # 首屏 limit 截断集合必不同——对拍对象是全集（差≤1 = 双侧容错粒度差：
+        # 脏行在 sqlite 侧被跳过、在 md 侧以归一形态装载成功）
+        m_all = {n.id for n in m_store.query_nodes(
+            layer=MemoryLayer.KNOWLEDGE, limit=1000000)}
+        ok(set(s_by_id) <= m_all and len(m_all - set(s_by_id)) <= 1,
+           f"query_nodes 全量集合（md {len(m_all)} >= sq {len(s_by_id)}，差<=1）")
+
+        # 双向边对拍（抽 10 节点）
+        bad_e = 0
+        for nid in sorted(inter)[:10]:
+            s_out = {(e.target_id, e.relation_type.value)
+                     for e in sq.store.get_outgoing_edges(nid)}
+            m_out = {(e.target_id, e.relation_type.value)
+                     for e in m_store.get_outgoing_edges(nid)}
+            if s_out != m_out:
+                bad_e += 1
+        ok(bad_e == 0, f"出边域对拍 10 节点零分叉（bad={bad_e}）")
+
+        # 检索对拍：同查询 top5 序列一致（同义词扩展+bigram 打分同构）
+        for q in ("存在论", "条件空间的适用边界", "反馈回路如何稳定系统"):
+            s_top = [n.id for n, _ in sq.store.search_content(
+                q, layers=[MemoryLayer.KNOWLEDGE], limit=5)]
+            m_top = [n.id for n, _ in m_store.search_content(
+                q, layers=[MemoryLayer.KNOWLEDGE], limit=5)]
+            ok(s_top == m_top, f"search_content top5 一致：{q!r}")
+    finally:
+        sq.close()
+        md.close()
+
+
+# ---------------------------------------------------------------------------
+# 组B · md 模式写闭环（tmp 语料根）
+# ---------------------------------------------------------------------------
+
+def group_b(tmp):
+    print("组B · md 模式写闭环")
+    root = os.path.join(tmp, "wb")
+    os.makedirs(root, exist_ok=True)
+    dex = ConditionDex(md_root=root)
+    ok(dex._md_mode and dex.engine is None, "md 模式构造（engine=None）")
+    r = dex.seed_base()
+    ok(r["status"] == "seeded", f"seed_base 首播 {r['status']}")
+
+    kdir = os.path.join(root, "knowledge")
+    import hashlib
+    ids_on_disk = set()
+    if os.path.isdir(kdir):
+        for dp, _dn, fs in os.walk(kdir):
+            for f in fs:
+                if f.endswith(".md"):
+                    ids_on_disk.add(f[:-3])
+    card_ids = {"wis_" + hashlib.sha1(n["name"].encode("utf-8")).hexdigest()[:12]
+                for n in BASE_ENTRIES}
+    missing = card_ids - ids_on_disk
+    ok(not missing,
+       f"26 张知识卡确定性 id 全部落盘（knowledge 层共 "
+       f"{len(ids_on_disk)} 文件=卡+seed 留痕，缺 {len(missing)}）")
+
+    # 重开读回：读写同源（新实例从 md 装载）
+    dex.close()
+    dex2 = ConditionDex(md_root=root)
+    ok(dex2.seed_base()["status"] == "already_seeded", "seed 幂等（重开判重）")
+    seeds = [n for n in dex2.store.query_nodes(
+        layer=MemoryLayer.KNOWLEDGE, limit=300)
+        if "domain:存在论" in (n.tags or [])]
+    ok(len(seeds) >= 1, f"知识层从 md 读回（{len(seeds)} 张 domain:存在论 卡）")
+
+    # add_relation → fm.edges 落盘（schema 与导出语料一致）
+    eid = dex2.add_relation("存在论", "条件论", "causal",
+                            note="条件论以存在论为前提", confidence=0.85)
+    ok(eid.startswith("mde_"), "add_relation 返回确定性边 id")
+    nid_a = dex2._by_name["存在论"]
+    fm_txt = ""
+    for dp, _dn, fs in os.walk(kdir):
+        for f in fs:
+            if not f.endswith(".md"):
+                continue
+            with open(os.path.join(dp, f), encoding="utf-8") as fh:
+                txt = fh.read()
+            if nid_a in txt:
+                fm_txt = txt.split("---")[1]
+                break
+        if fm_txt:
+            break
+    ok(fm_txt and '"relation_type": "causal"' in fm_txt
+       and '"verified": 1' in fm_txt,
+       "fm.edges 落盘含 causal 边（verified=1）")
+
+    # 读写同源：写入即可读（MdStore.reload 生效）
+    edges = dex2.store.get_outgoing_edges(nid_a)
+    ok(any(e.relation_type.value == "causal" and e.verified
+           for e in edges), "写入后 MdStore 即时可读（verified 边）")
+    return dex2
+
+
+# ---------------------------------------------------------------------------
+# 组C · 往返与留痕
+# ---------------------------------------------------------------------------
+
+def group_c(dex2):
+    print("组C · 往返与留痕")
+    cs = ConditionSpace(observation_position="测试台",
+                        observation_tool="test_wisdom_md_store",
+                        time_window=(0.0, 9999999999.0),
+                        existence_constraint="测试用")
+    nid = dex2.add_entry("测试学科X", "存在论", "测试断言：往返一致", cs,
+                         level=2, status="verified")
+    ok(nid.startswith("wis_"), "add_entry md 确定性 id")
+    node = dex2.store.get_node(nid)
+    ok(node is not None and node.state_attributes.get("name") == "测试学科X"
+       and node.layer.value == "knowledge", "写入即可读（state_attributes 往返）")
+
+    rec = dex2._perceive(json.dumps({"type": "测试留痕"}, ensure_ascii=False),
+                         condition_space=cs, importance=0.6,
+                         tags=["观测层", "测试"])
+    ok(len(rec.id) >= 12, "留痕节点返回 id")
+    found = False
+    for dp, _dn, fs in os.walk(os.path.join(dex2.md_root, "knowledge")):
+        if rec.id + ".md" in fs:
+            found = True
+            break
+    ok(found, "留痕落 knowledge 层（对齐 sqlite add_perception 自动进知识层）")
+    # 重开可读（留痕读回）
+    dex2.close()
+    dex3 = ConditionDex(md_root=dex2.md_root)
+    got = dex3.store.get_node(rec.id)
+    ok(got is not None and got.layer.value == "knowledge", "留痕重开可读")
+    return dex3
+
+
+# ---------------------------------------------------------------------------
+# 组D · 兼容回归（sqlite 形态 + duck conn）
+# ---------------------------------------------------------------------------
+
+def group_d(tmp):
+    print("组D · 兼容回归")
+    db = os.path.join(tmp, "legacy.db")
+    dex = ConditionDex(db_path=db)          # 显式 db_path → sqlite 形态
+    ok(not dex._md_mode and dex.engine is not None, "显式 db_path → sqlite 形态")
+    ok(dex.seed_base()["status"] == "seeded", "sqlite 形态 seed 正常")
+    nid = dex.add_entry("旧路学科Y", "条件论", "旧路断言", None, level=2,
+                        status="verified")
+    ok(bool(nid), "sqlite 形态 add_entry 正常")
+    ok(bool(dex.add_relation("旧路学科Y", "条件论", "similar")),
+       "sqlite 形态建边正常")
+    dex.close()
+
+    dex2 = ConditionDex(fresh=True, db_path=os.path.join(tmp, "fresh.db"))
+    ok(not dex2._md_mode, "fresh=True → sqlite 形态（兼容 api.py 兜底）")
+    dex2.close()
+
+    if corpus_ready(DEFAULT_MD_ROOT):
+        dex3 = ConditionDex(md_root=DEFAULT_MD_ROOT)
+        try:
+            conn = read_conn(dex3)           # duck：MdStore.conn → MdConn
+            ok(isinstance(conn, MdConn), "read_conn(dex) md 模式返回 MdConn")
+            rows = conn._all_rows()
+            ok(len(rows) > 0, f"duck conn 可查询（{len(rows)} 行）")
+        finally:
+            dex3.close()
+
+
+def main():
+    ok(len(BASE_ENTRIES) >= 20, f"BASE_ENTRIES {len(BASE_ENTRIES)} 在位")
+    ok(len(BASE_RELATIONS) >= 20, f"BASE_RELATIONS {len(BASE_RELATIONS)} 在位")
+    tmp = tempfile.mkdtemp(prefix="wb_md_test_")
+    try:
+        group_a(tmp)
+        dex2 = group_b(tmp)
+        dex3 = group_c(dex2)
+        dex3.close()
+        group_d(tmp)
+    finally:
+        shutil.rmtree(tmp, ignore_errors=True)
+    print(f"\nALL PASS · {_N} assertions")
+
+
+if __name__ == "__main__":
+    main()