@furongjun1999/dsh-memory 0.4.11 → 0.5.0

package/md_cg/test_retr_s8_time.py

@@ -1,316 +1,369 @@
-# -*- coding: utf-8 -*-
-"""S8 时间算子探针（阶段二 4.1，脚本式，exit code 定成败）。
-
-用法：python -X utf8 -m md_cg.test_retr_s8_time
-覆盖：
-- 纯函数：轴归一与封闭枚举、按轴取窗、重叠 vs 端点模式、入参 fail-closed
-- 候选层过滤：效力轴 fail-open（缺字段=沉默不是否认）/ 观察轴 fail-closed
-  （缺字段=数据异常，须计数可见）
-- meta["time_filter"] 审计块；默认关 meta 键集合不变
-- 绕开热路径缓存（**读+写双侧**）：缓存键不含时间参数，读侧失守会返回未过滤
-  结果、写侧失守会把过滤结果覆盖进缓存让后续默认查询少结果，两者都是静默错答
-- stg 四 op 轴感知（observed 缺省逐位兼容 / effective 生效 / 非法轴 fail-closed）
-- recall 透传
-
-语料 15 节点（同一正文 token「阿尔法」，保证都被召回，差异只在时间字段）：
-  效力组 7：eff_a[100,200] eff_b[300,400] eff_c[150,350] eff_none(无端点)
-            eff_halffrom[500,-] eff_halfuntil[-,50] eff_d[1000,1100]
-  观察组 7：obs_1000/2000/3000/8000/9000(temporal) obs_win_4k_5k/6k_7k(time_window)
-  旧库 1：legacy_noaxis（手工抹掉 temporal 与 time_window → 观察轴不可判定）
-"""
-import glob
-import os
-import sys
-import tempfile
-
-sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
-
-from md_cg import hotcache, nodefile, stg, trust          # noqa: E402
-from md_cg.mdcg import MdCG                               # noqa: E402
-from md_cg.mdcos import MdCGSecure                        # noqa: E402
-from md_cg.security import Principal                      # noqa: E402
-
-passed = 0
-failed = 0
-TEXT = "阿尔法 时间 探针"
-QUERY = "阿尔法"
-#: 效力轴上**无字段**的节点（按效力轴 fail-open 策略保留）：eff_none + 观察组 7 + 旧库 1
-_NO_EFF_AXIS = {"eff_none", "legacy_noaxis",
-                "obs_1000", "obs_2000", "obs_3000", "obs_t_8000", "obs_t_9000",
-                "obs_win_4k_5k", "obs_win_6k_7k"}
-
-
-def check(name, cond, detail=""):
-    global passed, failed
-    if cond:
-        passed += 1
-        print("  [PASS] " + name)
-    else:
-        failed += 1
-        print("  [FAIL] " + name + "  " + detail)
-
-
-def _ids(results):
-    return [r[0]["id"] for r in results]
-
-
-def _node_file(root, nid):
-    hits = glob.glob(os.path.join(root, "**", nid + ".md"), recursive=True)
-    return hits[0] if hits else None
-
-
-def _build(root):
-    cg = MdCG(root)
-    for nid, kw in (
-        ("eff_a", {"effective_from": 100, "effective_until": 200}),
-        ("eff_b", {"effective_from": 300, "effective_until": 400}),
-        ("eff_c", {"effective_from": 150, "effective_until": 350}),
-        ("eff_none", {}),
-        ("eff_halffrom", {"effective_from": 500}),
-        ("eff_halfuntil", {"effective_until": 50}),
-        ("eff_d", {"effective_from": 1000, "effective_until": 1100}),
-        ("obs_1000", {"temporal": 1000.0}),
-        ("obs_2000", {"temporal": 2000.0}),
-        ("obs_3000", {"temporal": 3000.0}),
-        ("obs_win_4k_5k", {"condition_space": {"time_window": [4000, 5000]}}),
-        ("obs_win_6k_7k", {"condition_space": {"time_window": [6000, 7000]}}),
-        ("obs_t_8000", {"temporal": 8000.0}),
-        ("obs_t_9000", {"temporal": 9000.0}),
-        ("legacy_noaxis", {}),
-    ):
-        cg.add(nid, TEXT, "knowledge", **kw)
-    cg.flush()
-    # 旧库形态：手工抹掉观察轴两字段（add 会强制补 time_window，只能改文件模拟）
-    p = _node_file(root, "legacy_noaxis")
-    fm, content = nodefile.loads(open(p, encoding="utf-8").read())
-    fm.pop("temporal", None)
-    cs = dict(fm.get("condition_space") or {})
-    cs.pop("time_window", None)
-    fm["condition_space"] = cs
-    open(p, "w", encoding="utf-8").write(nodefile.dumps(fm, content))
-    cg.rebuild_index()
-    return cg
-
-
-def main():
-    # ================= 1) 纯函数层 =================
-    check("轴归一：None → effective；大小写归一；believed 抛 ValueError",
-          trust.time_axis_of(None) == "effective"
-          and trust.time_axis_of("OBSERVED") == "observed"
-          and trust.time_axis_of("effective") == "effective"
-          and _raises(ValueError, lambda: trust.time_axis_of("believed")),
-          str(trust.TIME_AXES))
-
-    check("重叠模式：有交 True / 无交 False / 候选两端不可解析 False",
-          trust.window_match(10, 20, 15, 25) is True
-          and trust.window_match(10, 20, 30, 40) is False
-          and trust.window_match(None, None, 1, 2) is False)
-
-    check("端点模式：B1 缺省 gte/lte；单端算子不误杀全空",
-          trust.window_match(10, 20, 5, 25, "gte", "lte") is True
-          and trust.window_match(10, 20, 15, 25, "gte", "lte") is False
-          and trust.window_match(10, 20, 5, None, "gte", None) is True
-          and trust.window_match(3, 20, 5, None, "gte", None) is False)
-
-    check("入参 fail-closed 四则：未启用 / 孤算子 / 逆序 / 非法轴",
-          trust.check_time_args()[:2] == (False, None)
-          and trust.check_time_args(start_operator="gte")[2] != ""
-          and trust.check_time_args(start_time=300, end_time=100)[2] != ""
-          and trust.check_time_args(start_time=100, time_axis="believed")[2] != "")
-
-    check("按轴取窗：观察轴 temporal 优先于 time_window；效力轴别名回落 valid_from",
-          trust.time_window_of({"temporal": 7.0,
-                                "condition_space": {"time_window": [1, 2]}},
-                               "observed") == (7.0, 7.0)
-          and trust.time_window_of({"valid_from": 3, "valid_until": 9},
-                                   "effective") == (3, 9)
-          and trust.time_window_of({}, "effective") == (None, None))
-
-    _kept, _dr, _ms = trust.filter_by_time(
-        [{"effective_from": 1, "effective_until": 2}, {}], "effective", 5, 9)
-    _kept2, _dr2, _ms2 = trust.filter_by_time([{}], "observed", 5, 9)
-    check("轴策略：效力轴缺字段 fail-open 保留；观察轴缺字段 fail-closed 计入 missing",
-          _dr == 1 and _ms == 0 and _dr2 == 1 and _ms2 == 1)
-
-    # ================= 2) 建库 =================
-    root = tempfile.mkdtemp(prefix="retr_s8_")
-    cg = _build(root)
-
-    # 默认关（无时间入参）基线
-    # 注：热路径接入点是 `MdCGSecure.search_rrf`，`MdCG.search` **无缓存**——
-    # 故「cached 标记」不能在此处取证（否则恒真的弱断言），见第 9 节。
-    r0, m0 = cg.search(QUERY, k=20, judge=False, record=False)
-    base = sorted(_ids(r0))
-    r0b, m0b = cg.search(QUERY, k=20, judge=False, record=False)
-    check("默认关：meta 无 time_filter、候选 15 条、二次同口径可复算",
-          "time_filter" not in m0 and len(base) == 15
-          and sorted(_ids(r0b)) == base,
-          "n=%d keys=%s" % (len(base), sorted(m0.keys())))
-
-    # ================= 3) 效力轴（重叠模式）=================
-    r1, m1 = cg.search(QUERY, k=20, judge=False, record=False,
-                       start_time=150, end_time=250, time_axis="effective")
-    ids1 = set(_ids(r1))
-    tf1 = m1.get("time_filter") or {}
-    check("效力轴重叠：窗内相交保留、窗外剔除、无端点 fail-open 保留",
-          {"eff_a", "eff_c", "eff_none"} <= ids1
-          and not ({"eff_b", "eff_halffrom", "eff_halfuntil", "eff_d"} & ids1)
-          and len(ids1) == 11,
-          str(sorted(ids1)))
-    check("效力轴：审计块 axis/mode/applied 正确、dropped 与存活数记账自洽",
-          tf1.get("axis") == "effective" and tf1.get("mode") == "overlap"
-          and tf1.get("applied") is True and tf1.get("dropped") == 4
-          and tf1.get("dropped") + len(ids1) == 15
-          and tf1.get("start") == 150 and tf1.get("end") == 250,
-          str(tf1))
-
-    # ================= 4) 观察轴（窗口语义 + fail-closed）=================
-    r2, m2 = cg.search(QUERY, k=20, judge=False, record=False,
-                       start_time=2500, end_time=3500, time_axis="observed")
-    ids2 = _ids(r2)
-    tf2 = m2.get("time_filter") or {}
-    check("观察轴：temporal 命中唯一、效力组按观测窗不相交剔除",
-          ids2 == ["obs_3000"], str(ids2))
-    check("观察轴：无字段节点 fail-closed 并计入 axis_missing",
-          tf2.get("axis_missing") == 1 and tf2.get("dropped") == 14
-          and tf2.get("dropped") + len(ids2) == 15, str(tf2))
-
-    # ================= 5) 端点模式（单端算子）=================
-    r3, m3 = cg.search(QUERY, k=20, judge=False, record=False,
-                       start_time=150, start_operator="gte",
-                       time_axis="effective")
-    ids3 = set(_ids(r3))
-    tf3 = m3.get("time_filter") or {}
-    # 只给起点算子 → 只筛起点侧：eff_a（起点 100<150）与 eff_halfuntil（起点缺失、
-    # 端点模式下该侧不可比较）被剔除；eff_none 与全部观察组/旧库节点在效力轴**无字段**
-    # → fail-open 保留（沉默不是否认）。此处非空集本身即旧缺陷（单端算子恒空）的回归证据。
-    check("端点模式：单端 gte 只筛起点侧且不返回空集、无该轴字段者 fail-open 保留",
-          ids3 == {"eff_b", "eff_c", "eff_d", "eff_halffrom"} | _NO_EFF_AXIS
-          and tf3.get("mode") == "endpoint"
-          and tf3.get("start_operator") == "gte"
-          and tf3.get("dropped") == 2, str(sorted(ids3)))
-
-    # ================= 6) 误用 fail-closed =================
-    check("误用三则经库层抛 ValueError（孤算子 / 逆序 / 非法轴）",
-          _raises(ValueError, lambda: cg.search(
-              QUERY, k=5, judge=False, record=False, start_operator="gte"))
-          and _raises(ValueError, lambda: cg.search(
-              QUERY, k=5, judge=False, record=False, start_time=300, end_time=100))
-          and _raises(ValueError, lambda: cg.search(
-              QUERY, k=5, judge=False, record=False, start_time=100,
-              time_axis="believed")))
-
-    # ================= 7) stg 四 op 轴感知 =================
-    tl_obs = stg.timeline(cg, limit=100, time_axis="observed")
-    tl_eff = stg.timeline(cg, limit=100, time_axis="effective")
-    check("stg timeline：观察轴 14 条（旧库节点不可判定被跳过）、效力轴 4 条（两端齐备）",
-          tl_obs["count"] == 14 and tl_eff["count"] == 4,
-          "obs=%s eff=%s" % (tl_obs["count"], tl_eff["count"]))
-    check("stg timeline：缺省轴与显式 observed 逐位一致（旧行为兼容）",
-          stg.timeline(cg, limit=100)["count"] == tl_obs["count"])
-
-    # 返回结构取证：relation 的时间结果在 ["time"]["relation"]（非顶层 time_relation）
-    rel_eff = stg.relation(cg, "eff_a", "eff_b", time_axis="effective")
-    rel_obs = stg.relation(cg, "eff_a", "eff_b", time_axis="observed")
-    check("stg relation：效力轴取 effective 区间（eff_a before eff_b）；观察轴取写入观测时刻（区间不同）",
-          rel_eff["time"]["relation"] == "before"
-          and rel_eff["time"]["a"] == (100.0, 200.0)
-          and rel_obs["time"]["a"] != (100.0, 200.0),
-          "%s / %s" % (rel_eff["time"], rel_obs["time"]))
-
-    an_eff = stg.anchors(cg, time_window=[150, 250], limit=50,
-                         time_axis="effective")
-    an_obs = stg.anchors(cg, time_window=[150, 250], limit=50,
-                         time_axis="observed")
-    check("stg anchors：效力轴命中 2（eff_a overlaps + eff_c contains）、观察轴 0",
-          an_eff["count"] == 2 and an_obs["count"] == 0,
-          "eff=%s obs=%s" % (an_eff["count"], an_obs["count"]))
-
-    check("stg 非法轴 fail-closed（四 op 同源 trust.time_axis_of）",
-          _raises(ValueError, lambda: stg.consistency(cg, time_axis="believed"))
-          and _raises(ValueError, lambda: stg.relation(
-              cg, "eff_a", "eff_b", time_axis="believed")))
-
-    # ================= 8) 关→开→关 默认口径不变 =================
-    # 置于 Secure 层操作**之前**：那些调用会记录访问（recall 无法关 record），
-    # 可能与基线产生与时间算子无关的差异，污染「首末一致」这一断言。
-    r5, m5 = cg.search(QUERY, k=20, judge=False, record=False)
-    check("默认口径等价：关→开→关 首末一致且无 time_filter 键",
-          sorted(_ids(r5)) == base and "time_filter" not in m5)
-
-    # ================= 9) 热缓存：命中 + 时间算子读/写双侧绕行 =================
-    sec = MdCGSecure(root, principal=Principal(
-        actor="test-s8", clearance="secret", can_write=True,
-        role="designer", auth_mode="local-cli"))
-    hotcache.attach(sec)
-    _sa, sma = sec.search_rrf(QUERY, k=20, judge=False, record=False)
-    _sb, smb = sec.search_rrf(QUERY, k=20, judge=False, record=False)
-    check("热缓存：同参二次 search_rrf 命中（cached=True）且结果一致",
-          sma.get("cached") is not True and smb.get("cached") is True
-          and [r[0]["id"] for r in _sa] == [r[0]["id"] for r in _sb],
-          "a=%s b=%s" % (sma.get("cached"), smb.get("cached")))
-
-    # 上一步已把 15 条全量结果写进缓存键（键不含时间参数）。带时间算子的查询
-    # 若只绕开「读」而漏掉「写」，则返回的 11 条会把缓存覆盖成子集 —— 故下方
-    # 「随后默认查询仍为 15 条」是**写侧绕行**的判别性证据。
-    _t1, mt1 = sec.search_rrf(QUERY, k=20, judge=False, record=False,
-                              start_time=150, end_time=250,
-                              time_axis="effective")
-    check("时间算子启用时绕开热缓存：缓存已有 15 条全量，仍返回过滤后 11 条、meta 无 cached",
-          "cached" not in mt1 and len(_t1) == 11
-          and sorted(r[0]["id"] for r in _t1) == sorted(ids1),
-          "n=%d cached=%s" % (len(_t1), mt1.get("cached")))
-    _d2, md2 = sec.search_rrf(QUERY, k=20, judge=False, record=False)
-    check("时间算子查询不污染缓存（写侧同受绕行约束）：其后默认查询仍为全量 15 条",
-          len(_d2) == 15 and md2.get("cached") is True,
-          "n=%d cached=%s" % (len(_d2), md2.get("cached")))
-
-    # ================= 10) recall 透传 =================
-    out = sec.recall(QUERY, k=20, judge=False, start_time=2500,
-                     end_time=3500, time_axis="observed")
-    mr = out.get("meta") or {}
-    ids_r = [e["id"] for e in (out.get("pack") or [])]
-    check("recall 透传：meta 带 time_filter（观察轴）且过滤真实生效",
-          (mr.get("time_filter") or {}).get("axis") == "observed"
-          and ids_r == ["obs_3000"], str(ids_r))
-
-    # ================= 11) Secure.search 常规路径审计块（MCP cg(op=read) 实走此路）=================
-    # 判别性：本节的观测量=「过滤生效 ∧ meta 带审计块」**同时**成立，缺一即 FAIL。
-    # 缺陷史（2026-09-20 取证）：候选层过滤已在 `_candidates` 完成（结果正确），但
-    # `MdCGSecure.search` 的**常规返回路径**未把 `_time_filter_stat` 交给 `_emit` ——
-    # 审计块此前只在 no_candidates 分支落键 → 「过滤确实生效而审计不可见」（静默，
-    # 调用方无法区分「未启用」与「启用但候选全 fail-open 保留」）。
-    # 只用 `cg.search`（父类，本无此缺口）或 `search_rrf`（另行落键）**均测不到**该缺口，
-    # 故单列一节：本节的红/绿随该唯一缺口翻转，是它的判别性守卫。
-    _s11, m11 = sec.search(QUERY, k=20, judge=False, record=False,
-                           start_time=150, end_time=250, time_axis="effective")
-    tf11 = m11.get("time_filter") or {}
-    check("Secure.search 常规路径：过滤生效且 meta 带 time_filter（审计块不被吞）",
-          sorted(r[0]["id"] for r in _s11) == sorted(ids1)
-          and tf11.get("axis") == "effective" and tf11.get("mode") == "overlap"
-          and tf11.get("applied") is True and tf11.get("dropped") == 4,
-          "n=%d tf=%s" % (len(_s11), tf11))
-
-    _s11b, m11b = sec.search(QUERY, k=20, judge=False, record=False)
-    check("Secure.search 默认关：候选不变且 meta 键集合不含 time_filter（默认关零变更）",
-          "time_filter" not in m11b and len(_s11b) == 15,
-          "n=%d keys=%s" % (len(_s11b), sorted(m11b.keys())))
-
-    print("\ntest_retr_s8: %d 通过 / %d 失败" % (passed, failed))
-    return 0 if failed == 0 else 1
-
-
-def _raises(exc, fn):
-    try:
-        fn()
-    except exc:
-        return True
-    except Exception:
-        return False
-    return False
-
-
-if __name__ == "__main__":
-    if hasattr(sys.stdout, "reconfigure"):
-        sys.stdout.reconfigure(encoding="utf-8")
-    sys.exit(main())
+# -*- coding: utf-8 -*-
+"""S8 时间算子探针（阶段二 4.1，脚本式，exit code 定成败）。
+
+用法：python -X utf8 -m md_cg.test_retr_s8_time
+覆盖：
+- 纯函数：轴归一与封闭枚举、按轴取窗、重叠 vs 端点模式、入参 fail-closed
+- 候选层过滤：效力轴 fail-open（缺字段=沉默不是否认）/ 观察轴 fail-closed
+  （缺字段=数据异常，须计数可见）
+- meta["time_filter"] 审计块；默认关 meta 键集合不变
+- 绕开热路径缓存（**读+写双侧**）：缓存键不含时间参数，读侧失守会返回未过滤
+  结果、写侧失守会把过滤结果覆盖进缓存让后续默认查询少结果，两者都是静默错答
+- stg 四 op 轴感知（observed 缺省逐位兼容 / effective 生效 / 非法轴 fail-closed）
+- 单位归一（issue #23）：`trust.epoch_seconds` 唯一实现点；毫秒存量不改盘即可参与
+  timeline 排序与 S2 时间门控（旧行为下 1.6e12 恒排最前、秒窗被误判不相交）
+- recall 透传
+
+语料 15 节点（同一正文 token「阿尔法」，保证都被召回，差异只在时间字段）：
+  效力组 7：eff_a[100,200] eff_b[300,400] eff_c[150,350] eff_none(无端点)
+            eff_halffrom[500,-] eff_halfuntil[-,50] eff_d[1000,1100]
+  观察组 7：obs_1000/2000/3000/8000/9000(temporal) obs_win_4k_5k/6k_7k(time_window)
+  旧库 1：legacy_noaxis（手工抹掉 temporal 与 time_window → 观察轴不可判定）
+"""
+import glob
+import os
+import sys
+import tempfile
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from md_cg import hotcache, nodefile, stg, trust          # noqa: E402
+from md_cg.mdcg import MdCG                               # noqa: E402
+from md_cg.mdcos import MdCGSecure                        # noqa: E402
+from md_cg.security import Principal                      # noqa: E402
+
+passed = 0
+failed = 0
+TEXT = "阿尔法 时间 探针"
+QUERY = "阿尔法"
+#: 效力轴上**无字段**的节点（按效力轴 fail-open 策略保留）：eff_none + 观察组 7 + 旧库 1
+_NO_EFF_AXIS = {"eff_none", "legacy_noaxis",
+                "obs_1000", "obs_2000", "obs_3000", "obs_t_8000", "obs_t_9000",
+                "obs_win_4k_5k", "obs_win_6k_7k"}
+
+
+def check(name, cond, detail=""):
+    global passed, failed
+    if cond:
+        passed += 1
+        print("  [PASS] " + name)
+    else:
+        failed += 1
+        print("  [FAIL] " + name + "  " + detail)
+
+
+def _ids(results):
+    return [r[0]["id"] for r in results]
+
+
+def _node_file(root, nid):
+    hits = glob.glob(os.path.join(root, "**", nid + ".md"), recursive=True)
+    return hits[0] if hits else None
+
+
+def _build(root):
+    cg = MdCG(root)
+    for nid, kw in (
+        ("eff_a", {"effective_from": 100, "effective_until": 200}),
+        ("eff_b", {"effective_from": 300, "effective_until": 400}),
+        ("eff_c", {"effective_from": 150, "effective_until": 350}),
+        ("eff_none", {}),
+        ("eff_halffrom", {"effective_from": 500}),
+        ("eff_halfuntil", {"effective_until": 50}),
+        ("eff_d", {"effective_from": 1000, "effective_until": 1100}),
+        ("obs_1000", {"temporal": 1000.0}),
+        ("obs_2000", {"temporal": 2000.0}),
+        ("obs_3000", {"temporal": 3000.0}),
+        ("obs_win_4k_5k", {"condition_space": {"time_window": [4000, 5000]}}),
+        ("obs_win_6k_7k", {"condition_space": {"time_window": [6000, 7000]}}),
+        ("obs_t_8000", {"temporal": 8000.0}),
+        ("obs_t_9000", {"temporal": 9000.0}),
+        ("legacy_noaxis", {}),
+    ):
+        cg.add(nid, TEXT, "knowledge", **kw)
+    cg.flush()
+    # 旧库形态：手工抹掉观察轴两字段（add 会强制补 time_window，只能改文件模拟）
+    p = _node_file(root, "legacy_noaxis")
+    fm, content = nodefile.loads(open(p, encoding="utf-8").read())
+    fm.pop("temporal", None)
+    cs = dict(fm.get("condition_space") or {})
+    cs.pop("time_window", None)
+    fm["condition_space"] = cs
+    open(p, "w", encoding="utf-8").write(nodefile.dumps(fm, content))
+    cg.rebuild_index()
+    return cg
+
+
+def main():
+    # ================= 1) 纯函数层 =================
+    check("轴归一：None → effective；大小写归一；believed 抛 ValueError",
+          trust.time_axis_of(None) == "effective"
+          and trust.time_axis_of("OBSERVED") == "observed"
+          and trust.time_axis_of("effective") == "effective"
+          and _raises(ValueError, lambda: trust.time_axis_of("believed")),
+          str(trust.TIME_AXES))
+
+    check("重叠模式：有交 True / 无交 False / 候选两端不可解析 False",
+          trust.window_match(10, 20, 15, 25) is True
+          and trust.window_match(10, 20, 30, 40) is False
+          and trust.window_match(None, None, 1, 2) is False)
+
+    check("端点模式：B1 缺省 gte/lte；单端算子不误杀全空",
+          trust.window_match(10, 20, 5, 25, "gte", "lte") is True
+          and trust.window_match(10, 20, 15, 25, "gte", "lte") is False
+          and trust.window_match(10, 20, 5, None, "gte", None) is True
+          and trust.window_match(3, 20, 5, None, "gte", None) is False)
+
+    check("入参 fail-closed 四则：未启用 / 孤算子 / 逆序 / 非法轴",
+          trust.check_time_args()[:2] == (False, None)
+          and trust.check_time_args(start_operator="gte")[2] != ""
+          and trust.check_time_args(start_time=300, end_time=100)[2] != ""
+          and trust.check_time_args(start_time=100, time_axis="believed")[2] != "")
+
+    check("按轴取窗：观察轴 temporal 优先于 time_window；效力轴别名回落 valid_from",
+          trust.time_window_of({"temporal": 7.0,
+                                "condition_space": {"time_window": [1, 2]}},
+                               "observed") == (7.0, 7.0)
+          and trust.time_window_of({"valid_from": 3, "valid_until": 9},
+                                   "effective") == (3, 9)
+          and trust.time_window_of({}, "effective") == (None, None))
+
+    # 单位归一（issue #23）：唯一实现点 trust.epoch_seconds —— 秒过秒、毫秒 /1000、
+    # 全时窗哨兵与非数值不误判。阈值 1e11 的判别力全靠「哨兵与秒值都在阈下」这一条。
+    check("epoch_seconds：秒直通 / 毫秒归一 / 哨兵与非数值不误判",
+          trust.epoch_seconds(1700000000.0) == 1700000000.0
+          and trust.epoch_seconds(1700000000000) == 1700000000.0
+          and trust.epoch_seconds(1600003600000.0) == 1600003600.0
+          and trust.epoch_seconds(nodefile.FULL_TIME_WINDOW_MAX) == \
+          nodefile.FULL_TIME_WINDOW_MAX
+          and trust.epoch_seconds(None) is None
+          and trust.epoch_seconds("1700000000") is None,
+          str(trust._MS_EPOCH_THRESHOLD))
+
+    check("parse_time：数字与数字串两条路径都归一，且秒值直通不回退",
+          trust.parse_time(1700000000000) == 1700000000.0
+          and trust.parse_time("1700000000000") == 1700000000.0
+          and trust.parse_time(1700000000.0) == 1700000000.0
+          and trust.parse_time("1700000000") == 1700000000.0
+          and trust.parse_time(None) is None
+          and trust.parse_time("2020-01-01") is not None)
+
+    _kept, _dr, _ms = trust.filter_by_time(
+        [{"effective_from": 1, "effective_until": 2}, {}], "effective", 5, 9)
+    _kept2, _dr2, _ms2 = trust.filter_by_time([{}], "observed", 5, 9)
+    check("轴策略：效力轴缺字段 fail-open 保留；观察轴缺字段 fail-closed 计入 missing",
+          _dr == 1 and _ms == 0 and _dr2 == 1 and _ms2 == 1)
+
+    # ================= 2) 建库 =================
+    root = tempfile.mkdtemp(prefix="retr_s8_")
+    cg = _build(root)
+
+    # 默认关（无时间入参）基线
+    # 注：热路径接入点是 `MdCGSecure.search_rrf`，`MdCG.search` **无缓存**——
+    # 故「cached 标记」不能在此处取证（否则恒真的弱断言），见第 9 节。
+    r0, m0 = cg.search(QUERY, k=20, judge=False, record=False)
+    base = sorted(_ids(r0))
+    r0b, m0b = cg.search(QUERY, k=20, judge=False, record=False)
+    check("默认关：meta 无 time_filter、候选 15 条、二次同口径可复算",
+          "time_filter" not in m0 and len(base) == 15
+          and sorted(_ids(r0b)) == base,
+          "n=%d keys=%s" % (len(base), sorted(m0.keys())))
+
+    # ================= 3) 效力轴（重叠模式）=================
+    r1, m1 = cg.search(QUERY, k=20, judge=False, record=False,
+                       start_time=150, end_time=250, time_axis="effective")
+    ids1 = set(_ids(r1))
+    tf1 = m1.get("time_filter") or {}
+    check("效力轴重叠：窗内相交保留、窗外剔除、无端点 fail-open 保留",
+          {"eff_a", "eff_c", "eff_none"} <= ids1
+          and not ({"eff_b", "eff_halffrom", "eff_halfuntil", "eff_d"} & ids1)
+          and len(ids1) == 11,
+          str(sorted(ids1)))
+    check("效力轴：审计块 axis/mode/applied 正确、dropped 与存活数记账自洽",
+          tf1.get("axis") == "effective" and tf1.get("mode") == "overlap"
+          and tf1.get("applied") is True and tf1.get("dropped") == 4
+          and tf1.get("dropped") + len(ids1) == 15
+          and tf1.get("start") == 150 and tf1.get("end") == 250,
+          str(tf1))
+
+    # ================= 4) 观察轴（窗口语义 + fail-closed）=================
+    r2, m2 = cg.search(QUERY, k=20, judge=False, record=False,
+                       start_time=2500, end_time=3500, time_axis="observed")
+    ids2 = _ids(r2)
+    tf2 = m2.get("time_filter") or {}
+    check("观察轴：temporal 命中唯一、效力组按观测窗不相交剔除",
+          ids2 == ["obs_3000"], str(ids2))
+    check("观察轴：无字段节点 fail-closed 并计入 axis_missing",
+          tf2.get("axis_missing") == 1 and tf2.get("dropped") == 14
+          and tf2.get("dropped") + len(ids2) == 15, str(tf2))
+
+    # ================= 5) 端点模式（单端算子）=================
+    r3, m3 = cg.search(QUERY, k=20, judge=False, record=False,
+                       start_time=150, start_operator="gte",
+                       time_axis="effective")
+    ids3 = set(_ids(r3))
+    tf3 = m3.get("time_filter") or {}
+    # 只给起点算子 → 只筛起点侧：eff_a（起点 100<150）与 eff_halfuntil（起点缺失、
+    # 端点模式下该侧不可比较）被剔除；eff_none 与全部观察组/旧库节点在效力轴**无字段**
+    # → fail-open 保留（沉默不是否认）。此处非空集本身即旧缺陷（单端算子恒空）的回归证据。
+    check("端点模式：单端 gte 只筛起点侧且不返回空集、无该轴字段者 fail-open 保留",
+          ids3 == {"eff_b", "eff_c", "eff_d", "eff_halffrom"} | _NO_EFF_AXIS
+          and tf3.get("mode") == "endpoint"
+          and tf3.get("start_operator") == "gte"
+          and tf3.get("dropped") == 2, str(sorted(ids3)))
+
+    # ================= 6) 误用 fail-closed =================
+    check("误用三则经库层抛 ValueError（孤算子 / 逆序 / 非法轴）",
+          _raises(ValueError, lambda: cg.search(
+              QUERY, k=5, judge=False, record=False, start_operator="gte"))
+          and _raises(ValueError, lambda: cg.search(
+              QUERY, k=5, judge=False, record=False, start_time=300, end_time=100))
+          and _raises(ValueError, lambda: cg.search(
+              QUERY, k=5, judge=False, record=False, start_time=100,
+              time_axis="believed")))
+
+    # ================= 7) stg 四 op 轴感知 =================
+    tl_obs = stg.timeline(cg, limit=100, time_axis="observed")
+    tl_eff = stg.timeline(cg, limit=100, time_axis="effective")
+    check("stg timeline：观察轴 14 条（旧库节点不可判定被跳过）、效力轴 4 条（两端齐备）",
+          tl_obs["count"] == 14 and tl_eff["count"] == 4,
+          "obs=%s eff=%s" % (tl_obs["count"], tl_eff["count"]))
+    check("stg timeline：缺省轴与显式 observed 逐位一致（旧行为兼容）",
+          stg.timeline(cg, limit=100)["count"] == tl_obs["count"])
+
+    # 返回结构取证：relation 的时间结果在 ["time"]["relation"]（非顶层 time_relation）
+    rel_eff = stg.relation(cg, "eff_a", "eff_b", time_axis="effective")
+    rel_obs = stg.relation(cg, "eff_a", "eff_b", time_axis="observed")
+    check("stg relation：效力轴取 effective 区间（eff_a before eff_b）；观察轴取写入观测时刻（区间不同）",
+          rel_eff["time"]["relation"] == "before"
+          and rel_eff["time"]["a"] == (100.0, 200.0)
+          and rel_obs["time"]["a"] != (100.0, 200.0),
+          "%s / %s" % (rel_eff["time"], rel_obs["time"]))
+
+    an_eff = stg.anchors(cg, time_window=[150, 250], limit=50,
+                         time_axis="effective")
+    an_obs = stg.anchors(cg, time_window=[150, 250], limit=50,
+                         time_axis="observed")
+    check("stg anchors：效力轴命中 2（eff_a overlaps + eff_c contains）、观察轴 0",
+          an_eff["count"] == 2 and an_obs["count"] == 0,
+          "eff=%s obs=%s" % (an_eff["count"], an_obs["count"]))
+
+    check("stg 非法轴 fail-closed（四 op 同源 trust.time_axis_of）",
+          _raises(ValueError, lambda: stg.consistency(cg, time_axis="believed"))
+          and _raises(ValueError, lambda: stg.relation(
+              cg, "eff_a", "eff_b", time_axis="believed")))
+
+    # ========== 7b) 单位归一：毫秒存量不改盘即可正确排序（issue #23）==========
+    # 用**独立小库**取证，不动上面 15 节点语料（其计数断言是既有契约）。
+    # 复现旧版 sources 摄取出的存量节点形态：time_window 落 13 位毫秒。
+    ms_root = tempfile.mkdtemp(prefix="retr_s8_ms_")
+    ms_cg = MdCG(ms_root)
+    ms_cg.add("ms_old", TEXT, "knowledge",
+              condition_space={"time_window": [1600000000000, 1600003600000]})
+    ms_cg.add("sec_new", TEXT, "knowledge",
+              condition_space={"time_window": [1700000000, 1700003600]})
+    ms_cg.flush()
+    ms_cg.rebuild_index()
+    check("毫秒存量先决条件：索引快照确实落 13 位毫秒窗（否则本组断言退化为空转）",
+          ((ms_cg.index["nodes"].get("ms_old") or {}).get("time_window")
+           or [None])[0] == 1600000000000,
+          str((ms_cg.index["nodes"].get("ms_old") or {}).get("time_window")))
+    _msiv = stg._interval({"condition_space":
+                           {"time_window": [1600000000000, 1600003600000]}})
+    check("单位归一：_interval 与 trust.time_window_of 同口径（毫秒 /1000）",
+          _msiv == (1600000000.0, 1600003600.0), str(_msiv))
+    _mtl = stg.timeline(ms_cg, limit=10)
+    check("单位归一：毫秒存量不再顶到 timeline 头部（旧行为 1.6e12 恒排最前）",
+          [(i["id"], i["start"]) for i in _mtl["items"]] ==
+          [("sec_new", 1700000000.0), ("ms_old", 1600000000.0)],
+          str([(i["id"], i["start"]) for i in _mtl["items"]]))
+    from md_cg.mdcg import _cond_prefilter_pass
+    check("单位归一：S2 时间门控对毫秒节点不再误剔除（同根因）",
+          _cond_prefilter_pass({"time_window": [1600000000000, 1600003600000]},
+                               {"time_window": [1600000010, 1600000020]}) is True
+          and _cond_prefilter_pass({"time_window": [1600000000000, 1600003600000]},
+                                   {"time_window": [1700000000, 1700000020]}) is False)
+
+    # ================= 8) 关→开→关 默认口径不变 =================
+    # 置于 Secure 层操作**之前**：那些调用会记录访问（recall 无法关 record），
+    # 可能与基线产生与时间算子无关的差异，污染「首末一致」这一断言。
+    r5, m5 = cg.search(QUERY, k=20, judge=False, record=False)
+    check("默认口径等价：关→开→关 首末一致且无 time_filter 键",
+          sorted(_ids(r5)) == base and "time_filter" not in m5)
+
+    # ================= 9) 热缓存：命中 + 时间算子读/写双侧绕行 =================
+    sec = MdCGSecure(root, principal=Principal(
+        actor="test-s8", clearance="secret", can_write=True,
+        role="designer", auth_mode="local-cli"))
+    hotcache.attach(sec)
+    _sa, sma = sec.search_rrf(QUERY, k=20, judge=False, record=False)
+    _sb, smb = sec.search_rrf(QUERY, k=20, judge=False, record=False)
+    check("热缓存：同参二次 search_rrf 命中（cached=True）且结果一致",
+          sma.get("cached") is not True and smb.get("cached") is True
+          and [r[0]["id"] for r in _sa] == [r[0]["id"] for r in _sb],
+          "a=%s b=%s" % (sma.get("cached"), smb.get("cached")))
+
+    # 上一步已把 15 条全量结果写进缓存键（键不含时间参数）。带时间算子的查询
+    # 若只绕开「读」而漏掉「写」，则返回的 11 条会把缓存覆盖成子集 —— 故下方
+    # 「随后默认查询仍为 15 条」是**写侧绕行**的判别性证据。
+    _t1, mt1 = sec.search_rrf(QUERY, k=20, judge=False, record=False,
+                              start_time=150, end_time=250,
+                              time_axis="effective")
+    check("时间算子启用时绕开热缓存：缓存已有 15 条全量，仍返回过滤后 11 条、meta 无 cached",
+          "cached" not in mt1 and len(_t1) == 11
+          and sorted(r[0]["id"] for r in _t1) == sorted(ids1),
+          "n=%d cached=%s" % (len(_t1), mt1.get("cached")))
+    _d2, md2 = sec.search_rrf(QUERY, k=20, judge=False, record=False)
+    check("时间算子查询不污染缓存（写侧同受绕行约束）：其后默认查询仍为全量 15 条",
+          len(_d2) == 15 and md2.get("cached") is True,
+          "n=%d cached=%s" % (len(_d2), md2.get("cached")))
+
+    # ================= 10) recall 透传 =================
+    out = sec.recall(QUERY, k=20, judge=False, start_time=2500,
+                     end_time=3500, time_axis="observed")
+    mr = out.get("meta") or {}
+    ids_r = [e["id"] for e in (out.get("pack") or [])]
+    check("recall 透传：meta 带 time_filter（观察轴）且过滤真实生效",
+          (mr.get("time_filter") or {}).get("axis") == "observed"
+          and ids_r == ["obs_3000"], str(ids_r))
+
+    # ================= 11) Secure.search 常规路径审计块（MCP cg(op=read) 实走此路）=================
+    # 判别性：本节的观测量=「过滤生效 ∧ meta 带审计块」**同时**成立，缺一即 FAIL。
+    # 缺陷史（2026-09-20 取证）：候选层过滤已在 `_candidates` 完成（结果正确），但
+    # `MdCGSecure.search` 的**常规返回路径**未把 `_time_filter_stat` 交给 `_emit` ——
+    # 审计块此前只在 no_candidates 分支落键 → 「过滤确实生效而审计不可见」（静默，
+    # 调用方无法区分「未启用」与「启用但候选全 fail-open 保留」）。
+    # 只用 `cg.search`（父类，本无此缺口）或 `search_rrf`（另行落键）**均测不到**该缺口，
+    # 故单列一节：本节的红/绿随该唯一缺口翻转，是它的判别性守卫。
+    _s11, m11 = sec.search(QUERY, k=20, judge=False, record=False,
+                           start_time=150, end_time=250, time_axis="effective")
+    tf11 = m11.get("time_filter") or {}
+    check("Secure.search 常规路径：过滤生效且 meta 带 time_filter（审计块不被吞）",
+          sorted(r[0]["id"] for r in _s11) == sorted(ids1)
+          and tf11.get("axis") == "effective" and tf11.get("mode") == "overlap"
+          and tf11.get("applied") is True and tf11.get("dropped") == 4,
+          "n=%d tf=%s" % (len(_s11), tf11))
+
+    _s11b, m11b = sec.search(QUERY, k=20, judge=False, record=False)
+    check("Secure.search 默认关：候选不变且 meta 键集合不含 time_filter（默认关零变更）",
+          "time_filter" not in m11b and len(_s11b) == 15,
+          "n=%d keys=%s" % (len(_s11b), sorted(m11b.keys())))
+
+    print("\ntest_retr_s8: %d 通过 / %d 失败" % (passed, failed))
+    return 0 if failed == 0 else 1
+
+
+def _raises(exc, fn):
+    try:
+        fn()
+    except exc:
+        return True
+    except Exception:
+        return False
+    return False
+
+
+if __name__ == "__main__":
+    if hasattr(sys.stdout, "reconfigure"):
+        sys.stdout.reconfigure(encoding="utf-8")
+    sys.exit(main())