@furongjun1999/dsh-memory 0.4.7 → 0.4.9

package/md_cg/test_retr_s3.py

@@ -0,0 +1,194 @@
+# -*- coding: utf-8 -*-
+"""S3 图扩散激活探针（脚本式，exit code 定成败）。
+
+用法：python -X utf8 -m md_cg.test_retr_s3
+覆盖：
+- 默认关：不落 gates 键、结果与开启前后（关→开→关）一致
+- S3 开：沿 edges 双向扩散出候选（TIER_SPREAD），seeds/expanded 审计计数正确
+- hops 控制扩散半径；decay 控制跳数衰减（A > B > C 分数序）
+- 未过 S2 门控的邻居不得被扩散引入
+- 无 edges → expanded=0/reason=no_edges，并落回原 T2/T3 路径（召回不丢）
+"""
+import os
+import sys
+import tempfile
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from md_cg.mdcg import MdCG, TIER_SPREAD  # noqa: E402
+
+passed = 0
+failed = 0
+
+
+def check(name, cond, detail=""):
+    global passed, failed
+    if cond:
+        passed += 1
+        print("  [PASS] " + name)
+    else:
+        failed += 1
+        print("  [FAIL] " + name + "  " + detail)
+
+
+def _setenv(**kv):
+    old = {}
+    for k, v in kv.items():
+        old[k] = os.environ.get(k)
+        if v is None:
+            os.environ.pop(k, None)
+        else:
+            os.environ[k] = v
+    return old
+
+
+def _restore(old):
+    for k, v in old.items():
+        if v is None:
+            os.environ.pop(k, None)
+        else:
+            os.environ[k] = v
+
+
+def _build(root, edges=True, pos_b=None):
+    cg = MdCG(root)
+    cg.add("A", "工程 结构 应力 梁 截面", "knowledge", importance=0.5,
+           condition_space={"observation_position": "工程 结构"})
+    cg.add("B", "qxzq 与检索词无交集", "knowledge", importance=0.5,
+           condition_space={"observation_position": pos_b or "工程 结构"})
+    cg.add("C", "ppyx 另一段无交集文本", "knowledge", importance=0.5,
+           condition_space={"observation_position": "工程 结构"})
+    cg.add("D", "孤立节点 无任何边", "knowledge", importance=0.5,
+           condition_space={"observation_position": "工程 结构"})
+    if edges:
+        cg.append_edge("A", {"target": "B", "relation_type": "related"})
+        cg.append_edge("B", {"target": "C", "relation_type": "related"})
+    cg.flush()
+    return cg
+
+
+def _score_of(results, nid):
+    for r in results:
+        if r[0]["id"] == nid:
+            return r[1]
+    return None
+
+
+def main():
+    root = tempfile.mkdtemp(prefix="retr_s3_")
+    cg = _build(root)
+
+    # ---- 1) 默认关 ----
+    old = _setenv(MDCG_RETRIEVAL_PIPELINE=None, MDCG_GATE_S1_DOMAIN=None,
+                  MDCG_GATE_S2_COND=None, MDCG_GATE_S3_SPREAD=None)
+    r_off, meta_off = cg.search("工程 应力", k=5, judge=False, record=False)
+    check("默认关：meta 不含 gates 键", "gates" not in meta_off,
+          str(sorted(meta_off.keys())))
+    snap_off = [(r[0]["id"], round(float(r[1]), 6)) for r in r_off]
+
+    # 1b 总开关开、S1/S2 显式关、S3 子开关未设 → S3 必须仍为关（显式 =1 才生效）
+    _setenv(MDCG_RETRIEVAL_PIPELINE="1", MDCG_GATE_S1_DOMAIN="0",
+            MDCG_GATE_S2_COND="0", MDCG_GATE_S3_SPREAD=None)
+    r_m, m_m = cg.search("工程 应力", k=5, judge=False, record=False)
+    check("S3 子开关未设：不生效（无 gates / 结果不变）",
+          "gates" not in m_m
+          and [(r[0]["id"], round(float(r[1]), 6)) for r in r_m] == snap_off,
+          str(m_m.get("gates")) + " " + str([r[0]["id"] for r in r_m]))
+
+    # ---- 2) S3 开（S1/S2 关），hops=1 ----
+    _setenv(MDCG_RETRIEVAL_PIPELINE="1", MDCG_GATE_S1_DOMAIN="0",
+            MDCG_GATE_S2_COND="0", MDCG_GATE_S3_SPREAD="1",
+            MDCG_SPREAD_HOPS="1", MDCG_SPREAD_DECAY="0.5",
+            MDCG_SPREAD_GAIN="0.2")
+    r1, m1 = cg.search("工程 应力", k=5, judge=False, record=False)
+    s3 = (m1.get("gates") or {}).get("s3") or {}
+    check("S3 命中种子=1", s3.get("seeds") == 1, str(s3))
+    check("S3 hops=1 扩散出 1 个邻居", s3.get("expanded") == 1, str(s3))
+    check("S3 tier 为扩散层", m1.get("tier") == TIER_SPREAD, str(m1.get("tier")))
+    check("S3 邻居 B 进入结果", _score_of(r1, "B") is not None,
+          str([r[0]["id"] for r in r1]))
+    check("S3 未连通的 D 不进入", _score_of(r1, "D") is None,
+          str([r[0]["id"] for r in r1]))
+
+    # ---- 3) hops=2 扩散到 C，且衰减有序 A > B > C ----
+    _setenv(MDCG_SPREAD_HOPS="2")
+    r2, m2 = cg.search("工程 应力", k=5, judge=False, record=False)
+    s3 = (m2.get("gates") or {}).get("s3") or {}
+    check("S3 hops=2 扩散出 2 个", s3.get("expanded") == 2, str(s3))
+    sa, sb, sc = _score_of(r2, "A"), _score_of(r2, "B"), _score_of(r2, "C")
+    check("S3 衰减有序 A>B>C",
+          sa is not None and sb is not None and sc is not None
+          and sa > sb > sc, "A=%s B=%s C=%s" % (sa, sb, sc))
+
+    # ---- 4) 未过 S2 门控的邻居不得被扩散引入（部分被拦，避免全滤回退掩盖不变量）----
+    rroot = tempfile.mkdtemp(prefix="retr_s3g_")
+    rg = MdCG(rroot)
+    rg.add("A", "工程 结构 应力", "knowledge",
+           condition_space={"observation_position": "工程 结构"})
+    rg.add("B", "qxzq 无交集", "knowledge",
+           condition_space={"observation_position": "医学 疾病"})
+    rg.append_edge("A", {"target": "B", "relation_type": "related"})
+    rg.flush()
+    _setenv(MDCG_GATE_S2_COND="1")
+    r3, m3 = rg.search("工程 应力", k=5, judge=False, record=False,
+                       context={"observation_position": "工程 结构"})
+    g3 = m3.get("gates") or {}
+    ids3 = [r[0]["id"] for r in r3]
+    check("S2 确实拦掉了 B", (g3.get("s2") or {}).get("dropped") == 1,
+          str(g3.get("s2")))
+    check("扩散不得引入被 S2 拦掉的邻居 B",
+          "B" not in ids3 and (g3.get("s3") or {}).get("expanded", 0) == 0,
+          str(ids3) + " " + str(g3.get("s3")))
+    _setenv(MDCG_GATE_S2_COND="0")
+
+    # ---- 5) 无 edges → 不扩散且落回原路径 ----
+    root2 = tempfile.mkdtemp(prefix="retr_s3n_")
+    cg2 = _build(root2, edges=False)
+    r4, m4 = cg2.search("工程 应力", k=5, judge=False, record=False)
+    s3 = (m4.get("gates") or {}).get("s3") or {}
+    check("无 edges：expanded=0 且 reason=no_edges",
+          s3.get("expanded") == 0 and s3.get("reason") == "no_edges", str(s3))
+    check("无 edges：落回原路径且召回不丢",
+          _score_of(r4, "A") is not None and m4.get("tier") != TIER_SPREAD,
+          str(m4.get("tier")))
+
+    # ---- 5b) 幂等：重复检索 / 索引重建后同口径 ----
+    _setenv(MDCG_RETRIEVAL_PIPELINE="1", MDCG_GATE_S1_DOMAIN="0",
+            MDCG_GATE_S2_COND="0", MDCG_GATE_S3_SPREAD="1",
+            MDCG_SPREAD_HOPS="2", MDCG_SPREAD_DECAY="0.5",
+            MDCG_SPREAD_GAIN="0.2")
+    _snap = lambda rs: [(r[0]["id"], round(float(r[1]), 6)) for r in rs]
+    ra, ma = cg.search("工程 应力", k=5, judge=False, record=False)
+    rb, mb = cg.search("工程 应力", k=5, judge=False, record=False)
+    check("S3 幂等：重复检索同口径",
+          _snap(ra) == _snap(rb)
+          and (ma.get("gates") or {}) == (mb.get("gates") or {}),
+          str(_snap(ra)) + " vs " + str(_snap(rb)))
+    cg.rebuild_index()
+    rc, mc = cg.search("工程 应力", k=5, judge=False, record=False)
+    check("S3 幂等：索引重建后同口径",
+          _snap(rc) == _snap(ra)
+          and (mc.get("gates") or {}) == (ma.get("gates") or {}),
+          str(_snap(rc)) + " vs " + str(_snap(ra)))
+
+    # ---- 6) 默认口径等价（关→开→关）----
+    _setenv(MDCG_RETRIEVAL_PIPELINE=None, MDCG_GATE_S1_DOMAIN=None,
+            MDCG_GATE_S2_COND=None, MDCG_GATE_S3_SPREAD=None,
+            MDCG_SPREAD_HOPS=None, MDCG_SPREAD_DECAY=None,
+            MDCG_SPREAD_GAIN=None)
+    r_off2, meta_off2 = cg.search("工程 应力", k=5, judge=False, record=False)
+    snap_off2 = [(r[0]["id"], round(float(r[1]), 6)) for r in r_off2]
+    check("默认口径等价：关→开→关 首末一致", snap_off == snap_off2,
+          str(snap_off) + " vs " + str(snap_off2))
+    check("默认关：meta 仍不含 gates 键", "gates" not in meta_off2,
+          str(sorted(meta_off2.keys())))
+    _restore(old)
+
+    print("\ntest_retr_s3: %d 通过 / %d 失败" % (passed, failed))
+    return 0 if failed == 0 else 1
+
+
+if __name__ == "__main__":
+    if hasattr(sys.stdout, "reconfigure"):
+        sys.stdout.reconfigure(encoding="utf-8")
+    sys.exit(main())

package/md_cg/test_retr_s4.py

@@ -0,0 +1,163 @@
+# -*- coding: utf-8 -*-
+"""S4 层级激活优先级探针（脚本式，exit code 定成败）。
+
+用法：python -X utf8 -m md_cg.test_retr_s4
+覆盖：
+- layer_boosts 解析（默认副本 / 部分覆盖 / 非法项忽略 / 负值夹 0）
+- 默认关：无 gates 键；关→开→关 结果一致
+- 子开关未设（总开关开）→ S4 不生效
+- 层级加成生效：同内容不同层 → anchor/self > structural > knowledge > contextual
+- 自定义 MDCG_LAYER_BOOST 改变次序；分数上限仍 1.0
+- 只改排序不拉回被门控剔除的节点；幂等（重复检索 / 索引重建）
+"""
+import os
+import sys
+import tempfile
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from md_cg.mdcg import MdCG, layer_boosts, LAYER_BOOST_DEFAULT  # noqa: E402
+
+passed = 0
+failed = 0
+TEXT = "阿尔法 贝塔 伽马 德尔塔"
+# 查询只与正文部分重叠（sim≈0.5），否则精确匹配会把分数顶到 1.0、层级加成看不出来
+QUERY = "阿尔法 泽塔"
+
+
+def check(name, cond, detail=""):
+    global passed, failed
+    if cond:
+        passed += 1
+        print("  [PASS] " + name)
+    else:
+        failed += 1
+        print("  [FAIL] " + name + "  " + detail)
+
+
+def _setenv(**kv):
+    old = {}
+    for k, v in kv.items():
+        old[k] = os.environ.get(k)
+        if v is None:
+            os.environ.pop(k, None)
+        else:
+            os.environ[k] = v
+    return old
+
+
+def _restore(old):
+    for k, v in old.items():
+        if v is None:
+            os.environ.pop(k, None)
+        else:
+            os.environ[k] = v
+
+
+def _scores(results):
+    return {r[0]["id"]: round(float(r[1]), 6) for r in results}
+
+
+def main():
+    # ---- 1) 解析器 ----
+    check("默认表副本（不共享引用）",
+          layer_boosts("") == LAYER_BOOST_DEFAULT
+          and layer_boosts("") is not LAYER_BOOST_DEFAULT)
+    bo = layer_boosts("anchor=0.3,zzz=1,contextual=abc,knowledge=-5")
+    check("部分覆盖 + 非法项忽略 + 负值夹 0",
+          bo["anchor"] == 0.3 and bo["contextual"] == LAYER_BOOST_DEFAULT["contextual"]
+          and bo["knowledge"] == 0.0 and "zzz" in bo, str(bo))
+
+    # ---- 2) 建库：同内容、不同层 ----
+    root = tempfile.mkdtemp(prefix="retr_s4_")
+    cg = MdCG(root)
+    for layer in ("anchor", "self", "structural", "knowledge", "contextual"):
+        cg.add("n_" + layer, TEXT, layer)
+    cg.flush()
+
+    old = _setenv(MDCG_RETRIEVAL_PIPELINE=None, MDCG_GATE_S4_LAYER=None,
+                  MDCG_LAYER_BOOST=None)
+    r_off, m_off = cg.search(QUERY, k=10, judge=False, record=False)
+    snap_off = _scores(r_off)
+    check("默认关：meta 不含 gates 键", "gates" not in m_off,
+          str(sorted(m_off.keys())))
+    check("默认关：同内容各层分数相同（无加成）",
+          len(set(snap_off.values())) == 1 and 0 < max(snap_off.values()) < 1.0,
+          str(snap_off))
+
+    # 2b 子开关未设（总开关开、S1/S2 显式关）→ S4 不生效
+    _setenv(MDCG_RETRIEVAL_PIPELINE="1", MDCG_GATE_S4_LAYER=None,
+            MDCG_LAYER_BOOST=None, MDCG_GATE_S1_DOMAIN="0",
+            MDCG_GATE_S2_COND="0")
+    r_x, m_x = cg.search(QUERY, k=10, judge=False, record=False)
+    check("子开关未设：S4 不生效",
+          "s4" not in (m_x.get("gates") or {}) and _scores(r_x) == snap_off,
+          str(m_x.get("gates")))
+
+    # ---- 3) S4 开：层级次序 ----
+    _setenv(MDCG_RETRIEVAL_PIPELINE="1", MDCG_GATE_S4_LAYER="1",
+            MDCG_LAYER_BOOST=None, MDCG_GATE_S1_DOMAIN="0",
+            MDCG_GATE_S2_COND="0")
+    r4, m4 = cg.search(QUERY, k=10, judge=False, record=False)
+    sc = _scores(r4)
+    g4 = (m4.get("gates") or {}).get("s4") or {}
+    check("S4 审计含 boosts/layers", bool(g4.get("boosts")) and bool(g4.get("layers")),
+          str(g4)[:160])
+    check("anchor/self > structural > knowledge > contextual",
+          sc["n_anchor"] > sc["n_structural"] > sc["n_knowledge"] > sc["n_contextual"]
+          and sc["n_self"] == sc["n_anchor"],
+          str(sc))
+    check("加成上限：分数不超过 1.0", max(sc.values()) <= 1.0 + 1e-9, str(sc))
+
+    # ---- 4) 自定义加成表改变次序 ----
+    _setenv(MDCG_LAYER_BOOST="contextual=0.5,knowledge=0.0,anchor=0.0,self=0.0,structural=0.0")
+    r5, _m5 = cg.search(QUERY, k=10, judge=False, record=False)
+    sc5 = _scores(r5)
+    check("自定义加成：contextual 升到首位",
+          sc5["n_contextual"] > sc5["n_anchor"], str(sc5))
+
+    # ---- 5) 只改排序，不拉回被门控剔除的节点 ----
+    _setenv(MDCG_LAYER_BOOST=None, MDCG_GATE_S2_COND="1",
+            MDCG_GATE_S1_DOMAIN="0")
+    # 注意：本段临时改了 S1/S2，出段后必须还原，否则幂等段的 gates 不可比
+    root2 = tempfile.mkdtemp(prefix="retr_s4g_")
+    cgg = MdCG(root2)
+    cgg.add("g_ok", TEXT, "knowledge",
+            condition_space={"observation_position": "工程 结构"})
+    cgg.add("g_no", TEXT, "anchor",
+            condition_space={"observation_position": "医学 疾病"})
+    cgg.flush()
+    r6, m6 = cgg.search("阿尔法", k=10, judge=False, record=False,
+                        context={"observation_position": "工程 结构"})
+    ids6 = [r[0]["id"] for r in r6]
+    check("S2 拦掉的节点不因 S4 加成回归",
+          "g_no" not in ids6 and "g_ok" in ids6, str(ids6) + " " + str(m6.get("gates")))
+    _setenv(MDCG_GATE_S2_COND=None, MDCG_GATE_S1_DOMAIN="0")
+
+    # ---- 6) 幂等：重复检索 / 索引重建 ----
+    r7, m7 = cg.search(QUERY, k=10, judge=False, record=False)
+    check("S4 幂等：重复检索同口径",
+          _scores(r7) == sc and (m7.get("gates") or {}) == (m4.get("gates") or {}),
+          str(_scores(r7)))
+    cg.rebuild_index()
+    r8, m8 = cg.search(QUERY, k=10, judge=False, record=False)
+    check("S4 幂等：索引重建后同口径",
+          _scores(r8) == sc and (m8.get("gates") or {}) == (m4.get("gates") or {}),
+          str(_scores(r8)))
+
+    # ---- 7) 默认口径等价（关→开→关）----
+    _setenv(MDCG_RETRIEVAL_PIPELINE=None, MDCG_GATE_S4_LAYER=None,
+            MDCG_LAYER_BOOST=None, MDCG_GATE_S2_COND=None, MDCG_GATE_S1_DOMAIN=None)
+    r9, m9 = cg.search(QUERY, k=10, judge=False, record=False)
+    check("默认口径等价：关→开→关 首末一致",
+          _scores(r9) == snap_off and "gates" not in m9, str(_scores(r9)))
+    _restore(old)
+
+    print("\ntest_retr_s4: %d 通过 / %d 失败" % (passed, failed))
+    return 0 if failed == 0 else 1
+
+
+if __name__ == "__main__":
+    if hasattr(sys.stdout, "reconfigure"):
+        sys.stdout.reconfigure(encoding="utf-8")
+    sys.exit(main())

package/md_cg/test_retr_s5.py

@@ -0,0 +1,200 @@
+# -*- coding: utf-8 -*-
+"""S5 负记忆抑制探针（脚本式，exit code 定成败）。
+
+用法：python -X utf8 -m md_cg.test_retr_s5
+覆盖：
+- 默认关：无 gates 键；关→开→关 结果一致
+- 子开关未设（总开关开）→ S5 不生效
+- 重叠抑制：与 rejected 节点词面高度重叠的正候选 → w × (1-λ)
+- 相邻抑制：与 rejected 节点有边相连（文本不重叠）的正候选同样被抑制
+- 无关候选不被抑制；λ 可配（0=不抑制，1=压到 0 且不低于 0）
+- 阈值真的在过滤：部分重叠候选在 thr=0.6 下不抑制、thr=0.4 下被抑制
+- 负记忆条目在 S5 开时不再与正候选同权（1.0 → 1-λ）
+- 无负记忆命中时不落 s5 审计；幂等（重复检索 / 索引重建）
+"""
+import os
+import sys
+import tempfile
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from md_cg.mdcg import MdCG  # noqa: E402
+
+passed = 0
+failed = 0
+TEXT = "阿尔法 贝塔 伽马 德尔塔"
+# neg 文本必须与 pos_overlap 高度重叠（jaccard ≥ 0.5），否则重叠支路测不出来
+NEGTEXT = "阿尔法 贝塔 伽马 德尔塔 否决样本"
+QUERY = "阿尔法 泽塔"
+
+
+def check(name, cond, detail=""):
+    global passed, failed
+    if cond:
+        passed += 1
+        print("  [PASS] " + name)
+    else:
+        failed += 1
+        print("  [FAIL] " + name + "  " + detail)
+
+
+def _setenv(**kv):
+    old = {}
+    for k, v in kv.items():
+        old[k] = os.environ.get(k)
+        if v is None:
+            os.environ.pop(k, None)
+        else:
+            os.environ[k] = v
+    return old
+
+
+def _restore(old):
+    for k, v in old.items():
+        if v is None:
+            os.environ.pop(k, None)
+        else:
+            os.environ[k] = v
+
+
+def _scores(results):
+    return {r[0]["id"]: round(float(r[1]), 6) for r in results}
+
+
+def _build(root):
+    cg = MdCG(root)
+    # 三个正候选都含查询词（都会成为 T2 命中），差别只在「与负记忆的关系」：
+    #   pos_overlap：词面高度重叠（重叠支路）
+    #   pos_edge   ：文本不重叠但有边指向负记忆（相邻支路）
+    #   pos_clean  ：既无重叠也无边（不应被抑制）
+    cg.add("pos_overlap", TEXT, "knowledge")
+    cg.add("pos_edge", "泽塔 量子 比特 纠缠", "knowledge")
+    cg.add("pos_clean", "泽塔 海流 洋流 潮汐", "knowledge")
+    #   pos_mid    ：与负记忆部分重叠（包含度≈0.5）→ 用来验证阈值真的在过滤
+    cg.add("pos_mid", "阿尔法 贝塔 海流 洋流", "knowledge")
+    cg.add("h_rej", NEGTEXT, "rejected")
+    cg.append_edge("pos_edge", {"target": "h_rej", "relation_type": "related"})
+    cg.flush()
+    return cg
+
+
+def main():
+    root = tempfile.mkdtemp(prefix="retr_s5_")
+    cg = _build(root)
+
+    # ---- 1) 默认关 ----
+    old = _setenv(MDCG_RETRIEVAL_PIPELINE=None, MDCG_GATE_S5_NEG=None,
+                  MDCG_NEG_LAMBDA=None, MDCG_NEG_SIM=None,
+                  MDCG_GATE_S1_DOMAIN=None, MDCG_GATE_S2_COND=None)
+    r_off, m_off = cg.search(QUERY, k=10, judge=False, record=False)
+    base = _scores(r_off)
+    check("默认关：meta 不含 gates 键", "gates" not in m_off,
+          str(sorted(m_off.keys())))
+    check("默认关：负记忆条目与正候选同权（1.0）",
+          base.get("rejected/h_rej.md") == 1.0, str(base))
+
+    # ---- 2) 子开关未设（总开关开、S1/S2 显式关）→ S5 不生效 ----
+    _setenv(MDCG_RETRIEVAL_PIPELINE="1", MDCG_GATE_S5_NEG=None,
+            MDCG_GATE_S1_DOMAIN="0", MDCG_GATE_S2_COND="0")
+    r_x, m_x = cg.search(QUERY, k=10, judge=False, record=False)
+    check("子开关未设：S5 不生效",
+          "s5" not in (m_x.get("gates") or {}) and _scores(r_x) == base,
+          str(m_x.get("gates")))
+
+    # ---- 3) S5 开：重叠 + 相邻抑制，无关候选不动 ----
+    _setenv(MDCG_GATE_S5_NEG="1", MDCG_NEG_LAMBDA="0.5", MDCG_NEG_SIM="0.5")
+    r5, m5 = cg.search(QUERY, k=10, judge=False, record=False)
+    sc5 = _scores(r5)
+    g5 = (m5.get("gates") or {}).get("s5") or {}
+    check("S5 审计：λ/threshold/抑制数",
+          g5.get("lambda") == 0.5 and g5.get("threshold") == 0.5
+          and g5.get("suppressed", 0) >= 2, str(g5))
+    _sup = set(g5.get("suppressed_ids") or [])
+    check("重叠候选被抑制（×0.5）",
+          abs(sc5.get("pos_overlap", -1) - base["pos_overlap"] * 0.5) < 1e-9
+          and "pos_overlap" in _sup,
+          "%s vs %s  sup=%s" % (sc5.get("pos_overlap"), base.get("pos_overlap"), sorted(_sup)))
+    check("相邻候选被抑制（文本不重叠也生效）",
+          abs(sc5.get("pos_edge", -1) - base["pos_edge"] * 0.5) < 1e-9
+          and "pos_edge" in _sup,
+          "%s vs %s  sup=%s" % (sc5.get("pos_edge"), base.get("pos_edge"), sorted(_sup)))
+    check("无关候选不被抑制",
+          sc5.get("pos_clean") == base.get("pos_clean")
+          and "pos_clean" not in _sup,
+          "%s vs %s  sup=%s" % (sc5.get("pos_clean"), base.get("pos_clean"), sorted(_sup)))
+    check("负记忆条目不再同权（1.0 → 0.5）",
+          sc5.get("rejected/h_rej.md") == 0.5, str(sc5.get("rejected/h_rej.md")))
+
+    # ---- 4) λ 可配：0=不抑制；1=压到 0 且 ≥0 ----
+    _setenv(MDCG_NEG_LAMBDA="0")
+    r0, _m0 = cg.search(QUERY, k=10, judge=False, record=False)
+    check("λ=0：不抑制", _scores(r0) == base, str(_scores(r0)))
+    _setenv(MDCG_NEG_LAMBDA="1")
+    r1, _m1 = cg.search(QUERY, k=10, judge=False, record=False)
+    sc1 = _scores(r1)
+    check("λ=1：压到 0 且不为负",
+          sc1["pos_overlap"] == 0.0 and sc1["pos_edge"] == 0.0
+          and min(sc1.values()) >= 0.0, str(sc1))
+
+    # ---- 5) 阈值真的在过滤：部分重叠候选（包含度≈0.5）在 0.6 下不抑制、0.4 下被抑制 ----
+    _setenv(MDCG_NEG_LAMBDA="0.5", MDCG_NEG_SIM="0.6")
+    r2, m2 = cg.search(QUERY, k=10, judge=False, record=False)
+    sc2 = _scores(r2)
+    sup2 = set(((m2.get("gates") or {}).get("s5") or {}).get("suppressed_ids") or [])
+    check("thr=0.6：部分重叠不抑制、完全重叠仍抑制",
+          sc2["pos_mid"] == base["pos_mid"]
+          and "pos_mid" not in sup2 and "pos_overlap" in sup2,
+          str(sc2) + " " + str(sorted(sup2)))
+    _setenv(MDCG_NEG_SIM="0.4")
+    r3, m3 = cg.search(QUERY, k=10, judge=False, record=False)
+    sc3 = _scores(r3)
+    sup3 = set(((m3.get("gates") or {}).get("s5") or {}).get("suppressed_ids") or [])
+    check("thr=0.4：部分重叠也被抑制",
+          abs(sc3["pos_mid"] - base["pos_mid"] * 0.5) < 1e-9
+          and "pos_mid" in sup3,
+          str(sc3) + " " + str(sorted(sup3)))
+
+    # ---- 5b) 配置的 λ 必须真的生效（含审计）----
+    _setenv(MDCG_NEG_LAMBDA="0.25", MDCG_NEG_SIM="0.5")
+    r4, m4 = cg.search(QUERY, k=10, judge=False, record=False)
+    sc4 = _scores(r4)
+    g4 = (m4.get("gates") or {}).get("s5") or {}
+    check("λ=0.25：负记忆条目 0.75 且审计记 0.25",
+          sc4.get("rejected/h_rej.md") == 0.75 and g4.get("lambda") == 0.25,
+          "%s %s" % (sc4.get("rejected/h_rej.md"), g4))
+
+    # ---- 6) 无负记忆命中：不落 s5 审计 ----
+    r3, m3 = cg.search("西格玛 欧米伽", k=10, judge=False, record=False)
+    check("无负记忆命中：不落 s5 审计",
+          "s5" not in (m3.get("gates") or {}), str(m3.get("gates")))
+
+    # ---- 7) 幂等：重复检索 / 索引重建 ----
+    _setenv(MDCG_NEG_LAMBDA="0.5", MDCG_NEG_SIM="0.5")
+    ra, ma = cg.search(QUERY, k=10, judge=False, record=False)
+    rb, mb = cg.search(QUERY, k=10, judge=False, record=False)
+    check("S5 幂等：重复检索同口径",
+          _scores(ra) == _scores(rb)
+          and (ma.get("gates") or {}) == (mb.get("gates") or {}), str(_scores(ra)))
+    cg.rebuild_index()
+    rc, mc = cg.search(QUERY, k=10, judge=False, record=False)
+    check("S5 幂等：索引重建后同口径",
+          _scores(rc) == _scores(ra)
+          and (mc.get("gates") or {}) == (ma.get("gates") or {}), str(_scores(rc)))
+
+    # ---- 8) 默认口径等价（关→开→关）----
+    _setenv(MDCG_RETRIEVAL_PIPELINE=None, MDCG_GATE_S5_NEG=None,
+            MDCG_NEG_LAMBDA=None, MDCG_NEG_SIM=None,
+            MDCG_GATE_S1_DOMAIN=None, MDCG_GATE_S2_COND=None)
+    r9, m9 = cg.search(QUERY, k=10, judge=False, record=False)
+    check("默认口径等价：关→开→关 首末一致",
+          _scores(r9) == base and "gates" not in m9, str(_scores(r9)))
+    _restore(old)
+
+    print("\ntest_retr_s5: %d 通过 / %d 失败" % (passed, failed))
+    return 0 if failed == 0 else 1
+
+
+if __name__ == "__main__":
+    if hasattr(sys.stdout, "reconfigure"):
+        sys.stdout.reconfigure(encoding="utf-8")
+    sys.exit(main())