@furongjun1999/dsh-memory 0.9.23 → 0.10.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (94) hide show
  1. package/CHANGELOG.en.md +450 -404
  2. package/CHANGELOG.md +67 -0
  3. package/README.en.md +244 -240
  4. package/README.md +272 -262
  5. package/codebuddy/CODEBUDDY.md +1 -1
  6. package/docs/discipline/templates/full.md.tmpl +1 -1
  7. package/docs/discipline/templates/rules.mdc.tmpl +1 -1
  8. package/docs/en/mdcg/world-model-endpoint.md +1 -1
  9. package/docs/mdcg/cogmap_ops.md +16 -15
  10. package/lib/client.js +29 -3
  11. package/lib/hooks.d.ts +12 -0
  12. package/lib/hooks.js +43 -1
  13. package/md_cg/bucketindex.py +231 -0
  14. package/md_cg/bucketscope.py +339 -0
  15. package/md_cg/config_registry_bulk.py +19 -10
  16. package/md_cg/config_validate.py +14 -2
  17. package/md_cg/detrerank.py +137 -0
  18. package/md_cg/fsutil.py +135 -0
  19. package/md_cg/govern/__init__.py +23 -0
  20. package/md_cg/iface/__init__.py +46 -0
  21. package/md_cg/iface/_verify.py +609 -0
  22. package/md_cg/iface/contracts.py +337 -0
  23. package/md_cg/judgment_manifest.py +14 -0
  24. package/md_cg/kernel/__init__.py +27 -0
  25. package/md_cg/mcp_server.py +5991 -5761
  26. package/md_cg/mdcg.py +474 -16
  27. package/md_cg/mdcos.py +228 -11
  28. package/md_cg/metacognition.py +66 -13
  29. package/md_cg/mutate_extract_shipped.py +126 -0
  30. package/md_cg/mutate_ledger_e2e.py +100 -0
  31. package/md_cg/op_table.py +8 -0
  32. package/md_cg/probe_c_ledger_integrity.py +258 -0
  33. package/md_cg/probe_d_retirement_discipline.py +263 -0
  34. package/md_cg/review_cli.py +6 -1
  35. package/md_cg/run_tests.py +30 -1
  36. package/md_cg/scopeselect.py +214 -0
  37. package/md_cg/scrub.py +66 -9
  38. package/md_cg/state_extract.py +149 -0
  39. package/md_cg/state_slots.py +58 -10
  40. package/md_cg/structboost.py +179 -0
  41. package/md_cg/test_alternatives_historical.py +162 -0
  42. package/md_cg/test_b2_sensitivity_landing.py +226 -223
  43. package/md_cg/test_b3_merge_keeps_content.py +608 -583
  44. package/md_cg/test_bucket_cards.py +85 -0
  45. package/md_cg/test_bucket_day_order.py +74 -0
  46. package/md_cg/test_bucket_quota.py +79 -0
  47. package/md_cg/test_bucket_recency_guards.py +75 -0
  48. package/md_cg/test_bucket_three_causes.py +83 -0
  49. package/md_cg/test_bucketindex.py +130 -0
  50. package/md_cg/test_bucketscope_early.py +71 -0
  51. package/md_cg/test_cond_term_fold_gate.py +1210 -0
  52. package/md_cg/test_cond_term_neg_scope.py +147 -0
  53. package/md_cg/test_detrerank.py +157 -0
  54. package/md_cg/test_gate_four_kpi.py +136 -0
  55. package/md_cg/test_generation_guard.py +14 -2
  56. package/md_cg/test_issue192_move_layer_indices_guard.py +48 -0
  57. package/md_cg/test_issue39_utf8_stdio.py +572 -572
  58. package/md_cg/test_issue413_preimage_failclosed.py +147 -9
  59. package/md_cg/test_issue768_overwrite_time_window_anchor.py +135 -0
  60. package/md_cg/test_issue789_log_plaintext_probe.py +100 -0
  61. package/md_cg/test_issue790_verb_negation.py +77 -0
  62. package/md_cg/test_issue810_root_guard.py +71 -0
  63. package/md_cg/test_issue812_info_index_stale.py +78 -0
  64. package/md_cg/test_issue827_weaken_dedup.py +361 -0
  65. package/md_cg/test_issue829_calibration_positive_gate.py +473 -0
  66. package/md_cg/test_kernel_iface_contracts.py +18 -0
  67. package/md_cg/test_md_cg_no_hive_dep.py +238 -17
  68. package/md_cg/test_probe_cd_shipped.py +385 -0
  69. package/md_cg/test_recent_tail.py +93 -0
  70. package/md_cg/test_run_tests_entry_gates.py +110 -0
  71. package/md_cg/test_scopeselect.py +70 -0
  72. package/md_cg/test_scopeselect_clock.py +70 -0
  73. package/md_cg/test_session_attribution_consistency.py +835 -813
  74. package/md_cg/test_state_extract_shipped.py +162 -0
  75. package/md_cg/test_state_ledger_chain.py +207 -0
  76. package/md_cg/test_state_ledger_e2e.py +145 -0
  77. package/md_cg/test_structboost.py +102 -0
  78. package/md_cg/test_timeboost.py +83 -0
  79. package/md_cg/test_timeboost_relativedate.py +71 -0
  80. package/md_cg/test_topicagg.py +129 -0
  81. package/md_cg/timeboost.py +223 -0
  82. package/md_cg/topicagg.py +142 -0
  83. package/md_cg/whitebox_kb/wisdom/code_compose.py +33 -5
  84. package/md_cg/whitebox_kb/wisdom/compose_engine.py +5 -1
  85. package/md_cg/whitebox_kb/wisdom/role_compose.py +271 -268
  86. package/package.json +137 -137
  87. package/src/hooks.ts +43 -1
  88. package/ui/minimal/payload/build-ui.mjs +41 -2
  89. package/ui/minimal/payload/lib/client.js +29 -3
  90. package/ui/minimal/payload/web/app.js +3 -1
  91. package/ui/minimal/payload/web/client.js +30 -3
  92. package/web/app.js +3 -1
  93. package/web/client.js +30 -3
  94. package/zcode/AGENTS.md +1 -1
@@ -0,0 +1,147 @@
1
+ # -*- coding: utf-8 -*-
2
+ """#753 续:否定词 + **非连接字**(使用/采用/用)作用域的守卫。
3
+
4
+ 背景
5
+ ====
6
+ `_negated_at` 原先把「否定词与被否定成分之间」的允许字符限定为连接字
7
+ (`NEG_LINK_CHARS` = 的/这/种/是/…)。于是:
8
+
9
+ 条件写 `PE`,问句 `这里未使用 pe 估值`
10
+ #761 归一前:`pe` 不命中 `PE` → DEFER ✓
11
+ #761 归一后:命中 → ACCEPT ✗(否定句被当成"条件已确认")
12
+
13
+ 缺口在**语义**:`未`否定的是动词`使用`,`pe估值`是`使用`的宾语。
14
+ 宾语被否定动词管辖,故仍是"条件未确认"。只认连接字会把这类全漏掉。
15
+
16
+ 本文件钉三件事
17
+ ==============
18
+ 1. **动词字集生效**:`未使用X` / `不是用X` / `没有使用X` / `未采用X` 等必须判为否定。
19
+ 2. **不误伤肯定句**:`使用X` / `用X`(无否定词)必须判为命中。
20
+ 3. **不碰 #362 既有语义**:`不适合梅雨季节的施工` **照旧命中**(`适合`不是动词字,
21
+ 否定词作用的是`适合`而非`梅雨季节`);A-not-A(`是不是X`)照旧命中。
22
+
23
+ 判别力:删掉 `NEG_VERB_CHARS` 的任一字符 → 第 1 组红。
24
+ 把 `NEG_VERB_CHARS` 塞进 `NEG_LINK_CHARS`(合并两表)→ 结构断言红。
25
+ """
26
+ import sys
27
+
28
+ import md_cg.mdcg as M
29
+ from md_cg.mdcg import MdCG, neg_condition_hits
30
+
31
+ PASS = FAIL = 0
32
+
33
+
34
+ def ok(cond, label):
35
+ global PASS, FAIL
36
+ if cond:
37
+ PASS += 1
38
+ else:
39
+ FAIL += 1
40
+ print(" FAIL %s" % label)
41
+
42
+
43
+ print("=" * 68)
44
+ print("R14. 否定词 + 非连接字(虚化动词)作用域")
45
+ print("=" * 68)
46
+
47
+ # ---------------------------------------------------------------- R14a 结构
48
+ # 两表**分开**存放:连接字(纯语法)与虚化动词(承载语义)词性不同,
49
+ # 合并会掩盖「否定词作用于动作、动作的宾语才是词项」这层结构。
50
+ ok(hasattr(M, "NEG_VERB_CHARS"), "R14a NEG_VERB_CHARS 应存在(独立于 NEG_LINK_CHARS)")
51
+ ok(hasattr(M, "NEG_SCOPE_CHARS"), "R14a NEG_SCOPE_CHARS 应存在(判据准入集的单点)")
52
+ ok(
53
+ set(M.NEG_SCOPE_CHARS) == set(M.NEG_LINK_CHARS) | set(M.NEG_VERB_CHARS),
54
+ "R14a NEG_SCOPE_CHARS 必须恰为 连接字 ∪ 虚化动词字(不多不少)",
55
+ )
56
+ ok(
57
+ not (set(M.NEG_VERB_CHARS) <= set("是这那该此种个些的了在为之其")),
58
+ "R14a 虚化动词字不得全部落在连接字表内(否则分表无意义)",
59
+ )
60
+ # 逐字符断言,防止有人写成多字串(frozenset 里多字串永不匹配单个字符)
61
+ ok(
62
+ all(len(c) == 1 for c in M.NEG_VERB_CHARS),
63
+ "R14a NEG_VERB_CHARS 必须逐**单字**存放(多字串在逐字符判据里永不匹配)",
64
+ )
65
+ for ch in "使用采":
66
+ ok(ch in M.NEG_VERB_CHARS, "R14a NEG_VERB_CHARS 应含 %r" % ch)
67
+
68
+ # ------------------------------------------------------- R14b 否定句须判否定
69
+ NEG_SENTS = [
70
+ ("未使用pe估值", "pe"),
71
+ ("不是用pe估值", "pe"),
72
+ ("没有使用pe估值", "pe"),
73
+ ("未采用pe估值", "pe"),
74
+ ("不是采用pe估值", "pe"),
75
+ ("未使用这种pe估值", "pe"), # 动词 + 连接字混合
76
+ ("没有用这个pe估值", "pe"), # 动词 + 连接字混合
77
+ ]
78
+ for sent, needle in NEG_SENTS:
79
+ i = sent.find(needle)
80
+ ok(i >= 0, "R14b 用例自检:%r 应含 %r" % (sent, needle))
81
+ if i < 0:
82
+ continue
83
+ ok(M._negated_at(sent, i), "R14b %r 中 %r 应判为被否定" % (sent, needle))
84
+
85
+ # ---------------------------------------------- R14c 肯定句不得被误判否定
86
+ POS_SENTS = [
87
+ ("使用pe估值", "pe"),
88
+ ("用pe估值", "pe"),
89
+ ("采用pe估值", "pe"),
90
+ ("pe估值", "pe"),
91
+ ]
92
+ for sent, needle in POS_SENTS:
93
+ i = sent.find(needle)
94
+ ok(i >= 0, "R14c 用例自检:%r 应含 %r" % (sent, needle))
95
+ if i < 0:
96
+ continue
97
+ ok(not M._negated_at(sent, i), "R14c %r 中 %r 不应被判否定" % (sent, needle))
98
+
99
+ # ------------------------------------------------- R14d #362 既有语义不变
100
+ # 这是**回归防线**:本补丁只扩字符集,不改判定结构。
101
+ # `不适合梅雨季节的施工`:否定词是`不`,其后`适合`不在准入集 ⇒ 仍判「未被否定」。
102
+ ok(
103
+ bool(neg_condition_hits(["梅雨季节"], "不适合梅雨季节的施工", "")),
104
+ "R14d `不适合梅雨季节的施工` 必须**照旧命中**(#362 语义,不得因本补丁收紧)",
105
+ )
106
+ ok(
107
+ not bool(neg_condition_hits(["梅雨季节"], "不是梅雨季节", "")),
108
+ "R14d `不是梅雨季节` 仍须判为否定(取消命中)",
109
+ )
110
+ ok(
111
+ not bool(neg_condition_hits(["梅雨季节"], "不是这种梅雨季节", "")),
112
+ "R14d `不是这种梅雨季节` 仍须判为否定",
113
+ )
114
+ ok(
115
+ bool(neg_condition_hits(["梅雨季节"], "是不是梅雨季节", "")),
116
+ "R14d A-not-A(`是不是X`)仍须命中(不得被本补丁误判为否定)",
117
+ )
118
+
119
+ # ------------------------------------- R14e 端到端:条件面在否定句上须 DEFER
120
+ CONTENT = """# 功能名:T 节点
121
+ # 生效条件:PE
122
+ # 子功能:x
123
+ # 执行:y
124
+ # 验证方式:z
125
+ # 不适用条件:无
126
+ """
127
+ NODE = {
128
+ # verification_basis 取闭集内合法值 `test`(`nodefile.VERIFICATION_BASIS`);
129
+ # 不用自由文本:字面量须在 issue #466 登记表覆盖域内,否则 B01 判红。
130
+ "frontmatter": {"non_applicable_conditions": [], "verification_basis": "test"},
131
+ "content": CONTENT,
132
+ }
133
+ for q in ["未使用 pe 估值", "不是用 pe 估值", "没有使用 pe 估值",
134
+ "未采用 pe 估值", "不是采用 pe 估值"]:
135
+ st = MdCG.judge_qualification(NODE, q, None)["state"]
136
+ ok(st == "DEFER", "R14e 否定句 %r 应 DEFER(实得 %s)" % (q, st))
137
+
138
+ # 肯定句必须保持 ACCEPT(否则是矫枉过正)
139
+ for q in ["pe 估值", "使用 pe 估值", "用 pe 估值"]:
140
+ st = MdCG.judge_qualification(NODE, q, None)["state"]
141
+ ok(st == "ACCEPT", "R14e 肯定句 %r 应 ACCEPT(实得 %s)" % (q, st))
142
+
143
+ print()
144
+ print("=" * 68)
145
+ print("cond_term_neg_scope: %d passed, %d failed" % (PASS, FAIL))
146
+ print("=" * 68)
147
+ sys.exit(1 if FAIL else 0)
@@ -0,0 +1,157 @@
1
+ # -*- coding: utf-8 -*-
2
+ """确定性重排(detrerank)守卫。
3
+
4
+ 钉三件事:
5
+ 1. 四信号可数且确定(同输入同输出);
6
+ 2. **无模型分**——模块源码不得出现 embedding/similarity/模型调用;
7
+ 3. **时钟显式**——rerank 必须由调用方传 now,函数内不得取 time.time()。
8
+ """
9
+ from __future__ import annotations
10
+
11
+ import inspect
12
+ import io
13
+ import os
14
+ import sys
15
+
16
+ sys.stdout.reconfigure(encoding="utf-8", errors="replace")
17
+ _REPO = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
18
+ sys.path.insert(0, _REPO)
19
+
20
+ from md_cg import detrerank as D # noqa: E402
21
+
22
+ FAILS = []
23
+ NOW = 1_800_000_000.0
24
+
25
+
26
+ def ok(cond, name, detail=""):
27
+ if cond:
28
+ print(" PASS " + name)
29
+ else:
30
+ print(" FAIL " + name + " " + str(detail)[:200])
31
+ FAILS.append(name)
32
+
33
+
34
+ def _code_only(path):
35
+ """把源码解析成 AST 后**去掉 docstring**再 unparse —— 只留代码。
36
+
37
+ 为什么需要:本模块的说明文字会提到「不得出现 embedding / time.time」这类词,
38
+ 直接 grep 源码会把**守卫自己的说明**判成违规(假红)。守卫要盯实现,不是措辞。
39
+ """
40
+ import ast
41
+ tree = ast.parse(io.open(path, encoding="utf-8").read())
42
+ for nd in ast.walk(tree):
43
+ body = getattr(nd, "body", None)
44
+ if (isinstance(body, list) and body
45
+ and isinstance(body[0], ast.Expr)
46
+ and isinstance(body[0].value, ast.Constant)
47
+ and isinstance(body[0].value.value, str)):
48
+ body[0] = ast.Pass()
49
+ return ast.unparse(tree)
50
+
51
+
52
+ def node(nid, content, tags=None, created_at=None, edges=None, layer="knowledge", session=None):
53
+ fm = {"layer": layer, "tags": tags or []}
54
+ if created_at is not None:
55
+ fm["created_at"] = created_at
56
+ if edges is not None:
57
+ fm["edges"] = edges
58
+ if session:
59
+ fm["session"] = session
60
+ return {"id": nid, "content": content, "frontmatter": fm}
61
+
62
+
63
+ def main():
64
+ print("=" * 66)
65
+ print("确定性重排(detrerank)")
66
+ print("=" * 66)
67
+ q = "灯箱 报价 制作"
68
+ terms = D._terms(q)
69
+ ok(terms == ["灯箱", "报价", "制作"], "D1 分词确定", terms)
70
+
71
+ a = node("a", "灯箱已有报价,尚未安排制作", ["灯箱"], created_at=NOW - 3600)
72
+ b = node("b", "无关内容", [], created_at=NOW - 86400 * 365)
73
+ sa = D.signals(a, terms, NOW)
74
+ sb = D.signals(b, terms, NOW)
75
+ ok(sa["lexical"] == 3 and sb["lexical"] == 0,
76
+ "D2 词法信号 = query 词命中数", (sa["lexical"], sb["lexical"]))
77
+ ok(0.0 < sa["temporal"] <= 1.0, "D3 时间信号落在 (0,1]", sa["temporal"])
78
+ ok(sb["temporal"] < sa["temporal"], "D4 旧条目时间分更低", (sa["temporal"], sb["temporal"]))
79
+
80
+ n_edge = node("e", "灯箱 报价", ["灯箱"], created_at=NOW - 10,
81
+ edges=[{"target": "x", "relation_type": "causal"}])
82
+ se = D.signals(n_edge, terms, NOW)
83
+ ok(se["interaction"] == 1, "D5 交互信号 = 出边数(真的发生过作用)", se["interaction"])
84
+
85
+ n_same = node("s", "灯箱", ["灯箱"], created_at=NOW, session="sess-1")
86
+ ss = D.signals(n_same, terms, NOW, session="sess-1")
87
+ sn = D.signals(n_same, terms, NOW, session="sess-2")
88
+ ok(ss["spatial"] > sn["spatial"], "D6 空间信号含同会话加成", (ss["spatial"], sn["spatial"]))
89
+
90
+ # D7 全序 + 幂等:同输入两次结果逐位相同
91
+ items = [(a, 0.9, {}, []), (b, 0.9, {}, []), (n_edge, 0.9, {}, [])]
92
+ r1 = D.rerank(items, q, NOW)
93
+ r2 = D.rerank(items, q, NOW)
94
+ ok([x[0][0]["id"] for x in r1] == [x[0][0]["id"] for x in r2], "D7 同输入同输出(可回放)")
95
+ ok([x[0][0]["id"] for x in r1][0] != "b", "D8 词法强的排在前面", [x[0][0]["id"] for x in r1])
96
+
97
+ # D9 分数相同按 id 字典序(全序兜底)
98
+ twins = [(node("z", "同样内容", []), 0.5, {}, []),
99
+ (node("a", "同样内容", []), 0.5, {}, [])]
100
+ rt = D.rerank(twins, q, NOW)
101
+ ok([x[0][0]["id"] for x in rt] == ["a", "z"], "D9 同分按 id 字典序",
102
+ [x[0][0]["id"] for x in rt])
103
+
104
+ # D10/D11b:只扫**代码**(去 docstring/注释)——说明文字里出现这些词是允许的,
105
+ # 守卫要盯的是「实现里有没有真的引入模型分/隐式取时钟」。
106
+ code = _code_only(os.path.join(_REPO, "md_cg", "detrerank.py")).lower()
107
+ banned = [w for w in ("embedding", "cosine", "similarity", "openai", "model.")
108
+ if w in code]
109
+ ok(not banned, "D10 代码内无模型分/相似度项", banned)
110
+
111
+ sig = inspect.signature(D.rerank)
112
+ ok("now" in sig.parameters, "D11a rerank 形参含 now(时钟显式)", list(sig.parameters))
113
+ ok("time.time" not in code and "datetime.now" not in code,
114
+ "D11b 代码内不取系统时钟")
115
+
116
+ # D12 权重是固定常量(改值必须走评审)
117
+ ok(isinstance(D.WEIGHTS, dict) and set(D.WEIGHTS) == {"lexical", "interaction", "spatial", "temporal"},
118
+ "D12 四信号权重常量齐全", D.WEIGHTS)
119
+
120
+ # ---------- 端到端:接线层(recall 的 MDCG_DET_RERANK 开关) ----------
121
+ import tempfile
122
+ from md_cg.mdcos import MdCGOS
123
+ root = tempfile.mkdtemp(prefix="det_e2e_")
124
+ g2 = MdCGOS(root, actor="test")
125
+ ccg = ("# 功能名:{n}\n# 生效条件:{c}\n# 子功能:s\n# 执行:e\n"
126
+ "# 验证方式:test\n# 不适用条件:无\n\n{b}\n")
127
+ g2.add("a", ccg.format(n="灯箱报价记录", c="灯箱 报价",
128
+ b="灯箱 已有报价 制作 待安排"), tags=["灯箱"])
129
+ g2.add("b", ccg.format(n="灯箱制作安排", c="灯箱 制作",
130
+ b="制作 安排 灯箱"), tags=["灯箱"])
131
+ g2.add("c", ccg.format(n="其它项目", c="别的",
132
+ b="别的内容 灯箱 提了一句"), tags=["其它"])
133
+ os.environ["MDCG_DET_RERANK"] = "1"
134
+ try:
135
+ p1 = g2.recall("灯箱 报价 制作", budget_tokens=3000, k=10)
136
+ ids1 = [e["id"] for e in p1["pack"]]
137
+ ok(p1["meta"].get("det_reranked") is True, "E1 开启后 meta.det_reranked=True")
138
+ ok(all("det" in e for e in p1["pack"]), "E2 每条带四信号读数(可审计)")
139
+ ok(ids1[0] == "a", "E3 词法最强的排第一", ids1)
140
+ sig0 = p1["pack"][0]["det"]["signals"]
141
+ ok(set(sig0) == {"lexical", "interaction", "spatial", "temporal"},
142
+ "E4 信号四键齐全", sig0)
143
+ p2 = g2.recall("灯箱 报价 制作", budget_tokens=3000, k=10)
144
+ ok([e["id"] for e in p2["pack"]] == ids1, "E5 同输入同输出(可回放)")
145
+ finally:
146
+ os.environ.pop("MDCG_DET_RERANK", None)
147
+ p3 = g2.recall("灯箱 报价 制作", budget_tokens=3000, k=10)
148
+ ok(p3["meta"].get("det_reranked") in (False, None), "E6 缺省关闭")
149
+ ok(not any("det" in e for e in p3["pack"]), "E7 关闭时不注入 det 读数")
150
+
151
+ print()
152
+ print("detrerank:%d 项失败" % len(FAILS) if FAILS else "detrerank:全部通过")
153
+ return 1 if FAILS else 0
154
+
155
+
156
+ if __name__ == "__main__":
157
+ sys.exit(main())
@@ -0,0 +1,136 @@
1
+ # -*- coding: utf-8 -*-
2
+ """W5 四 KPI 门禁的守卫:门禁自己必须可红可绿。
3
+
4
+ 真源:docs/architecture/灵枢1.0_建设路线_v0.1.md §9 W5「每 KPI 一条可红可绿的检查」。
5
+
6
+ # 功能名:W5 门禁自证守卫
7
+ # 生效条件:改动 scripts/gate_four_kpi.py、或有人怀疑「门禁恒绿」时
8
+ # 子功能:①四条 KPI 齐全且名字与路线稿一字不差
9
+ # ②每条都绑定真实存在的证据模块(防证据路径写错导致永远 INVALID 却不自知)
10
+ # ③变异体必须变红(模块不存在 / 恒真不打印 / 打印 0 条 / 自报失败)
11
+ # ④真证据必须全绿(防门禁恒红,恒红与恒绿同属无判别力)
12
+ # ⑤判据锚点在场(min_pass 下限、正则、fail-closed 的 INVALID 分支)
13
+ # 执行:python -X utf8 -m md_cg.test_gate_four_kpi
14
+ # 验证方式:本文件全绿即通过;任一条红退出码 1
15
+ # 不适用条件:不判 KPI 语义是否"足够好";不替代全量套件
16
+
17
+ 为什么要单独守这个门禁
18
+ 本会话已实证:断言写错对象后,删掉被测判据用例仍 PASS(恒绿)。
19
+ 门禁是「判别的判别」,它一旦恒绿,下游全部结论失效且无人察觉。
20
+ 故本守卫不只跑一遍门禁,而是**要求门禁对变异体必须变红**。
21
+ """
22
+ from __future__ import annotations
23
+
24
+ import io
25
+ import os
26
+ import re
27
+ import sys
28
+ import tempfile
29
+
30
+ HERE = os.path.dirname(os.path.abspath(__file__))
31
+ REPO = os.path.dirname(HERE)
32
+ sys.path.insert(0, REPO)
33
+ sys.path.insert(0, os.path.join(REPO, "scripts"))
34
+
35
+ PASS = 0
36
+ FAIL = 0
37
+ FAILS = []
38
+
39
+
40
+ def check(name, ok, detail=""):
41
+ global PASS, FAIL
42
+ if ok:
43
+ PASS += 1
44
+ print(" PASS %s" % name)
45
+ else:
46
+ FAIL += 1
47
+ FAILS.append(name)
48
+ print(" FAIL %s %s" % (name, detail))
49
+
50
+
51
+ def main():
52
+ import gate_four_kpi as G
53
+
54
+ print("=" * 78)
55
+ print("W5 四 KPI 门禁 · 守卫")
56
+ print("=" * 78)
57
+
58
+ # ---- A 四个 KPI 齐全,名字与路线稿一致 ----
59
+ names = [s["kpi"] for s in G.KPIS]
60
+ want = ["记得住", "查得到", "改得回", "账不丢"]
61
+ check("A1 恰好四个 KPI 且顺序与 §0 原文一致", names == want, "实得 %r" % (names,))
62
+
63
+ src = io.open(os.path.join(REPO, "docs", "architecture",
64
+ "灵枢1.0_建设路线_v0.1.md"), encoding="utf-8").read()
65
+ missing = [n for n in want if n not in src]
66
+ check("A2 四个名字都能在路线稿原文中找到", not missing, "缺:%r" % (missing,))
67
+
68
+ # ---- B 证据模块真实存在 ----
69
+ for spec in G.KPIS:
70
+ mod = spec["evidence"]
71
+ rel = os.path.join(REPO, "md_cg", mod.split(".")[-1] + ".py")
72
+ check("B 证据存在:%s → %s" % (spec["kpi"], mod), os.path.isfile(rel), "缺文件 %s" % rel)
73
+
74
+ # ---- C 每条都有可判参数(防漏填导致空转)----
75
+ for spec in G.KPIS:
76
+ ok = (spec["pattern"] and spec["min_pass"] > 0 and spec["why"].strip())
77
+ check("C 判据参数齐备:%s(min_pass=%s)" % (spec["kpi"], spec["min_pass"]), ok)
78
+
79
+ # ---- D 变异体必须变红(核心)----
80
+ tmp = tempfile.mkdtemp(prefix="kpi_gate_guard_")
81
+ os.environ["PYTHONPATH"] = tmp + os.pathsep + os.environ.get("PYTHONPATH", "")
82
+
83
+ def write_mod(name, body):
84
+ with open(os.path.join(tmp, name + ".py"), "w", encoding="utf-8") as f:
85
+ f.write(body)
86
+ return name
87
+
88
+ m_missing = "zzz_no_such_module_xyz"
89
+ m_pass = write_mod("mut_always_pass", "import sys\nsys.exit(0)\n")
90
+ m_zero = write_mod("mut_zero_assert", "print('0 通过,0 失败')\n")
91
+ m_bad = write_mod("mut_failed", "print('3 通过,2 失败')\nimport sys\nsys.exit(1)\n")
92
+
93
+ cases = [
94
+ ("D1 模块不存在 → INVALID", m_missing, "INVALID"),
95
+ ("D2 恒真但不打印条数 → INVALID", m_pass, "INVALID"),
96
+ ("D3 打印 0 条断言 → INVALID", m_zero, "INVALID"),
97
+ ("D4 自报失败 → FAIL", m_bad, "FAIL"),
98
+ ]
99
+ for label, module, expect in cases:
100
+ spec = dict(G.KPIS[0])
101
+ spec["evidence"] = module
102
+ spec["pattern"] = r"(\d+)\s*通过,\s*(\d+)\s*失败"
103
+ spec["min_pass"] = 4
104
+ r = G.evaluate(spec, timeout=120)
105
+ check("%s(实得 %s)" % (label, r["state"]), r["state"] == expect, r["reason"])
106
+
107
+ # ---- E 反向:真证据必须 PASS(防恒红)----
108
+ ev = G.evaluate(G.KPIS[2], timeout=900) # 改得回:单条最快
109
+ check("E1 真证据应 PASS(防门禁恒红,实得 %s)" % ev["state"], ev["state"] == "PASS", ev["reason"])
110
+
111
+ # ---- F 结构锚点:fail-closed 分支必须在场 ----
112
+ gsrc = io.open(os.path.join(REPO, "scripts", "gate_four_kpi.py"), encoding="utf-8").read()
113
+ for anchor, why in (
114
+ ("INVALID", "解析不出条数/fail-closed 必须用 INVALID 而非 PASS"),
115
+ ("min_pass", "低于断言下限必须判 INVALID(防恒绿)"),
116
+ ("TimeoutExpired", "超时必须不判 PASS"),
117
+ ("len(groups)", "必须真的解析出条数,不能只看退出码"),
118
+ ):
119
+ check("F 结构锚点在场:%s(%s)" % (anchor, why), anchor in gsrc)
120
+
121
+ # ---- G verdict 聚合规则 ----
122
+ rows = [{"state": "PASS"}, {"state": "INVALID"}]
123
+ check("G1 有 INVALID 时整体 INVALID", G.verdict_of(rows) == "INVALID", G.verdict_of(rows))
124
+ rows = [{"state": "INVALID"}, {"state": "FAIL"}]
125
+ check("G2 FAIL 优先于 INVALID", G.verdict_of(rows) == "FAIL", G.verdict_of(rows))
126
+ check("G3 空集判 INVALID(不判 PASS)", G.verdict_of([]) == "INVALID", G.verdict_of([]))
127
+
128
+ print()
129
+ print("===== W5 门禁守卫:%d 通过,%d 失败 =====" % (PASS, FAIL))
130
+ if FAILS:
131
+ print(" 红项:" + ";".join(FAILS))
132
+ return 0 if FAIL == 0 else 1
133
+
134
+
135
+ if __name__ == "__main__":
136
+ sys.exit(main())
@@ -55,7 +55,11 @@ PINNED_INFO_KEYS = frozenset({
55
55
 
56
56
  NEW_INFO_KEYS = ("code_generation", "disk_generation", "stale_on_disk",
57
57
  "restart_required", "hint",
58
- "write_gate") # issue #708:分层写入闸的进程内计数(warn 观察期)
58
+ "write_gate", # issue #708:分层写入闸的进程内计数(warn 观察期)
59
+ # issue #812(2026-11-07):内存索引 vs 落盘索引的偏差面——纯增量扩键。
60
+ # 因由:批量软删 1987 条后 node_count 不降反升,真因是本进程内存索引陈旧;
61
+ # 没有这组键,调用方无法判定「数字到底信哪个」。
62
+ "index_mtime", "node_count_scope")
59
63
 
60
64
  _FAILED = []
61
65
 
@@ -286,6 +290,13 @@ FAST_INFO_KEYS = set(NEW_INFO_KEYS) | {
286
290
  #: 快照模式**新增**的键(除 NEW_INFO_KEYS 外)——只允许这三个出现在 full 的「多出」里。
287
291
  FAST_ONLY_KEYS = {"flush_state", "full", "node_count"}
288
292
 
293
+ #: issue #812:**只在 full 面**出现的键(快照面不解析 58MB 落盘索引,故不能要求快照齐备)。
294
+ #: 它们仍需登记,否则「full 多出未登记键」会红——这正是扩键流程要的。
295
+ #: 只在 full=true 时出现的 info 键(新增键必须登记,否则「快照缺键/多键」守卫会红)。
296
+ #: `bucket_auto`:桶索引每日自动重建的读数(2026-11-09 引入,info 触发点)。
297
+ FULL_ONLY_INFO_KEYS = {"index_node_count", "index_stale", "index_stale_error",
298
+ "bucket_auto"}
299
+
289
300
 
290
301
  def _parse_info(stdout, rid):
291
302
  for line in stdout.splitlines():
@@ -346,7 +357,8 @@ def check_info_keys(tmp):
346
357
  if lost:
347
358
  _fail("五·键集", "full 丢失既有 20 键:%s" % lost)
348
359
  return
349
- extra = sorted(set(full) - PINNED_INFO_KEYS - set(NEW_INFO_KEYS) - FAST_ONLY_KEYS)
360
+ extra = sorted(set(full) - PINNED_INFO_KEYS - set(NEW_INFO_KEYS) - FAST_ONLY_KEYS
361
+ - FULL_ONLY_INFO_KEYS)
350
362
  if extra:
351
363
  _fail("五·键集", "full 多出未登记键:%s" % extra)
352
364
  return
@@ -256,3 +256,51 @@ def test_g3_existing_move_behavior_zero_regression(tmp_path):
256
256
  == buckets_before.get("orphan", 0) - 1, "桶归属迁移票数不符"
257
257
  assert lifecycle.state_of(after_entry) == lifecycle.state_of(before_entry)
258
258
  cg.close()
259
+
260
+
261
+ # ---------------------------------------------------------------- 直跑入口
262
+ # 为什么需要(issue #807 同族,2026-11-07 实测):本件的 6 个用例都是
263
+ # `def test_*(tmp_path)`,**没有模块级断言、没有 __main__**。CI 的门禁
264
+ # (scripts/run_tests.py)对 md_cg 组只发 `python -X utf8 -m <module>` ——
265
+ # `-m` 对纯 pytest-style 模块只做 import+define,跑**零个**用例、RC 恒 0。
266
+ # 实测:把 test_g1 的首条断言改成 `assert False` 后,直跑仍 RC=0,而
267
+ # `-m pytest` 下是 1 failed 5 passed。⇒ 本件在门禁里**从未真正执行过**,
268
+ # 是一条假绿守卫。补此入口使其成为 script-style,让直跑真的跑断言。
269
+ #
270
+ # tmp_path 是 pytest 的 fixture,直跑路径下用 tempfile.mkdtemp 等价替代
271
+ # (用例只把它当目录路径用:str(tmp_path) / os.path.join(root, ...))。
272
+ _ENTRY_CASES = (
273
+ ("G1 搬迁携带全部 7 键", test_g1_move_carries_all_seven_keys),
274
+ ("G1b archived lifecycle 键存活", test_g1b_archived_lifecycle_key_survives),
275
+ ("G1c 搬迁条目形态等于重扫", test_g1c_moved_entry_form_equals_rescan),
276
+ ("G1d 上游 verify 面", test_g1d_upstream_verify_face),
277
+ ("G2 旧层残留清零", test_g2_no_old_layer_residue),
278
+ ("G3 既有搬迁行为零回归", test_g3_existing_move_behavior_zero_regression),
279
+ )
280
+
281
+
282
+ def main():
283
+ import tempfile
284
+ import traceback
285
+
286
+ failed = []
287
+ for name, fn in _ENTRY_CASES:
288
+ d = tempfile.mkdtemp(prefix="mdcg_i192_")
289
+ try:
290
+ fn(d)
291
+ print(" PASS %s" % name)
292
+ except Exception as exc: # noqa: BLE001 —— 直跑要报全,不吞
293
+ failed.append(name)
294
+ print(" FAIL %s: %s: %s" % (name, type(exc).__name__, exc))
295
+ traceback.print_exc()
296
+ finally:
297
+ shutil.rmtree(d, ignore_errors=True)
298
+ total = len(_ENTRY_CASES)
299
+ print("")
300
+ print("===== SUMMARY %d/%d 通过 =====" % (total - len(failed), total))
301
+ print("VERDICT=%s" % ("PASS" if not failed else "FAIL"))
302
+ return 0 if not failed else 1
303
+
304
+
305
+ if __name__ == "__main__":
306
+ raise SystemExit(main())