@furongjun1999/dsh-memory 0.9.6 → 0.9.8

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 (233) hide show
  1. package/README.md +150 -662
  2. package/codebuddy/CODEBUDDY.md +333 -204
  3. package/data/policy.json +38 -38
  4. package/docs/README.md +18 -0
  5. package/docs/discipline/harnesses.yaml +11 -3
  6. package/docs/discipline/templates/compact.txt.tmpl +20 -18
  7. package/docs/discipline/templates/full.md.tmpl +63 -61
  8. package/docs/discipline/templates/rules.mdc.tmpl +5 -3
  9. package/docs/discipline/templates/skill.md.tmpl +25 -23
  10. package/docs/discipline/templates/spec.txt.tmpl +2 -2
  11. package/docs/discipline/templates/zcode-user.md.tmpl +1 -1
  12. package/dsh/cordis.yml.example +139 -139
  13. package/lib/hooks.d.ts +19 -0
  14. package/lib/hooks.js +30 -7
  15. package/lib/index.js +18 -3
  16. package/lib/init.js +19 -4
  17. package/lib/lib/hook_audit.js +64 -2
  18. package/lib/lib/prompt_safety.js +14 -3
  19. package/lib/tools.d.ts +7 -0
  20. package/lib/tools.js +22 -1
  21. package/md_cg/_guard_kit.py +121 -121
  22. package/md_cg/anchor.py +114 -114
  23. package/md_cg/audit.py +15 -2
  24. package/md_cg/chain.py +3 -2
  25. package/md_cg/config_registry_guard.py +487 -487
  26. package/md_cg/conformance.py +747 -747
  27. package/md_cg/consolidate.py +1793 -1793
  28. package/md_cg/crypto.py +794 -794
  29. package/md_cg/hyperedge.py +262 -262
  30. package/md_cg/interface_schema.py +146 -146
  31. package/md_cg/lifecycle.py +426 -426
  32. package/md_cg/mdcos.py +12 -8
  33. package/md_cg/predict.py +1103 -1103
  34. package/md_cg/protect.py +654 -654
  35. package/md_cg/provenance.py +606 -606
  36. package/md_cg/queue_health.py +130 -130
  37. package/md_cg/refine.py +613 -613
  38. package/md_cg/retrieve.py +128 -128
  39. package/md_cg/revise.py +140 -140
  40. package/md_cg/rollback.py +657 -657
  41. package/md_cg/rollback_cli.py +124 -124
  42. package/md_cg/rotate.py +346 -346
  43. package/md_cg/routing.py +490 -490
  44. package/md_cg/self_surface.py +50 -50
  45. package/md_cg/semantic/canonical.py +144 -144
  46. package/md_cg/semantic/zh_en_atoms.py +141 -141
  47. package/md_cg/sources.py +1091 -1091
  48. package/md_cg/stg.py +3 -2
  49. package/md_cg/sustain.py +10 -1
  50. package/md_cg/test_audit_rotate_parity.py +279 -279
  51. package/md_cg/test_autonomy_modes.py +2013 -2013
  52. package/md_cg/test_cn_recall_grams.py +206 -206
  53. package/md_cg/test_code_kind_forbidden.py +101 -0
  54. package/md_cg/test_condaccept.py +156 -156
  55. package/md_cg/test_condgate.py +141 -141
  56. package/md_cg/test_heartbeat_ledger.py +9 -3
  57. package/md_cg/test_i50e_readside_protection.py +658 -658
  58. package/md_cg/test_issue111_gate_order.py +76 -76
  59. package/md_cg/test_issue111_scoring_contract.py +86 -86
  60. package/md_cg/test_issue121_calibration_bounds.py +293 -293
  61. package/md_cg/test_issue145_sleep_no_silent_rollback.py +167 -167
  62. package/md_cg/test_issue147_pack_id_escape.py +378 -378
  63. package/md_cg/test_issue148_merge_downgrade_leak.py +192 -192
  64. package/md_cg/test_issue149_accel_default_face.py +248 -248
  65. package/md_cg/test_issue150_empty_query_terms.py +328 -328
  66. package/md_cg/test_issue151_153_157_write_cache_index.py +151 -151
  67. package/md_cg/test_issue152_sealed_postings.py +309 -309
  68. package/md_cg/test_issue153b_index_generation_invalidation.py +142 -142
  69. package/md_cg/test_issue157b_index_remainder.py +344 -344
  70. package/md_cg/test_issue189_conclusion_bidirectional.py +223 -223
  71. package/md_cg/test_issue191_import_trust.py +382 -382
  72. package/md_cg/test_issue193_write_scope.py +378 -378
  73. package/md_cg/test_issue212_writeface_gaps.py +349 -349
  74. package/md_cg/test_issue214_reject_sensitivity.py +266 -266
  75. package/md_cg/test_issue215_condition_source_gap.py +272 -272
  76. package/md_cg/test_issue216_clash_routes.py +214 -214
  77. package/md_cg/test_issue217_slot_routes.py +248 -248
  78. package/md_cg/test_issue218_postcondition_value_blind.py +222 -222
  79. package/md_cg/test_issue258_pack_sensitivity.py +468 -468
  80. package/md_cg/test_issue265_coldverify_multiproc.py +438 -438
  81. package/md_cg/test_issue274_read_node_id_none.py +341 -341
  82. package/md_cg/test_issue275_layer_param.py +250 -250
  83. package/md_cg/test_issue277_node_id_traversal.py +384 -384
  84. package/md_cg/test_issue278_neg_separated_tension.py +271 -271
  85. package/md_cg/test_issue278_neg_window.py +294 -294
  86. package/md_cg/test_issue278_negative_layering.py +83 -83
  87. package/md_cg/test_issue284_skip_not_pass.py +291 -291
  88. package/md_cg/test_issue294_flywheel_sensitivity.py +573 -573
  89. package/md_cg/test_issue301_baseline_corpus_gate.py +245 -245
  90. package/md_cg/test_issue303_design_boundaries.py +262 -262
  91. package/md_cg/test_issue315_merge_branch_sensitivity.py +383 -383
  92. package/md_cg/test_issue319_tenant_unregistered_bypass.py +256 -256
  93. package/md_cg/test_issue321_mcp_deep_json.py +304 -304
  94. package/md_cg/test_issue322_node_write_lock.py +727 -727
  95. package/md_cg/test_issue324_write_size_fields.py +214 -214
  96. package/md_cg/test_issue325_audit_poison.py +229 -229
  97. package/md_cg/test_issue326_ingest_auto_whitelist.py +331 -331
  98. package/md_cg/test_issue326_key_perms.py +355 -355
  99. package/md_cg/test_issue326_twin_parser_parity.py +323 -323
  100. package/md_cg/test_issue326_vm_call_frame.py +569 -569
  101. package/md_cg/test_issue345_writelimit_effective_sens.py +233 -233
  102. package/md_cg/test_issue346_overwrite_sensitivity_gate.py +424 -424
  103. package/md_cg/test_issue348_insight_verify_op.py +199 -199
  104. package/md_cg/test_issue348_merge_rmw_lock.py +345 -345
  105. package/md_cg/test_issue350_casefold_collision.py +235 -235
  106. package/md_cg/test_issue351_ingest_root_fallback.py +201 -201
  107. package/md_cg/test_issue352_imgaudit.py +418 -418
  108. package/md_cg/test_issue352_pathsjson.py +537 -537
  109. package/md_cg/test_issue352_refindex.py +412 -412
  110. package/md_cg/test_issue352_rollback.py +604 -604
  111. package/md_cg/test_issue354_p2_writelimit_toctou.py +327 -327
  112. package/md_cg/test_issue354_rust_index_containment.py +571 -571
  113. package/md_cg/test_issue354_rust_wal_containment.py +606 -606
  114. package/md_cg/test_issue358_truncation_observable.py +247 -247
  115. package/md_cg/test_issue360_tidy_private_excerpt.py +195 -195
  116. package/md_cg/test_issue361_blindspot_not_silent_drop.py +494 -494
  117. package/md_cg/test_issue362_neg_condition.py +468 -468
  118. package/md_cg/test_issue363_lexical_idf.py +275 -275
  119. package/md_cg/test_issue370_readcache_assembly.py +488 -488
  120. package/md_cg/test_issue376_heartbeat_failed.py +293 -293
  121. package/md_cg/test_issue377_layer_protect_stamp.py +324 -324
  122. package/md_cg/test_issue378_protected_floor.py +310 -310
  123. package/md_cg/test_issue383_state_event_write_gate.py +462 -462
  124. package/md_cg/test_issue397_sens_eff_wiring.py +435 -435
  125. package/md_cg/test_issue398_toolface_op_parity.py +511 -511
  126. package/md_cg/test_issue399_hive_json_shape.py +317 -317
  127. package/md_cg/test_issue401_wisdom_codegen.py +295 -295
  128. package/md_cg/test_issue404_heal_receipt.py +389 -389
  129. package/md_cg/test_issue409_compact_index_completeness.py +509 -509
  130. package/md_cg/test_issue410_verify_demote_write.py +366 -366
  131. package/md_cg/test_issue411_scrub_idempotence.py +575 -575
  132. package/md_cg/test_issue412_discipline_producer_parity.py +311 -311
  133. package/md_cg/test_issue413_preimage_failclosed.py +239 -239
  134. package/md_cg/test_issue413_preimage_twin.py +339 -339
  135. package/md_cg/test_issue415_deps_gate_probe.py +357 -357
  136. package/md_cg/test_issue416_entry_key_subset.py +525 -525
  137. package/md_cg/test_issue417_twophase_merge_landed.py +423 -423
  138. package/md_cg/test_issue425_ingest_library_path.py +331 -331
  139. package/md_cg/test_issue426c_enqueue_not_silent.py +346 -346
  140. package/md_cg/test_issue426e_restore_overwrite_live.py +310 -310
  141. package/md_cg/test_issue426f_paths_gbk_failclosed.py +301 -301
  142. package/md_cg/test_issue427_serve_status_shape.py +191 -191
  143. package/md_cg/test_issue431_432_468_protect_single_point.py +322 -322
  144. package/md_cg/test_issue433_outlook_threshold_single_point.py +278 -278
  145. package/md_cg/test_issue434_435_float_gate.py +90 -90
  146. package/md_cg/test_issue439_sibling_sensitivity_inherit.py +448 -448
  147. package/md_cg/test_issue43_default_policy.py +874 -874
  148. package/md_cg/test_issue444_scrub_evidence_plaintext.py +281 -281
  149. package/md_cg/test_issue445_metacognition_ledger_gate.py +454 -454
  150. package/md_cg/test_issue446_453_403_verified_evidence.py +382 -382
  151. package/md_cg/test_issue450_sensitivity_fallback_reason.py +51 -51
  152. package/md_cg/test_issue450_writegate_unknown_sens.py +203 -203
  153. package/md_cg/test_issue451_audit_actor.py +313 -313
  154. package/md_cg/test_issue452_ingest_root_sensitivity.py +299 -299
  155. package/md_cg/test_issue463_ghost_prune_guard.py +259 -259
  156. package/md_cg/test_issue466_basis_registry.py +469 -469
  157. package/md_cg/test_issue467_identity_sensitivity_parity.py +360 -360
  158. package/md_cg/test_issue469_scrub_override_passthrough.py +126 -126
  159. package/md_cg/test_issue470_wal_event_durable.py +427 -427
  160. package/md_cg/test_issue471_identity_readable.py +136 -136
  161. package/md_cg/test_issue472_audit_chain.py +90 -90
  162. package/md_cg/test_issue472_prune_ledger.py +105 -105
  163. package/md_cg/test_issue472_rollback_degraded.py +81 -81
  164. package/md_cg/test_issue491_compose_sandbox.py +460 -460
  165. package/md_cg/test_issue498_merge_trace_stale_index.py +280 -280
  166. package/md_cg/test_issue511_scoreboard_bad_files.py +256 -256
  167. package/md_cg/test_issue521_rotate_lock_suffix.py +110 -110
  168. package/md_cg/test_issue524_orphan_noise_no_demote.py +88 -88
  169. package/md_cg/test_issue525_read_null_semantics.py +52 -52
  170. package/md_cg/test_issue547_append_jsonl_whitelist.py +114 -114
  171. package/md_cg/test_issue547_audit_durable.py +96 -96
  172. package/md_cg/test_issue557_r2_write_entries_size_gate.py +83 -83
  173. package/md_cg/test_issue559_refine_prefix.py +69 -69
  174. package/md_cg/test_issue562_retired_faces.py +150 -0
  175. package/md_cg/test_issue71_queue_backlog.py +376 -376
  176. package/md_cg/test_issue72_coldverify_controlled_drain.py +420 -420
  177. package/md_cg/test_issue82_credential_rules.py +271 -271
  178. package/md_cg/test_issue91_verifier_sandbox.py +314 -314
  179. package/md_cg/test_issue93_remember_redaction.py +282 -282
  180. package/md_cg/test_logref.py +1141 -1141
  181. package/md_cg/test_md_access_parity.py +330 -330
  182. package/md_cg/test_md_cg_no_hive_dep.py +139 -139
  183. package/md_cg/test_mutation_residue.py +119 -119
  184. package/md_cg/test_mutation_rollback.py +835 -835
  185. package/md_cg/test_n123_derive_expiry_chain.py +210 -210
  186. package/md_cg/test_n184_keys_concurrent_provision.py +1 -1
  187. package/md_cg/test_n209_verify_write_face_gates.py +466 -466
  188. package/md_cg/test_n62_tenant_bind_failclosed.py +200 -200
  189. package/md_cg/test_neg_tail_honesty.py +733 -733
  190. package/md_cg/test_p21_tokens.py +294 -294
  191. package/md_cg/test_p27_docindex.py +792 -792
  192. package/md_cg/test_p28_refcheck.py +1053 -1053
  193. package/md_cg/test_p2_six_elements.py +476 -476
  194. package/md_cg/test_p33_ccg_wiring.py +300 -300
  195. package/md_cg/test_p3_legacy_closure.py +438 -438
  196. package/md_cg/test_p40_refine_worklist.py +244 -244
  197. package/md_cg/test_p4_freshness.py +781 -781
  198. package/md_cg/test_p9c_dedup_hints.py +288 -288
  199. package/md_cg/test_policy_required_ccg.py +444 -444
  200. package/md_cg/test_read_scope_b27.py +2 -2
  201. package/md_cg/test_retr_s5.py +312 -312
  202. package/md_cg/test_s1_ast_judge.py +109 -109
  203. package/md_cg/test_security_audit_v21.py +249 -249
  204. package/md_cg/test_state_event_op.py +521 -521
  205. package/md_cg/test_stgidx_index_parity.py +1024 -1024
  206. package/md_cg/test_sustain_bad_stamp.py +178 -0
  207. package/md_cg/test_sustain_bounded.py +652 -652
  208. package/md_cg/test_trust_edge_reachability.py +405 -405
  209. package/md_cg/test_w7_redlines.py +659 -659
  210. package/md_cg/test_write_return_shape.py +919 -919
  211. package/md_cg/test_writepipe.py +228 -228
  212. package/md_cg/trust.py +1114 -1114
  213. package/md_cg/twophase.py +303 -303
  214. package/md_cg/whitebox.py +536 -536
  215. package/md_cg/whitebox_kb/wisdom/compose_engine.py +1634 -1634
  216. package/md_cg/whitebox_kb/wisdom/core/condition_vm.py +318 -318
  217. package/md_cg/whitebox_kb/wisdom/core/parser.py +1082 -1082
  218. package/md_cg/whitebox_kb/wisdom/md_access.py +490 -490
  219. package/md_cg/whitebox_kb/wisdom/migrate_wisdom.py +497 -497
  220. package/md_cg/whitebox_kb/wisdom/mini_python.py +995 -995
  221. package/md_cg/whitebox_kb/wisdom/net_units.py +2900 -2900
  222. package/md_cg/whitebox_kb/wisdom/semantic_translate.py +4060 -4060
  223. package/md_cg/writepipe.py +1337 -1337
  224. package/package.json +3 -2
  225. package/scripts/workspace_index.py +328 -0
  226. package/src/hooks.ts +932 -908
  227. package/src/index.ts +21 -6
  228. package/src/init.ts +16 -4
  229. package/src/lib/hook_audit.ts +48 -2
  230. package/src/lib/mdcg_client.ts +469 -469
  231. package/src/lib/prompt_safety.ts +17 -3
  232. package/src/tools.ts +23 -1
  233. package/zcode/AGENTS.md +333 -204
@@ -1,1053 +1,1053 @@
1
- # -*- coding: utf-8 -*-
2
- """ref 增量索引 + 漂移/悬空巡检(P28 · R3 改造验收)。
3
-
4
- 对照 `docs/mdcg/认知图_索引与工程规范化_计划_v0.1.md` 的 R3:
5
-
6
- ① 统一 ref 协议 + 提取器注册表:调用方只说 `kind`,`refindex` 决定用哪个
7
- 提取器;无提取器后缀显式报错。
8
- ② `_refindex.json` 增量(修 F):`incremental=True` 时未变文件**不再读盘重切**
9
- (`skipped_unchanged`),节点仍在(幂等);默认全量,保证「不漏召回」。
10
- ③ `op=ref action=check` 漂移/悬空巡检(修 D):改源一行 → 报 `stale`;删除源
11
- 文件 → 报 `dangling`;**只读、不抛、不改源文件**。
12
- ④ `op=ref action=stat`:看水位,最近一次索引是否被截断(`truncated` 不再静默)。
13
- ⑤ 接入 `sustain.diagnose`:`ref_stale` / `ref_dangling` 进入体检;`heal` 的
14
- `rebuild_refs` 重建派生物后漂移消除(源文件不动)。
15
- ⑥ 检索结果带 `ref` / `ref_kind`(读侧只加字段,不改召回逻辑)。
16
- ⑦ 派生物不膨胀:`_refindex.json` 体积有界、重跑不涨。
17
- ⑧ **判据分工**(#4):漂移 / 悬空只由**节点自身 frontmatter ref** 判,水位只判
18
- 「这个节点能不能跳过、不读源」(五项全等 + 源 size/mtime 未变 + 同 id 只判一次);
19
- 水位与节点不一致记 `ledger_mismatch` 归因行(带两侧 root),**不进 ok、不重复报红**。
20
- ⑨ **撞 id 现场**(D1 现状钉子):两域同 rel path + 同 name ⇒ 同 id、节点被后写域覆盖,
21
- 但归因行可见。本组断言**现状事实**(+ 去重 + 归因字段),不作期望语义断言——
22
- `node_id` 一旦改为含 root,① 会转红以提示同步迁移。
23
- ⑩ **口径钉子**(#5):id 是**位置化**寻址(`sha1(路径::符号名)` / `sha1(路径#标题路径)`),
24
- 幂等判据是**区间哈希 + 行位全等**(不是「id 存在即已索引」)。两侧都钉——实现侧
25
- 手算 sha1 独立复算 + 对 root / 正文 / 行位不敏感 + 五项逐项不跳过;文档侧
26
- `docs/plans/真源索引_通用机制_v0.3.md` §3.1 的口径字面、整块指纹(**陈化硬失败**)
27
- 与行级引用必须仍与实现同源(引用漂移即红,失败信息附现址)。
28
- ⑪ **常驻循环止血**:`SustainLoop._tick_heal` 按 `repeat_guard=True` 调 heal——同因重复
29
- 失败被拦下(`repeat_failure` + 退避)而**诊断面照旧报病灶**(判据不被削弱)、
30
- `heals[*].failed` 单列(没修好不静默、不假装成功)。
31
- 红基线(判别力自证,见 `_MUTATIONS` / `EXPECTED_RED`):`--head-baseline` 对
32
- **工作区 `md_cg/refindex.py`** 做定点变异(快路径探测数据改回水位条目=v1 旧行为),
33
- 断言红项集合**恰好**等于预期三项(⑩ / ⑪ 两段钉的是别的语义,变异态下**仍绿**,
34
- 故不入红项表——「恰好」二字正是这么被守住的)。
35
-
36
- 运行:python -m md_cg.test_p28_refcheck [--head-baseline]
37
- """
38
- from __future__ import annotations
39
-
40
- import argparse
41
- import hashlib
42
- import os
43
- import re
44
- import shutil
45
- import sys
46
- import tempfile
47
- import time
48
-
49
- from . import corpus, refindex, sources, sustain, tokens
50
- from .mdcg import MdCG
51
- from .mcp_server import call_tool
52
-
53
- PASS = FAIL = 0
54
- FAILS = []
55
- RED = [] # 转红项 cid(红基线模式比对用;绿态为空)
56
-
57
- _BASE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
58
- ROOT = os.path.join(_BASE, "_md_cg_p28")
59
-
60
-
61
- def check(name, cond, detail="", cid=None):
62
- global PASS, FAIL
63
- if cond:
64
- PASS += 1
65
- print(f" [PASS] {name}" + (f" · {detail}" if detail else ""))
66
- else:
67
- FAIL += 1
68
- FAILS.append(name)
69
- # 无 cid 的失败记哨兵:红基线模式下它会与预期红项不等 ⇒ fail-closed,
70
- # 不静默放过「变异波及了没被登记的断言」(同 test_logref.py:66 的口径)。
71
- RED.append(cid or ("?" + name))
72
- print(f" [FAIL] {name} · {detail}")
73
-
74
-
75
- ALPHA = ('"""能量计算模块。"""\n'
76
- 'import math\n'
77
- '\n'
78
- '\n'
79
- 'def compute_energy(mass, speed):\n'
80
- ' """计算动能。"""\n'
81
- ' return mass * speed * speed\n'
82
- '\n'
83
- '\n'
84
- 'class Reactor:\n'
85
- ' """反应堆模型。"""\n'
86
- '\n'
87
- ' def ignite(self):\n'
88
- ' """点火。"""\n'
89
- ' self.state = "on"\n')
90
-
91
- # 同函数同区间,只改一行实现 → 行数不变,只有区间哈希变(最强的漂移信号)
92
- ALPHA_DRIFT = ALPHA.replace("return mass * speed * speed",
93
- "return 0.5 * mass * speed * speed")
94
- # 第二个漂移变体:段⑤ 已把源改成 ALPHA_DRIFT 并重建索引,段⑨ 要再制造一次真实漂移,
95
- # 必须写出**与当前索引内容不同**的正文;否则只是 mtime 变了、内容没变,报 ok 才对。
96
- ALPHA_DRIFT2 = ALPHA.replace("return mass * speed * speed",
97
- "return mass * speed * speed / 2.0")
98
- ALPHA_MARK2 = "speed * speed / 2.0"
99
-
100
- BETA = ('"""辅助模块。"""\n'
101
- '\n'
102
- '\n'
103
- 'def helper_scale(x):\n'
104
- ' """缩放。"""\n'
105
- ' return x * 2\n')
106
-
107
-
108
- def rd(path):
109
- with open(path, encoding="utf-8") as f:
110
- return f.read()
111
-
112
-
113
- def wr(path, text, bump=0.0):
114
- with open(path, "w", encoding="utf-8") as f:
115
- f.write(text)
116
- if bump:
117
- # 显式推后 mtime:粗粒度文件系统上保证「变了」能被水位看见
118
- t = os.path.getmtime(path) + bump
119
- os.utime(path, (t, t))
120
-
121
-
122
- def index_read(cg, code_dir, **extra):
123
- a = {"op": "index_code", "path": code_dir}
124
- a.update(extra)
125
- return call_tool(cg, "cg", a)
126
-
127
-
128
- def ref_read(cg, **args):
129
- a = {"op": "ref"}
130
- a.update(args)
131
- return call_tool(cg, "cg", a)
132
-
133
-
134
- # ---------------- 6 判据分工:节点 ref 判漂移,水位只判跳过 ----------------
135
-
136
- #: 合成源:改一行实现 → 行数不变、只有区间哈希变(最强的漂移信号,同 ALPHA_DRIFT)。
137
- PROBE_SRC = ('"""probe 模块。"""\n'
138
- '\n'
139
- '\n'
140
- 'def probe_fn(x):\n'
141
- ' """probe 函数。"""\n'
142
- ' return x + 1\n')
143
- PROBE_SRC2 = PROBE_SRC.replace("return x + 1", "return x + 2")
144
-
145
-
146
- def _mk(tmp, name):
147
- d = os.path.join(tmp, name)
148
- os.makedirs(d)
149
- return d
150
-
151
-
152
- def _index_into(cg, src, ledger=None):
153
- """直调 refindex 的两步(index_dir → add_items),返回本轮节点 id。
154
-
155
- 必须直调:`op=index_code` 是这两步的**连写**,表达不了「只写一侧」的分歧。
156
- """
157
- items, _errors, _stats = refindex.index_dir(src, kind="code_ref",
158
- ledger=ledger)
159
- ids, _sens = refindex.add_items(cg, items, kind="code_ref", root=src)
160
- return ids
161
-
162
-
163
- def _tamper_ref_hash(node_path, fake="deadbeefcafe"):
164
- """把节点 .md 里 frontmatter 的 ref 哈希改成假值(源与水位都不动)。
165
-
166
- `cg.get` 每次从盘上读节点,故改 .md 即改「节点自述」;水位条目不受影响——
167
- 这正是「漂移只由节点 ref 判」要区分的场景(v1 拿水位当探测数据 ⇒ 静默)。
168
- """
169
- with open(node_path, encoding="utf-8") as f:
170
- txt = f.read()
171
- out = re.sub(r'"hash": "[0-9a-f]{12}"', '"hash": "%s"' % fake, txt, count=1)
172
- if out == txt:
173
- raise RuntimeError("夹具失效:节点 frontmatter 里没有可篡改的 ref.hash")
174
- with open(node_path, "w", encoding="utf-8") as f:
175
- f.write(out)
176
-
177
-
178
- def _verdict(res):
179
- """巡检结果 → 去重/归因视角的摘要(红基线断言只看这里)。"""
180
- ids = [r.get("node_id") for r in list(res.get("stale") or [])
181
- + list(res.get("dangling") or [])]
182
- return {"ok": res["ok"], "ids": ids, "dup": len(ids) - len(set(ids)),
183
- "mm": res.get("ledger_mismatch") or [],
184
- "mm_count": res.get("ledger_mismatch_count", 0)}
185
-
186
-
187
- def guard_verdict_owner(tmp):
188
- """三组构造:水位与节点分别推到不同代,看快路径认谁(直调 refindex)。"""
189
- # ①水位新、节点旧 —— 等价 mcp_server.py 的 index_dir→add_items 之间被杀
190
- lib = _mk(tmp, "fo_lib")
191
- src = _mk(tmp, "fo_src")
192
- p = os.path.join(src, "probe_widget.py")
193
- wr(p, PROBE_SRC)
194
- cg = MdCG(lib)
195
- ids = _index_into(cg, src, refindex.Ledger(lib))
196
- cg.flush()
197
- wr(p, PROBE_SRC2, bump=5.0)
198
- refindex.index_dir(src, kind="code_ref", ledger=refindex.Ledger(lib))
199
- res = refindex.check_refs(cg, ledger=refindex.Ledger(lib))
200
- v = _verdict(res)
201
- node_hashes = {refindex.ref_of(cg.get(n))[1]["hash"] for n in ids}
202
- check("6①只推水位:ok=False(漂移由节点 ref 判,水位说了不算)",
203
- res["ok"] is False, f"ok={res['ok']} stale={len(res['stale'])}", cid="A1")
204
- check("6①stale 覆盖该文件全部节点,hash_expected 取节点 ref(不是水位新哈希)",
205
- {r["node_id"] for r in res["stale"]} == set(ids)
206
- and {r["hash_expected"] for r in res["stale"]} == node_hashes,
207
- f"stale={sorted(r['node_id'] for r in res['stale'])}", cid="A1")
208
- check("6①水位/节点不一致记归因行 watermark_behind(写序断裂可见,不进 ok)",
209
- v["mm_count"] >= 1
210
- and all(r["kind_of_gap"] == "watermark_behind" for r in v["mm"]),
211
- str(v["mm"][:1]), cid="A2")
212
-
213
- # ②节点新、水位旧 —— ledger 根本不参与,才算「只推节点」
214
- lib = _mk(tmp, "fn_lib")
215
- src = _mk(tmp, "fn_src")
216
- p = os.path.join(src, "probe_widget.py")
217
- wr(p, PROBE_SRC)
218
- cg = MdCG(lib)
219
- _index_into(cg, src, refindex.Ledger(lib))
220
- cg.flush()
221
- wr(p, PROBE_SRC2, bump=5.0)
222
- _index_into(cg, src, ledger=None) # 全量重切,水位不动
223
- cg.flush()
224
- res = refindex.check_refs(cg, ledger=refindex.Ledger(lib))
225
- v = _verdict(res)
226
- check("6②只推节点:ok=True 且 stale/dangling 空(旧水位不得误报漂移)",
227
- res["ok"] is True and not res["stale"] and not res["dangling"],
228
- f"ok={res['ok']} stale={sorted(v['ids'])}", cid="B1")
229
- # 归因面与「红不红」解耦:这里只断言「不一致被记下来了」(ok 由 B1 断),
230
- # 免得同一次误报把两条断言一起打红、红项表失去「一项一行为」的对应。
231
- check("6②旧水位与节点不一致仍归因(watermark_behind)",
232
- v["mm_count"] >= 1
233
- and all(r["kind_of_gap"] == "watermark_behind" for r in v["mm"]),
234
- str(v["mm"][:1]), cid="B2")
235
-
236
- # ③篡改节点 ref.hash(源与水位都不动)—— 只有节点自述变了
237
- lib = _mk(tmp, "ft_lib")
238
- src = _mk(tmp, "ft_src")
239
- p = os.path.join(src, "probe_widget.py")
240
- wr(p, PROBE_SRC)
241
- cg = MdCG(lib)
242
- ids = _index_into(cg, src, refindex.Ledger(lib))
243
- cg.flush()
244
- nid = sorted(ids)[0]
245
- _tamper_ref_hash(os.path.join(lib, cg.index["nodes"][nid]["path"]))
246
- res = refindex.check_refs(cg, ledger=refindex.Ledger(lib))
247
- v = _verdict(res)
248
- check("6③篡改节点 ref.hash(源与水位都不动):必须报 stale",
249
- res["ok"] is False and any(r["node_id"] == nid for r in res["stale"]),
250
- f"ok={res['ok']} stale={v['ids']}", cid="C1")
251
- check("6③同一 node_id 至多一行(去重,不重复报红)", v["dup"] == 0,
252
- str(v["ids"]), cid="C2")
253
-
254
-
255
- def guard_collision(tmp):
256
- """撞 id 现场(D1 状态):钉现状事实 + 去重 + 归因,不作期望语义断言。
257
-
258
- `node_id` 不含 root ⇒ 两域同 rel path + 同 name 时撞 id、节点被**后写域**覆盖
259
- (`add_items` 同 id upsert 整条替换)。本组断言这个**现状**(以及归因行可见、
260
- 不重复报红);键一旦改为含 root,第一条会转红以提示同步迁移——那是迁移触发器,
261
- 不是要实现的期望语义(见 docs/plans/真源索引_通用机制_v0.3.md「id 语义与设计债」)。
262
- """
263
- lib = _mk(tmp, "fc_lib")
264
- domA = _mk(tmp, "fc_domA")
265
- domB = _mk(tmp, "fc_domB")
266
- for d in (domA, domB):
267
- wr(os.path.join(d, "probe_widget.py"), PROBE_SRC)
268
- cg = MdCG(lib)
269
- ids_a = _index_into(cg, domA, refindex.Ledger(lib))
270
- # #352-E(2026-10-15):id 键**仍未**含 root,但写入面加了冲突闸——B 的同名条目
271
- # 不再覆写 A,改为「不写入 + 回执 conflicts 点名」。故本钉从「两域 id 集相等」
272
- # (=后写域覆盖)改为「两域算出的 id 仍相同,但 B 一个都不写」。原钉的用途
273
- # (键一旦改为含 root 即转红以提示同步迁移)保留:改 id 键后这里仍会转红。
274
- _items_b, _eb, _sb = refindex.index_dir(domB, kind="code_ref",
275
- ledger=refindex.Ledger(lib))
276
- _cf = []
277
- ids_b, _ = refindex.add_items(cg, _items_b, kind="code_ref", root=domB,
278
- conflicts=_cf)
279
- cg.flush()
280
- nid = sorted(ids_a)[0]
281
- ref = refindex.ref_of(cg.get(nid))[1] or {}
282
- check("7①两域同 rel path + 同 name ⇒ 仍算出同一 id,但后写域被冲突闸拦下不写入",
283
- bool(ids_a) and ids_b == []
284
- and {c.get("id") for c in _cf} == set(ids_a),
285
- f"A={ids_a} B={ids_b} conflicts={_cf}", cid="D1")
286
- check("7①同一 id 只留一条节点,ref.root 仍属**先写域**(#352-E 后的钉子)",
287
- len(cg.index["nodes"]) == len(ids_a)
288
- and refindex._same_root(ref.get("root"), domA),
289
- f"nodes={len(cg.index['nodes'])} root={ref.get('root')}", cid="D1")
290
- _tamper_ref_hash(os.path.join(lib, cg.index["nodes"][nid]["path"]))
291
- res = refindex.check_refs(cg, ledger=refindex.Ledger(lib))
292
- v = _verdict(res)
293
- check("7②该现场篡改节点 ref.hash:同一 node_id 至多一行(dup=0)",
294
- v["dup"] == 0, str(v["ids"]), cid="D2")
295
- check("7②水位/节点不一致记 ledger_mismatch 归因行(撞 id 不再静默)",
296
- v["mm_count"] >= 1, str(v["mm_count"]), cid="D2")
297
- # #352-E:撞 id 的归因从**巡检面**(watermark/node 两侧 root 不一致)前移到
298
- # **写入面回执**(conflicts 带 existing_root/root)——写入面不再让撞 id 落盘,
299
- # 巡检面自然不再有「节点归了后写域」这个事实可归因。归因字段本身(两侧 root)
300
- # 一个没少,只是换了承担者。
301
- check("7②撞 id 归因改由写入面回执承担(conflicts 带两侧 root,可归因到域)",
302
- bool(_cf) and all(c.get("existing_root") and c.get("root")
303
- and c.get("id") for c in _cf),
304
- str(_cf[:1]), cid="D2")
305
- # 该现场再推一步:**两域的源都改掉**(合法漂移,不是篡改)。v1 逐条水位条目各自
306
- # 探测 ⇒ 同一个 node_id 被反复报(本夹具实测 HEAD 4 行 / dup=2 ⇒ v2 2 行 / dup=0,
307
- # 两个数是该文件的两个节点各一行)。这一步才真正钉住「去重」,且不依赖篡改。
308
- wr(os.path.join(domA, "probe_widget.py"), PROBE_SRC2, bump=5.0)
309
- wr(os.path.join(domB, "probe_widget.py"), PROBE_SRC2, bump=5.0)
310
- v = _verdict(refindex.check_refs(cg, ledger=refindex.Ledger(lib)))
311
- check("7③两域源都改:同一 node_id 至多一行(v1 会按水位条目重复报)",
312
- v["dup"] == 0 and bool(v["ids"]), str(v["ids"]), cid="D3")
313
- _items_b2, _eb2, _sb2 = refindex.index_dir(domB, kind="code_ref",
314
- ledger=refindex.Ledger(lib))
315
- _cf2 = []
316
- _ids_b2, _ = refindex.add_items(cg, _items_b2, kind="code_ref",
317
- root=domB, conflicts=_cf2)
318
- check("7③两域源都改后再索引 B:撞 id 仍由回执点名(带两侧 root,可归因到域)",
319
- _ids_b2 == [] and bool(_cf2)
320
- and all(c.get("existing_root") and c.get("root") for c in _cf2),
321
- f"ids={_ids_b2} conflicts={_cf2[:1]}", cid="D3")
322
-
323
-
324
- # ---------------- 12 自愈止血 + 水位保全(H1/H1b/H2/H3/H4/H5/H6 + 写序) ----------------
325
-
326
- def guard_heal_bounded(tmp):
327
- """水位不被写空 / 重建失败不许记成成功 / 同因不重试(直调 refindex|sustain)。
328
-
329
- 全部用对称小夹具(每域 1 个源文件、2 个节点),断言的是**行为**而不是文案。
330
- """
331
- # ①水位保全(H1):目录仍在、零文件 ⇒ 空扫不对账,水位不被剪空
332
- lib = _mk(tmp, "hb1_lib")
333
- src = _mk(tmp, "hb1_src")
334
- p = os.path.join(src, "hb1_widget.py")
335
- wr(p, PROBE_SRC)
336
- cg = MdCG(lib)
337
- ids = _index_into(cg, src, refindex.Ledger(lib))
338
- cg.flush()
339
- os.remove(p) # 目录仍在、零文件
340
- refindex.index_dir(src, kind="code_ref", ledger=refindex.Ledger(lib))
341
- s = refindex.Ledger(lib).summary()
342
- check("12①空扫不对账:水位条目与节点数保住([C1] 目录仍在零文件)",
343
- s["files"] == 1 and s["nodes"] == len(ids),
344
- f"files={s['files']} nodes={s['nodes']} ids={len(ids)}", cid="E1")
345
- check("12①两面不打架:last_index 如实记 files=0 且 empty_scan=True",
346
- (s.get("last_index") or {}).get("files") == 0
347
- and (s.get("last_index") or {}).get("empty_scan") is True,
348
- str(s.get("last_index"))[:120], cid="E1")
349
-
350
- # ②悬空 root 不重建(H1b):不调 index_dir ⇒ 水位不被写成空
351
- lib = _mk(tmp, "hb2_lib")
352
- src = _mk(tmp, "hb2_src")
353
- wr(os.path.join(src, "hb2_widget.py"), PROBE_SRC)
354
- cg = MdCG(lib)
355
- _index_into(cg, src, refindex.Ledger(lib))
356
- cg.flush()
357
- n_files = refindex.Ledger(lib).summary()["files"]
358
- shutil.rmtree(src) # 源大域整个没了
359
- r = refindex.rebuild(cg, ledger=refindex.Ledger(lib))
360
- check("12②rebuild 遇悬空 root:不重建、ok=False、roots_missing 如实记",
361
- r["ok"] is False and len(r["roots_missing"]) == 1 and r["indexed"] == 0
362
- and any("已不是目录" in e for e in r["errors"]),
363
- f"ok={r['ok']} missing={r['roots_missing']} err={r['errors'][:1]}", cid="E2")
364
- check("12②悬空 root 不再把水位写成空([C3] files 保住 vs 现状 0)",
365
- refindex.Ledger(lib).summary()["files"] == n_files,
366
- f"files={refindex.Ledger(lib).summary()['files']} 期望 {n_files}", cid="E2")
367
-
368
- # ③空扫组如实置 ok=False(H2)
369
- lib = _mk(tmp, "hb3_lib")
370
- src = _mk(tmp, "hb3_src")
371
- p = os.path.join(src, "hb3_widget.py")
372
- wr(p, PROBE_SRC)
373
- cg = MdCG(lib)
374
- _index_into(cg, src, refindex.Ledger(lib))
375
- cg.flush()
376
- os.remove(p)
377
- r = refindex.rebuild(cg, ledger=refindex.Ledger(lib))
378
- check("12③rebuild 空扫组:ok=False + errors + 逐组 empty_scan 如实回报",
379
- r["ok"] is False and r["errors"] and r["rebuilt"]
380
- and r["rebuilt"][0]["empty_scan"] is True
381
- and r["rebuilt"][0]["files"] == 0,
382
- str(r["rebuilt"])[:140], cid="E3")
383
-
384
- # ④只有悬空 ⇒ heal 不重建(H3):止血点在「按 code 分派」
385
- lib = _mk(tmp, "hb4_lib")
386
- domA = _mk(tmp, "hb4_domA")
387
- domB = _mk(tmp, "hb4_domB")
388
- for d, nm in ((domA, "hb4_a.py"), (domB, "hb4_b.py")):
389
- wr(os.path.join(d, nm), PROBE_SRC)
390
- cg = MdCG(lib)
391
- _index_into(cg, domA, refindex.Ledger(lib))
392
- _index_into(cg, domB, refindex.Ledger(lib))
393
- cg.flush()
394
- shutil.rmtree(domB)
395
- codes = {i["code"] for i in sustain.diagnose(cg)["issues"]}
396
- acts = [a["code"] for a in sustain.heal(cg)["actions"]]
397
- check("12④ref_dangling 不再触发重建(fix 字段无调度器消费,改它无效)",
398
- "ref_dangling" in codes and "rebuild_refs" not in acts,
399
- f"codes={sorted(codes)} actions={acts}", cid="E4")
400
-
401
- # ⑤重建失败如实记 ok=False(H4)+ ⑥同因不重试(H6)
402
- lib = _mk(tmp, "hb5_lib")
403
- domA = _mk(tmp, "hb5_domA")
404
- domB = _mk(tmp, "hb5_domB")
405
- for d, nm in ((domA, "hb5_a.py"), (domB, "hb5_b.py")):
406
- wr(os.path.join(d, nm), PROBE_SRC)
407
- cg = MdCG(lib)
408
- _index_into(cg, domA, refindex.Ledger(lib))
409
- _index_into(cg, domB, refindex.Ledger(lib))
410
- cg.flush()
411
- key = (os.path.abspath(lib), "rebuild_refs")
412
- sustain._HEAL_MEMO.pop(key, None)
413
- p = os.path.join(domA, "hb5_a.py")
414
- wr(p, PROBE_SRC2, bump=5.0) # 一个域漂移
415
- shutil.rmtree(domB) # 另一个域悬空(重建必带 roots_missing)
416
- h1 = sustain.heal(cg, repeat_guard=True, heal_interval=300.0)
417
- row = next((a for a in h1["actions"] if a["code"] == "rebuild_refs"), None)
418
- check("12⑤重建失败不得记成成功:applied=True 且 ok=False,并带结果摘要",
419
- bool(row) and row["applied"] is True and row["ok"] is False
420
- and bool((row.get("result") or {}).get("roots_missing")),
421
- str(row)[:170], cid="E5")
422
- wr(p, PROBE_SRC, bump=5.0) # 回到与上次相同的诊断信号
423
- h2 = sustain.heal(cg, repeat_guard=True, heal_interval=300.0)
424
- row2 = next((a for a in h2["actions"] if a["code"] == "rebuild_refs"), None)
425
- check("12⑥同一信号重复失败:拦下并如实记 repeat_failure(不是静默、不重刷)",
426
- bool(row2) and row2["applied"] is False
427
- and row2.get("reason") == "repeat_failure"
428
- and row2.get("streak") == 1 and row2.get("wait_s") == 300.0,
429
- str(row2)[:170], cid="E6")
430
- sustain._HEAL_MEMO[key]["ts"] = time.time() - 400 # 模拟退避窗已过
431
- h3 = sustain.heal(cg, repeat_guard=True, heal_interval=300.0)
432
- row3 = next((a for a in h3["actions"] if a["code"] == "rebuild_refs"), None)
433
- check("12⑥退避窗过后放行重试(退避 ≠ 永久静默)",
434
- bool(row3) and row3["applied"] is True, str(row3)[:170], cid="E6")
435
- sustain._HEAL_MEMO.clear()
436
-
437
- # ⑦sources 的水位面(H5):Ledger 没有 stat(),写错被 except 吞成 {}
438
- st = sources.FileDispatcher(cg)._ledger_stat()
439
- check("12⑦ingest stat 的水位面不再被 AttributeError 吞成空",
440
- isinstance(st, dict) and "files" in st and "path" in st,
441
- str(sorted(st))[:120], cid="E7")
442
-
443
- # ⑧写序(#2):节点写入失败 ⇒ 水位不得先落盘
444
- lib = _mk(tmp, "hb6_lib")
445
- src = _mk(tmp, "hb6_src")
446
- wr(os.path.join(src, "hb6_widget.py"), PROBE_SRC)
447
- cg = MdCG(lib)
448
- orig = refindex.add_items
449
-
450
- # 生效条件:任何入参都抛 RuntimeError(把「节点写入失败」注入 op 面),用于验证水位未先落盘;
451
- def _boom(*_a, **_k):
452
- raise RuntimeError("夹具:节点写入失败")
453
-
454
- refindex.add_items = _boom
455
- raised = False
456
- try:
457
- index_read(cg, src)
458
- except RuntimeError:
459
- raised = True
460
- finally:
461
- refindex.add_items = orig
462
- led = refindex.Ledger(lib)
463
- check("12⑧节点写入失败时水位不落盘(残留方向=节点新/水位旧,不是反向)",
464
- raised and not os.path.isfile(led.path),
465
- f"raised={raised} ledger_exists={os.path.isfile(led.path)}", cid="E8")
466
-
467
-
468
- # ---------------- 13 口径钉子:id 位置化 + 幂等判据由区间哈希承担(#5) ----------------
469
-
470
- #: 口径真源(文档侧)。措辞与实现必须同源——改一边不改另一边即转红。
471
- DOC_PLAN = os.path.join(_BASE, "docs", "plans", "真源索引_通用机制_v0.3.md")
472
- #: 口径块定位:起=该标题行,止=下一个**二级**标题(`### ` 不算终止)。
473
- DOC_SEC_HEAD = "### 3.1 "
474
-
475
- #: §3.1(归一化后)的 sha1 前 16 位:**判据侧陈化硬失败**——口径一被改写(无论改真源
476
- #: 还是改守卫)就必须同步本行,才谈得上「措辞与实现同源」。归一化=空白折叠 ⇒ 换行 /
477
- #: 缩进 / 表格对齐**不**误红,只有措辞变化才红。失败信息里打印现算值,照抄即恢复。
478
- DOC_SEC_FP = "f373ca6380e3debf"
479
-
480
- #: §3.1 里必须**逐字**出现的口径字面(缺一条 ⇒ 措辞已从实现侧漂开)。
481
- DOC_REQUIRED = (
482
- "路径-标题寻址", # id 是位置化寻址(不是内容寻址)
483
- "sha1(路径::符号名)", # code_ 的口径公式
484
- "sha1(路径#标题路径)", # doc_ 的口径公式
485
- "不含 root、不含正文", # 位置化的两个「不含」
486
- "id 存在即已索引", # D2:明令禁止的判据(幂等不许看 id)
487
- "hash + 行位全等", # 幂等判据的实际承担者
488
- )
489
-
490
- #: 行级引用 → 实现符号:(文档里必须出现的引用字面, 文件, 行号, 该行应以什么开头)。
491
- #: 「文档说谁在哪一行」是措辞里最硬的一句——引用漂移(代码增删行)即红,失败信息附
492
- #: 该符号**现址**,照抄回文档与本表即恢复(陈化可自愈到「人只需抄一次」)。
493
- DOC_CITATIONS = (
494
- ("refindex.py:577", "md_cg/refindex.py", 577, "def _code_ref("),
495
- ("refindex.py:591", "md_cg/refindex.py", 591, "def _doc_ref("),
496
- ("refindex.py:635", "md_cg/refindex.py", 635,
497
- "got = codeindex.region_hash(lines, lineno, end)"),
498
- ("refindex.py:652", "md_cg/refindex.py", 652, "def read_ref("),
499
- ("refindex.py:720", "md_cg/refindex.py", 720, "def check_refs("),
500
- ("refindex.py:931", "md_cg/refindex.py", 931, "def prune_orphans("),
501
- ("srcindex.py:71", "md_cg/srcindex.py", 71, "def src_id("),
502
- ("srcindex.py:86", "md_cg/srcindex.py", 86, "def file_hash("),
503
- ("logref.py:119", "md_cg/logref.py", 119, "def node_state("),
504
- ("docindex.py:294", "md_cg/docindex.py", 294, "def node_id("),
505
- )
506
-
507
-
508
- def _doc_block(path: str, head: str) -> str:
509
- """取文档里 `head` 起、到下一个**二级**标题前的正文块,空白折叠后返回。
510
-
511
- 起行不在文档里返回 ""(由调用方判红,不留静默分支)。归一化只折空白——
512
- 重排 / 换行 / 表格对齐不影响指纹,只有**措辞**变化才改指纹。
513
- """
514
- try:
515
- with open(path, encoding="utf-8") as f:
516
- lines = f.read().replace("\r\n", "\n").split("\n")
517
- except OSError:
518
- return "" # 文档缺失=口径真源不可读,交由调用方判红
519
- start = next((i for i, ln in enumerate(lines) if ln.startswith(head)), None)
520
- if start is None:
521
- return ""
522
- i = start + 1
523
- while i < len(lines) and not (lines[i].startswith("## ")
524
- and not lines[i].startswith("### ")):
525
- i += 1
526
- return re.sub(r"\s+", " ", "\n".join(lines[start:i])).strip()
527
-
528
-
529
- def guard_id_and_idempotence(tmp):
530
- """#5 口径钉子:id 位置化(不含 root / 正文 / 行位)、幂等判据由区间哈希 + 行位承担。
531
-
532
- 两侧都钉,谁单方面漂开都转红:①**实现侧**——id 用手算 sha1 独立复算(不借实现),
533
- 对 root / 正文哈希 / 行位不敏感;水位与节点 ref 五项全等才是跳过的前提(「id 存在」
534
- 不构成已索引);「同 id 而内容变」必须判 stale。②**文档侧**——`v0.3` 计划 §3.1 的
535
- 口径字面齐全、整块指纹不陈化、行级引用仍指向所声称的符号。
536
- """
537
- from . import codeindex, docindex
538
-
539
- # ---- ①id 口径:位置化,公式可手算复算 ----
540
- it = {"path": "pkg/probe_widget.py", "name": "probe_fn", "kind": "function",
541
- "lineno": 4, "end": 6, "hash": "aaaa1111bbbb", "lang": "py",
542
- "precise": True}
543
- want_code = "code_" + hashlib.sha1(
544
- b"pkg/probe_widget.py::probe_fn").hexdigest()[:12]
545
- check("13①code_ == sha1(路径::符号名)[:12](手算交叉核验,非复用实现)",
546
- codeindex.node_id(it) == want_code
547
- and refindex.node_id_of(it, "code_ref") == want_code,
548
- f"{codeindex.node_id(it)} vs {want_code}", cid="F1")
549
- dit = {"path": "docs/probe.md", "heading_path": ["一", "二"], "heading": "二",
550
- "level": 2, "lineno": 10, "end": 20, "hash": "cccc2222", "dup": 1}
551
- want_doc = "doc_" + hashlib.sha1(
552
- "docs/probe.md#一/二".encode("utf-8")).hexdigest()[:12]
553
- check("13①doc_ == sha1(路径#标题路径)[:12](手算交叉核验)",
554
- docindex.node_id(dit) == want_doc
555
- and refindex.node_id_of(dit, "doc_ref") == want_doc,
556
- f"{docindex.node_id(dit)} vs {want_doc}", cid="F1")
557
- check("13②id 不读 root(跨域同 rel path 必然同 id —— D1 的机械根)",
558
- codeindex.node_id(dict(it, root="/domA")) == want_code
559
- and codeindex.node_id(dict(it, root="/domB")) == want_code
560
- and docindex.node_id(dict(dit, root="/domA")) == want_doc,
561
- cid="F2")
562
- check("13②id 不读正文哈希与行位(改内容 / 挪行位 id 不变 ⇒ 位置化而非内容寻址)",
563
- codeindex.node_id(dict(it, hash="deadbeefdead", lineno=99, end=120)) == want_code
564
- and docindex.node_id(dict(dit, hash="zzzz9999", lineno=1, end=2)) == want_doc,
565
- cid="F2")
566
-
567
- # ---- ②幂等判据:水位只记「能不能跳过」,判据键 = 节点 ref 的五项 + 源 size/mtime ----
568
- lib = _mk(tmp, "fx_lib")
569
- src = _mk(tmp, "fx_src")
570
- p = os.path.join(src, "probe_widget.py")
571
- wr(p, PROBE_SRC)
572
- cg = MdCG(lib)
573
- ids = _index_into(cg, src, refindex.Ledger(lib))
574
- cg.flush()
575
- nid = sorted(ids)[0]
576
- _k0, ref0 = refindex.ref_of(cg.get(nid))
577
- ents = [e for e in (refindex.Ledger(lib).load().get("files") or {}).values()
578
- if (e.get("path") or "").endswith("probe_widget.py")]
579
- e = ents[0] if ents else {}
580
- n0 = (e.get("nodes") or [{}])[0]
581
- check("13③水位键 = 源文件绝对路径且只有一条(跨域不撞名)",
582
- len(ents) == 1, f"{len(ents)} 条", cid="F3")
583
- check("13③水位条目记的正是判据要用的五项(root/path/lineno/end/hash)+ 源 size/mtime",
584
- {"root", "path", "size", "mtime", "kind", "nodes"} <= set(e)
585
- and {"id", "lineno", "end", "hash"} <= set(n0), sorted(e), cid="F3")
586
- check("13③稳态:水位与节点 ref 五项全等(跳过的前提真的成立)",
587
- refindex._same_root(e.get("root"), (ref0 or {}).get("root"))
588
- and refindex._norm_rel(e.get("path")) == refindex._norm_rel((ref0 or {}).get("path"))
589
- and n0.get("hash") == (ref0 or {}).get("hash")
590
- and n0.get("lineno") == (ref0 or {}).get("lineno")
591
- and n0.get("end") == (ref0 or {}).get("end"),
592
- f"watermark={n0} ref={ref0}", cid="F3")
593
- # 逐项:五项全等 + 源未变 ⇒ 跳过;任一项不等 ⇒ 一律进探测(「id 存在」不构成已索引)
594
- n_rec = {"id": nid, "lineno": n0.get("lineno"), "end": n0.get("end"),
595
- "hash": n0.get("hash")}
596
- e_rec = {"root": e.get("root"), "path": e.get("path")}
597
- rel = refindex._norm_rel(e.get("path"))
598
- base = refindex._fast_path_verdict(cg, n_rec, e_rec, rel, e.get("root"), True)
599
- check("13④五项全等 + 源 size/mtime 未变 ⇒ 跳过(probe=None,不读源)",
600
- base["covered"] is True and base["probe"] is None, str(base), cid="F4")
601
- holes = []
602
- for k, v in (("hash", "000000000000"), ("lineno", 999), ("end", 1000)):
603
- n2 = dict(n_rec)
604
- n2[k] = v
605
- if refindex._fast_path_verdict(cg, n2, e_rec, rel, e.get("root"),
606
- True)["probe"] is None:
607
- holes.append(k)
608
- if refindex._fast_path_verdict(cg, n_rec, e_rec, "other_widget.py",
609
- e.get("root"), True)["probe"] is None:
610
- holes.append("path")
611
- v_root = refindex._fast_path_verdict(cg, n_rec, e_rec, rel,
612
- os.path.join(src, "elsewhere"), True)
613
- if v_root["probe"] is None:
614
- holes.append("root")
615
- check("13④五项任一项不等 ⇒ 一律不跳过(「id 存在」不构成已索引)", not holes,
616
- f"漏判项={holes}", cid="F4")
617
- check("13④跨域不一致归因 collision_multi_root 且带两侧 root(不静默、不改判定)",
618
- (v_root.get("mismatch") or {}).get("kind_of_gap") == "collision_multi_root"
619
- and bool((v_root.get("mismatch") or {}).get("watermark_root"))
620
- and bool((v_root.get("mismatch") or {}).get("node_root")),
621
- str(v_root.get("mismatch")), cid="F4")
622
-
623
- # ---- ③同 id 而内容变:必须判漂移(id 只做身份,不承担幂等) ----
624
- wr(p, PROBE_SRC2, bump=5.0) # 行数不变,只有区间哈希变
625
- res = refindex.check_refs(cg, ledger=refindex.Ledger(lib))
626
- row = next((r for r in res["stale"] if r["node_id"] == nid), None)
627
- check("13⑤同 id 而区间哈希变(源改一行)⇒ 必须判 stale,且 hash_expected 取节点 ref",
628
- res["ok"] is False and row is not None
629
- and row.get("hash_expected") == (ref0 or {}).get("hash"),
630
- f"ok={res['ok']} stale={sorted(r['node_id'] for r in res['stale'])}",
631
- cid="F5")
632
-
633
- # ---- ④文档侧:措辞 / 指纹 / 行级引用都必须与实现同源 ----
634
- sec = _doc_block(DOC_PLAN, DOC_SEC_HEAD)
635
- check("13⑥口径真源可读:文档 §3.1 块非空", bool(sec), DOC_PLAN, cid="F6")
636
- missing = [lit for lit in DOC_REQUIRED if lit not in sec]
637
- check("13⑥文档 §3.1 逐字含口径字面(id 位置化 + 幂等判据不看 id)", not missing,
638
- f"缺字面={missing}", cid="F6")
639
- fp = hashlib.sha1(sec.encode("utf-8")).hexdigest()[:16]
640
- check("13⑥文档 §3.1 指纹未陈化(措辞一改即红,须同步 DOC_SEC_FP)",
641
- fp == DOC_SEC_FP, f"现={fp} 判据侧={DOC_SEC_FP}(照抄现算值即恢复)",
642
- cid="F6")
643
- doc_text = ""
644
- try:
645
- with open(DOC_PLAN, encoding="utf-8") as f:
646
- doc_text = f.read()
647
- except OSError:
648
- pass # 缺文档 ⇒ 每条引用都记「未出现」,逐条报红
649
- bad = []
650
- for literal, fpath, lineno, claim in DOC_CITATIONS:
651
- if literal not in doc_text:
652
- bad.append(f"{literal} 未出现在文档")
653
- continue
654
- lines = rd(os.path.join(_BASE, fpath)).replace("\r\n", "\n").split("\n")
655
- got = lines[lineno - 1] if 0 < lineno <= len(lines) else ""
656
- # 判据取 strip 后前缀:行内缩进不算「引用漂移」(只钉「这一行是不是这句」)
657
- if not got.strip().startswith(claim):
658
- now = next((i for i, ln in enumerate(lines, 1)
659
- if ln.strip().startswith(claim)), "?")
660
- bad.append(f"{fpath}:{lineno} 现为 {got.strip()[:36]!r}"
661
- f"({claim.strip()} 现在 {fpath}:{now})")
662
- check("13⑦文档行级引用仍指向所声称的符号(引用漂移即红,失败信息附现址)",
663
- not bad, ";".join(bad[:3]), cid="F7")
664
-
665
-
666
- # ---------------- 14 常驻循环止血(heal 在役面) ----------------
667
-
668
- def guard_loop_tick_heal(tmp):
669
- """`SustainLoop._tick_heal`:空转被止住 + 判据未被削弱(不静默、不假装修好)。
670
-
671
- 直调 `_tick_heal()`(**不起线程、不打心跳戳**),`d` 显式落到临时目录——
672
- 绝不写 ~/.mdcg。落点取循环自己的账(`heals` / `last_diagnose`)与 heal 的
673
- 返回值,不断言内部实现细节。
674
- """
675
- lib = _mk(tmp, "lt_lib")
676
- domA = _mk(tmp, "lt_domA")
677
- domB = _mk(tmp, "lt_domB")
678
- for d, nm in ((domA, "lt_a.py"), (domB, "lt_b.py")):
679
- wr(os.path.join(d, nm), PROBE_SRC)
680
- cg = MdCG(lib)
681
- _index_into(cg, domA, refindex.Ledger(lib))
682
- _index_into(cg, domB, refindex.Ledger(lib))
683
- cg.flush()
684
- sustain._HEAL_MEMO.pop((os.path.abspath(lib), "rebuild_refs"), None)
685
- pa = os.path.join(domA, "lt_a.py")
686
- wr(pa, PROBE_SRC2, bump=5.0) # 一域漂移
687
- shutil.rmtree(domB) # 一域悬空 ⇒ 重建必带 roots_missing
688
- loop = sustain.SustainLoop(cg, "p28_guard_loop", heal_interval=300.0,
689
- auto_heal=True, d=os.path.join(tmp, "lt_sustain"))
690
- orig = sustain.heal
691
- seen = {}
692
-
693
- def _spy(cg_, **kw):
694
- res = orig(cg_, **kw)
695
- seen["kw"] = kw
696
- seen["actions"] = res.get("actions")
697
- return res
698
-
699
- sustain.heal = _spy
700
- try:
701
- loop._tick_heal()
702
- finally:
703
- sustain.heal = orig
704
- check("14①常驻循环按 repeat_guard=True + heal_interval 调 heal(空转止在这一面)",
705
- (seen.get("kw") or {}).get("repeat_guard") is True
706
- and (seen.get("kw") or {}).get("heal_interval") == 300.0,
707
- str(sorted((seen.get("kw") or {}))), cid="G1")
708
- h1 = loop.heals[-1]
709
- check("14②首轮真跑并如实记没修好:actions 含 rebuild_refs 且 failed 单列该码",
710
- "rebuild_refs" in h1["actions"] and "rebuild_refs" in h1["failed"],
711
- str(h1), cid="G2")
712
- check("14③判据未被削弱:诊断面照旧报出病灶(ref_stale 仍在 last_diagnose)",
713
- "ref_stale" in ((loop.last_diagnose or {}).get("issues") or []),
714
- str(loop.last_diagnose), cid="G2")
715
- # 同信号第二轮:先让信号与上次一致(重建已把索引推到新代 ⇒ 写回旧正文再漂一次)
716
- wr(pa, PROBE_SRC, bump=5.0)
717
- sustain.heal = _spy
718
- try:
719
- loop._tick_heal()
720
- finally:
721
- sustain.heal = orig
722
- row2 = next((a for a in (seen.get("actions") or [])
723
- if a.get("code") == "rebuild_refs"), None)
724
- check("14④同一信号重复失败 ⇒ 被拦下并如实记 repeat_failure(不是又重跑一次)",
725
- bool(row2) and row2.get("applied") is False
726
- and row2.get("reason") == "repeat_failure" and row2.get("streak") == 1,
727
- str(row2), cid="G3")
728
- h2 = loop.heals[-1]
729
- check("14④拦下也照样进账:heals 记该动作 + failed 单列,诊断面仍报病灶(都不静默)",
730
- "rebuild_refs" in h2["actions"] and "rebuild_refs" in h2["failed"]
731
- and "ref_stale" in ((loop.last_diagnose or {}).get("issues") or []),
732
- str(h2), cid="G3")
733
- sustain._HEAL_MEMO.clear()
734
-
735
-
736
- # ---------------- 红基线(判别力自证) ----------------
737
-
738
- #: 定点变异表:(标签, 现实现锚点, 改回后的旧行为)。
739
- #: **源 = 工作区 md_cg/refindex.py 的文本**(不取 git HEAD:HEAD 在改动提交那一刻就等于
740
- #: 实现本身,判别力静默失效——同款教训见 md_cg/test_logref.py 的 _MUTATIONS 头注)。
741
- #: 锚点必须**恰好命中 1 次**:0 次=漂移、≥2 次=歧义,皆 ANCHOR-MISS → fail-closed。
742
- _MUTATIONS = (
743
- ("快路径探测数据:节点自身 frontmatter ref(现行)→ 水位条目(v1 旧行为)",
744
- ' out["probe"] = {**ref, "path": ref.get("path") or rel,\n'
745
- ' "root": ref.get("root") or src_root}',
746
- ' out["probe"] = {**n, "path": rel, "root": src_root}'),
747
- )
748
-
749
- #: 变异态下**必须**转红的项:恰好命中这个集合(多一项少一项都报红):
750
- #: A1 6①只推水位 ⇒ v1 拿水位当探测数据 ⇒ 哈希「匹配」而静默 ok=True(漏报)
751
- #: B1 6②只推节点 ⇒ v1 拿旧水位当探测数据 ⇒ 旧哈希对不上新源(误报 stale)
752
- #: C1 6③篡改节点 ref.hash ⇒ v1 根本不看节点自述 ⇒ 静默 ok=True
753
- EXPECTED_RED = ("A1", "B1", "C1")
754
-
755
-
756
- # 生效条件:锚点命中次数不是恰好 1 次时由 _mutated_verdict 抛出(fail-closed)。
757
- class AnchorMiss(RuntimeError):
758
- pass
759
-
760
-
761
- #: 导入时的实现体字节码(红基线模式会把变异体挂回该属性,故不能与属性现值比)
762
- _IMPL_CODEC = refindex._fast_path_verdict.__code__.co_code
763
-
764
-
765
- def _anchor_selfcheck() -> None:
766
- """锚点仍在、且 EXPECTED_RED 都是本文件里真实存在的 cid → 静默通过。"""
767
- _mutated_verdict() # 锚点漂移/歧义 → AnchorMiss
768
- with open(os.path.abspath(__file__), encoding="utf-8") as f:
769
- src = f.read()
770
- gone = [i for i in EXPECTED_RED if ('cid="%s"' % i) not in src]
771
- if gone:
772
- raise AnchorMiss("EXPECTED_RED 含本文件里不存在的 check id(红项表已陈化):%s"
773
- % "、".join(gone))
774
-
775
-
776
- def _mutated_verdict():
777
- """读工作区 `md_cg/refindex.py`(**只读,不落盘**)按 _MUTATIONS 逐处文本替换,
778
- 把变异后的模块体 exec 进真模块 globals 的副本,返回变异后的 _fast_path_verdict。
779
- 锚点漂移/歧义时抛 AnchorMiss。"""
780
- p = os.path.join(_BASE, "md_cg", "refindex.py")
781
- with open(p, encoding="utf-8") as f:
782
- text = f.read()
783
- for label, anchor, repl in _MUTATIONS:
784
- n = text.count(anchor)
785
- if n != 1:
786
- raise AnchorMiss(
787
- "ANCHOR-MISS:定点变异锚点漂移/歧义 —— %s 命中 %d 次(应恰好 1 次)"
788
- "|锚点首行:%s" % (label, n, anchor.splitlines()[0].strip()))
789
- text = text.replace(anchor, repl)
790
- g = dict(vars(refindex)) # __name__ / __package__ 保留 ⇒ 相对 import 照旧解析
791
- exec(compile(text, "<mutated md_cg/refindex.py>", "exec"), g)
792
- fn = g["_fast_path_verdict"]
793
- # 与**导入时**的实现体比(不能比当前属性:红基线模式会把变异体挂回该属性)
794
- if fn.__code__.co_code == _IMPL_CODEC:
795
- raise AnchorMiss("ANCHOR-MISS:变异后字节码与实现相同(替换未生效)")
796
- return fn
797
-
798
-
799
- def main(argv=None):
800
- ap = argparse.ArgumentParser(
801
- description="P28 守卫:ref 增量索引 + 漂移/悬空巡检(R3)")
802
- ap.add_argument("--head-baseline", action="store_true",
803
- help="红基线:对工作区 md_cg/refindex.py 做定点变异(快路径探测"
804
- "数据改回水位条目=v1 旧行为),断言红项集合 == 预期红项")
805
- args = ap.parse_args(argv)
806
- # 锚点自证:**两态都查**(绿态也查)——锚点漂移后若只查红基线,绿线会继续绿着
807
- # 而判别力已死(本守卫的判别力全靠这张表)。漂移即 fail-closed。
808
- try:
809
- _anchor_selfcheck()
810
- except AnchorMiss as exc:
811
- print("!! %s" % exc)
812
- print("结果:锚点自证未过(fail-closed)——定点变异表须同步实现")
813
- return 2
814
- if args.head_baseline:
815
- # 变异体注入(只在本进程内,不落盘)
816
- refindex._fast_path_verdict = _mutated_verdict()
817
- print("!! 红基线模式:md_cg/refindex.py 经定点变异 —— %s ⇒ 快路径改用水位条目"
818
- % ";".join(lbl for lbl, _a, _r in _MUTATIONS))
819
-
820
- print("=" * 68)
821
- print("md 认知图 P28 验收 · ref 增量索引 + 漂移/悬空巡检(R3)")
822
- print("=" * 68)
823
-
824
- tmp = tempfile.mkdtemp(prefix="mdcg_refcheck_")
825
- # 部署声明:本用例的读写白名单根(2026-10-10 设计者裁定 dsh #85 选 A 后,
826
- # MDCG_INGEST_ROOT / MDCG_EXPORT_ROOT 未配置**不再放开**,会回落 mdcg 记忆库根
827
- # 与工作区;本用例沙箱在两者之外,故必须显式声明——这正是新语义要求的
828
- # 「部署用环境变量声明可读写的根」。
829
- os.environ["MDCG_INGEST_ROOT"] = tmp
830
- os.environ["MDCG_EXPORT_ROOT"] = tmp
831
- src = os.path.join(tmp, "pkg")
832
- os.makedirs(src)
833
- alpha = os.path.join(src, "alpha.py")
834
- beta = os.path.join(src, "beta.py")
835
- wr(alpha, ALPHA)
836
- wr(beta, BETA)
837
-
838
- corpus.reset_root(ROOT)
839
- cg = MdCG(ROOT)
840
-
841
- try:
842
- # ============================================ ① 统一 ref 协议 + 注册表
843
- print("\n【1】统一 ref 协议与提取器注册表")
844
- check("tokens.ALL_OPS 仍收录 ref / index_code",
845
- "ref" in tokens.ALL_OPS and "index_code" in tokens.ALL_OPS)
846
- check("REF_KEYS = (code_ref, doc_ref)",
847
- refindex.REF_KEYS == ("code_ref", "doc_ref"), str(refindex.REF_KEYS))
848
- reg = refindex.registry()
849
- check("注册表覆盖 code_ref / doc_ref",
850
- "code_ref" in reg and "doc_ref" in reg, str(sorted(reg)))
851
- check("kind_of_path 按后缀判 kind",
852
- refindex.kind_of_path("a.py") == "code_ref"
853
- and refindex.kind_of_path("a.md") == "doc_ref"
854
- and refindex.kind_of_path("a.txt") == "",
855
- f"py={refindex.kind_of_path('a.py')} md={refindex.kind_of_path('a.md')}")
856
- try:
857
- refindex.extract("x = 1", "foo.xyz")
858
- raised = False
859
- except ValueError as exc:
860
- raised = "无提取器" in str(exc) or "无索引提取器" in str(exc)
861
- check("无提取器后缀显式报错(不静默)", raised)
862
-
863
- # ============================================ ② 全量索引 + 水位
864
- print("\n【2】index_code 全量 + _refindex.json 水位")
865
- out1 = index_read(cg, src)
866
- check("op=index_code ok", out1.get("ok"), str(out1)[:120])
867
- check("files == 2(命中后缀)", out1.get("files") == 2, str(out1.get("files")))
868
- check("indexed > 0 且 error_count == 0",
869
- out1.get("indexed", 0) > 0 and out1.get("error_count") == 0)
870
- cg.flush()
871
- led = refindex.Ledger(ROOT)
872
- summ = led.summary()
873
- check("_refindex.json 落盘", summ["exists"], summ["path"])
874
- check("水位 files == 2", summ["files"] == 2, str(summ["files"]))
875
- check("水位 nodes == 本轮 indexed",
876
- summ["nodes"] == out1.get("indexed"), f"{summ['nodes']} == {out1.get('indexed')}")
877
- n_before = len((cg.index.get("nodes") or {}))
878
- ids1 = set(out1.get("ids") or [])
879
-
880
- # ============================================ ③ 增量:未变则跳过
881
- print("\n【3】增量索引:未变文件不重切(修 F)")
882
- out2 = index_read(cg, src, incremental=True)
883
- check("incremental 跳过未变文件",
884
- out2.get("skipped_unchanged") == 2, str(out2.get("skipped_unchanged")))
885
- check("incremental 下 indexed == 0(没有重复条目)",
886
- out2.get("indexed") == 0, str(out2.get("indexed")))
887
- cg.flush()
888
- check("增量后节点数不变(幂等)",
889
- len((cg.index.get("nodes") or {})) == n_before,
890
- f"{len((cg.index.get('nodes') or {}))} == {n_before}")
891
- out3 = index_read(cg, src)
892
- check("默认(不传 incremental)仍是全量:不跳过",
893
- out3.get("skipped_unchanged") == 0, str(out3.get("skipped_unchanged")))
894
- check("重跑 id 集合稳定", set(out3.get("ids") or []) == ids1,
895
- f"{len(set(out3.get('ids') or []))} vs {len(ids1)}")
896
-
897
- # ============================================ ④ 回读 + 检索带 ref
898
- print("\n【4】op=ref 回读 + 检索结果带 ref / ref_kind")
899
- rr = ref_read(cg, node_id=(out1.get("ids") or [""])[0])
900
- check("op=ref read ok", rr.get("ok") is True, str(rr)[:120])
901
- check("ref_kind == code_ref", rr.get("ref_kind") == "code_ref",
902
- str(rr.get("ref_kind")))
903
- check("回读带 hash_match(源未动 → True)",
904
- rr.get("hash_match") is True, str(rr.get("hash_match")))
905
- read = call_tool(cg, "cg", {"op": "read", "query": "compute_energy", "k": 10})
906
- hits = [r for r in read.get("results", []) if r.get("ref_kind") == "code_ref"]
907
- check("检索结果带 ref / ref_kind 字段", bool(hits), f"hits={len(hits)}")
908
- check("检索的 ref 指回源文件",
909
- bool(hits) and isinstance(hits[0].get("ref"), dict)
910
- and hits[0]["ref"].get("path") == "alpha.py",
911
- str(hits[0].get("ref") if hits else None))
912
-
913
- # ============================================ ⑤ 漂移:stale
914
- print("\n【5】漂移检测:改源一行 → stale(只读、不抛)")
915
- wr(alpha, ALPHA_DRIFT, bump=5.0)
916
- mt_before = {p: os.path.getmtime(p) for p in (alpha, beta)}
917
- chk = ref_read(cg, action="check")
918
- mt_after = {p: os.path.getmtime(p) for p in (alpha, beta)}
919
- check("action=check 返回 ok", isinstance(chk, dict), str(type(chk)))
920
- check("check 判定 ok=False(有漂移)", chk.get("ok") is False, str(chk.get("ok")))
921
- check("报出 stale 且指向 alpha.py",
922
- any((r.get("path") or "").endswith("alpha.py") for r in chk.get("stale", [])),
923
- str([r.get("path") for r in chk.get("stale", [])]))
924
- check("巡检只读:源文件 mtime 不变", mt_before == mt_after)
925
- stale_rr = ref_read(cg, ref={"path": "alpha.py", "lineno": 5, "end": 7,
926
- "hash": "deadbeef", "root": src, "precise": True})
927
- check("回读陈旧 ref → hash_match=False / stale=True",
928
- stale_rr.get("hash_match") is False and stale_rr.get("stale") is True,
929
- str(stale_rr.get("hash_match")))
930
- index_read(cg, src) # 重建:漂移消除
931
- chk2 = ref_read(cg, action="check")
932
- check("重跑 index_code 后 stale 消除", chk2.get("ok") is True,
933
- str([r.get("path") for r in chk2.get("stale", [])]))
934
-
935
- # ============================================ 6 判据分工(#4)
936
- print("\n【6】判据分工:漂移只由节点 ref 判,水位只判「能不能跳过」")
937
- guard_verdict_owner(tmp)
938
-
939
- # ============================================ 7 撞 id 现场(D1 现状钉子)
940
- print("\n【7】撞 id 现场:归因可见 + 不重复报红(现状钉子 / 迁移触发器)")
941
- guard_collision(tmp)
942
-
943
- # ============================================ 8 悬空:dangling
944
- print("\n【8】悬空检测:删源文件 → dangling")
945
- os.remove(beta)
946
- chk3 = ref_read(cg, action="check")
947
- check("报出 dangling 且指向 beta.py",
948
- any((r.get("path") or "").endswith("beta.py")
949
- for r in chk3.get("dangling", [])),
950
- str([r.get("path") for r in chk3.get("dangling", [])]))
951
- check("dangling 时 ok=False", chk3.get("ok") is False)
952
- wr(beta, BETA) # 恢复源 + 重建
953
- index_read(cg, src)
954
- chk4 = ref_read(cg, action="check")
955
- check("恢复源并重建后 dangling 消除",
956
- chk4.get("ok") is True and not chk4.get("dangling"),
957
- str([r.get("path") for r in chk4.get("dangling", [])]))
958
-
959
- # ============================================ 9 截断不静默 + 水位留痕
960
- print("\n【9】截断显式上报 + 写进水位")
961
- tr = index_read(cg, src, max_files=1)
962
- check("max_files=1 → truncated=True", tr.get("truncated") is True)
963
- check("truncated_reason 指出 max_files",
964
- "max_files" in (tr.get("truncated_reason") or ""),
965
- str(tr.get("truncated_reason")))
966
- stt = ref_read(cg, action="stat")
967
- check("action=stat 返回水位与 last_index",
968
- isinstance(stt.get("ledger"), dict)
969
- and "last_index" in stt["ledger"], str(stt.get("ledger"))[:120])
970
- check("水位记下 last_index.truncated=True",
971
- (stt["ledger"].get("last_index") or {}).get("truncated") is True)
972
- index_read(cg, src)
973
- check("重跑全量后 last_index.truncated 回 False",
974
- (ref_read(cg, action="stat")["ledger"].get("last_index") or {})
975
- .get("truncated") is False)
976
-
977
- # ============================================ 10 接入 sustain.diagnose / heal
978
- print("\n【10】接入 sustain:diagnose 报 ref_stale,heal 的 rebuild_refs 修复")
979
- wr(alpha, ALPHA_DRIFT2, bump=5.0)
980
- diag = call_tool(cg, "cg", {"op": "sustain", "action": "diagnose"})
981
- codes = {i["code"] for i in diag.get("issues", [])}
982
- check("diagnose issues 含 ref_stale", "ref_stale" in codes, str(sorted(codes)))
983
- check("diagnose stats 报 ref_stale / ref_checked",
984
- diag.get("stats", {}).get("ref_stale", 0) >= 1
985
- and "ref_checked" in diag.get("stats", {}),
986
- str(diag.get("stats", {}).get("ref_stale")))
987
- check("巡检不改源文件(改动仍在)", ALPHA_MARK2 in rd(alpha))
988
- heal = call_tool(cg, "cg", {"op": "sustain", "action": "heal"})
989
- check("heal 动作含 rebuild_refs",
990
- any(a.get("code") == "rebuild_refs" for a in heal.get("actions", [])),
991
- str([a.get("code") for a in heal.get("actions", [])]))
992
- diag2 = call_tool(cg, "cg", {"op": "sustain", "action": "diagnose"})
993
- codes2 = {i["code"] for i in diag2.get("issues", [])}
994
- check("heal 后 ref_stale 消除", "ref_stale" not in codes2, str(sorted(codes2)))
995
- check("heal 不修改源文件", ALPHA_MARK2 in rd(alpha))
996
-
997
- # ============================================ 11 派生物不膨胀 + 未知 action
998
- print("\n【11】_refindex.json 不膨胀 / 未知 action 不静默")
999
- led2 = refindex.Ledger(ROOT)
1000
- p = led2.path
1001
- size1 = os.path.getsize(p)
1002
- for _ in range(3):
1003
- index_read(cg, src)
1004
- size2 = os.path.getsize(p)
1005
- check("_refindex.json 体积有界(< 20KB)", size2 < 20000, f"{size2}B")
1006
- check("重跑 3 次不膨胀", size2 == size1, f"{size1} -> {size2}")
1007
- check("水位 files 仍 == 2(无幽灵条目)", led2.summary()["files"] == 2,
1008
- str(led2.summary()["files"]))
1009
- try:
1010
- r = ref_read(cg, action="nope")
1011
- silent = (isinstance(r, dict) and r.get("ok") is True)
1012
- except Exception: # noqa: BLE001
1013
- silent = False
1014
- check("未知 action 不静默成功", silent is False)
1015
-
1016
- # ============================================ 12 自愈止血 + 水位保全
1017
- print("\n【12】自愈止血与水位保全(水位不被写空 / 失败不记成成功 / 同因不重试)")
1018
- guard_heal_bounded(tmp)
1019
-
1020
- # ============================================ 13 口径钉子(#5)
1021
- print("\n【13】口径钉子:id 位置化 + 幂等判据由区间哈希承担(文档↔实现同源)")
1022
- guard_id_and_idempotence(tmp)
1023
-
1024
- # ============================================ 14 常驻循环止血
1025
- print("\n【14】常驻循环止血(repeat_guard / 不静默 / 不假装修好)")
1026
- guard_loop_tick_heal(tmp)
1027
-
1028
- finally:
1029
- shutil.rmtree(tmp, ignore_errors=True)
1030
-
1031
- print("\n" + "=" * 68)
1032
- print(f"通过 {PASS} / 失败 {FAIL}")
1033
- if FAILS:
1034
- print("失败项:\n - " + "\n - ".join(FAILS))
1035
- print("=" * 68)
1036
- if args.head_baseline:
1037
- # 短码集合比对:变异后**恰好**命中预期红项(多一项少一项都报红)。无 cid 的
1038
- # 失败在 check 里记了哨兵 "?<name>",故「波及到没登记的断言」也在此报红。
1039
- got, exp = sorted(set(RED)), sorted(EXPECTED_RED)
1040
- if got == exp:
1041
- print("红基线符合预期:变异后恰好命中 %d 项 —— %s"
1042
- % (len(exp), "、".join(exp)))
1043
- return 0
1044
- print("红基线与预期不符(fail-closed)——预期红项:%s;实得红项:%s"
1045
- % ("、".join(exp) or "(无)", "、".join(got) or "(无)"))
1046
- return 1
1047
- return 1 if FAIL else 0
1048
-
1049
-
1050
- if __name__ == "__main__":
1051
- if hasattr(sys.stdout, "reconfigure"):
1052
- sys.stdout.reconfigure(encoding="utf-8", errors="replace")
1053
- sys.exit(main())
1
+ # -*- coding: utf-8 -*-
2
+ """ref 增量索引 + 漂移/悬空巡检(P28 · R3 改造验收)。
3
+
4
+ 对照 `docs/mdcg/认知图_索引与工程规范化_计划_v0.1.md` 的 R3:
5
+
6
+ ① 统一 ref 协议 + 提取器注册表:调用方只说 `kind`,`refindex` 决定用哪个
7
+ 提取器;无提取器后缀显式报错。
8
+ ② `_refindex.json` 增量(修 F):`incremental=True` 时未变文件**不再读盘重切**
9
+ (`skipped_unchanged`),节点仍在(幂等);默认全量,保证「不漏召回」。
10
+ ③ `op=ref action=check` 漂移/悬空巡检(修 D):改源一行 → 报 `stale`;删除源
11
+ 文件 → 报 `dangling`;**只读、不抛、不改源文件**。
12
+ ④ `op=ref action=stat`:看水位,最近一次索引是否被截断(`truncated` 不再静默)。
13
+ ⑤ 接入 `sustain.diagnose`:`ref_stale` / `ref_dangling` 进入体检;`heal` 的
14
+ `rebuild_refs` 重建派生物后漂移消除(源文件不动)。
15
+ ⑥ 检索结果带 `ref` / `ref_kind`(读侧只加字段,不改召回逻辑)。
16
+ ⑦ 派生物不膨胀:`_refindex.json` 体积有界、重跑不涨。
17
+ ⑧ **判据分工**(#4):漂移 / 悬空只由**节点自身 frontmatter ref** 判,水位只判
18
+ 「这个节点能不能跳过、不读源」(五项全等 + 源 size/mtime 未变 + 同 id 只判一次);
19
+ 水位与节点不一致记 `ledger_mismatch` 归因行(带两侧 root),**不进 ok、不重复报红**。
20
+ ⑨ **撞 id 现场**(D1 现状钉子):两域同 rel path + 同 name ⇒ 同 id、节点被后写域覆盖,
21
+ 但归因行可见。本组断言**现状事实**(+ 去重 + 归因字段),不作期望语义断言——
22
+ `node_id` 一旦改为含 root,① 会转红以提示同步迁移。
23
+ ⑩ **口径钉子**(#5):id 是**位置化**寻址(`sha1(路径::符号名)` / `sha1(路径#标题路径)`),
24
+ 幂等判据是**区间哈希 + 行位全等**(不是「id 存在即已索引」)。两侧都钉——实现侧
25
+ 手算 sha1 独立复算 + 对 root / 正文 / 行位不敏感 + 五项逐项不跳过;文档侧
26
+ `docs/plans/真源索引_通用机制_v0.3.md` §3.1 的口径字面、整块指纹(**陈化硬失败**)
27
+ 与行级引用必须仍与实现同源(引用漂移即红,失败信息附现址)。
28
+ ⑪ **常驻循环止血**:`SustainLoop._tick_heal` 按 `repeat_guard=True` 调 heal——同因重复
29
+ 失败被拦下(`repeat_failure` + 退避)而**诊断面照旧报病灶**(判据不被削弱)、
30
+ `heals[*].failed` 单列(没修好不静默、不假装成功)。
31
+ 红基线(判别力自证,见 `_MUTATIONS` / `EXPECTED_RED`):`--head-baseline` 对
32
+ **工作区 `md_cg/refindex.py`** 做定点变异(快路径探测数据改回水位条目=v1 旧行为),
33
+ 断言红项集合**恰好**等于预期三项(⑩ / ⑪ 两段钉的是别的语义,变异态下**仍绿**,
34
+ 故不入红项表——「恰好」二字正是这么被守住的)。
35
+
36
+ 运行:python -m md_cg.test_p28_refcheck [--head-baseline]
37
+ """
38
+ from __future__ import annotations
39
+
40
+ import argparse
41
+ import hashlib
42
+ import os
43
+ import re
44
+ import shutil
45
+ import sys
46
+ import tempfile
47
+ import time
48
+
49
+ from . import corpus, refindex, sources, sustain, tokens
50
+ from .mdcg import MdCG
51
+ from .mcp_server import call_tool
52
+
53
+ PASS = FAIL = 0
54
+ FAILS = []
55
+ RED = [] # 转红项 cid(红基线模式比对用;绿态为空)
56
+
57
+ _BASE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
58
+ ROOT = os.path.join(_BASE, "_md_cg_p28")
59
+
60
+
61
+ def check(name, cond, detail="", cid=None):
62
+ global PASS, FAIL
63
+ if cond:
64
+ PASS += 1
65
+ print(f" [PASS] {name}" + (f" · {detail}" if detail else ""))
66
+ else:
67
+ FAIL += 1
68
+ FAILS.append(name)
69
+ # 无 cid 的失败记哨兵:红基线模式下它会与预期红项不等 ⇒ fail-closed,
70
+ # 不静默放过「变异波及了没被登记的断言」(同 test_logref.py:66 的口径)。
71
+ RED.append(cid or ("?" + name))
72
+ print(f" [FAIL] {name} · {detail}")
73
+
74
+
75
+ ALPHA = ('"""能量计算模块。"""\n'
76
+ 'import math\n'
77
+ '\n'
78
+ '\n'
79
+ 'def compute_energy(mass, speed):\n'
80
+ ' """计算动能。"""\n'
81
+ ' return mass * speed * speed\n'
82
+ '\n'
83
+ '\n'
84
+ 'class Reactor:\n'
85
+ ' """反应堆模型。"""\n'
86
+ '\n'
87
+ ' def ignite(self):\n'
88
+ ' """点火。"""\n'
89
+ ' self.state = "on"\n')
90
+
91
+ # 同函数同区间,只改一行实现 → 行数不变,只有区间哈希变(最强的漂移信号)
92
+ ALPHA_DRIFT = ALPHA.replace("return mass * speed * speed",
93
+ "return 0.5 * mass * speed * speed")
94
+ # 第二个漂移变体:段⑤ 已把源改成 ALPHA_DRIFT 并重建索引,段⑨ 要再制造一次真实漂移,
95
+ # 必须写出**与当前索引内容不同**的正文;否则只是 mtime 变了、内容没变,报 ok 才对。
96
+ ALPHA_DRIFT2 = ALPHA.replace("return mass * speed * speed",
97
+ "return mass * speed * speed / 2.0")
98
+ ALPHA_MARK2 = "speed * speed / 2.0"
99
+
100
+ BETA = ('"""辅助模块。"""\n'
101
+ '\n'
102
+ '\n'
103
+ 'def helper_scale(x):\n'
104
+ ' """缩放。"""\n'
105
+ ' return x * 2\n')
106
+
107
+
108
+ def rd(path):
109
+ with open(path, encoding="utf-8") as f:
110
+ return f.read()
111
+
112
+
113
+ def wr(path, text, bump=0.0):
114
+ with open(path, "w", encoding="utf-8") as f:
115
+ f.write(text)
116
+ if bump:
117
+ # 显式推后 mtime:粗粒度文件系统上保证「变了」能被水位看见
118
+ t = os.path.getmtime(path) + bump
119
+ os.utime(path, (t, t))
120
+
121
+
122
+ def index_read(cg, code_dir, **extra):
123
+ a = {"op": "index_code", "path": code_dir}
124
+ a.update(extra)
125
+ return call_tool(cg, "cg", a)
126
+
127
+
128
+ def ref_read(cg, **args):
129
+ a = {"op": "ref"}
130
+ a.update(args)
131
+ return call_tool(cg, "cg", a)
132
+
133
+
134
+ # ---------------- 6 判据分工:节点 ref 判漂移,水位只判跳过 ----------------
135
+
136
+ #: 合成源:改一行实现 → 行数不变、只有区间哈希变(最强的漂移信号,同 ALPHA_DRIFT)。
137
+ PROBE_SRC = ('"""probe 模块。"""\n'
138
+ '\n'
139
+ '\n'
140
+ 'def probe_fn(x):\n'
141
+ ' """probe 函数。"""\n'
142
+ ' return x + 1\n')
143
+ PROBE_SRC2 = PROBE_SRC.replace("return x + 1", "return x + 2")
144
+
145
+
146
+ def _mk(tmp, name):
147
+ d = os.path.join(tmp, name)
148
+ os.makedirs(d)
149
+ return d
150
+
151
+
152
+ def _index_into(cg, src, ledger=None):
153
+ """直调 refindex 的两步(index_dir → add_items),返回本轮节点 id。
154
+
155
+ 必须直调:`op=index_code` 是这两步的**连写**,表达不了「只写一侧」的分歧。
156
+ """
157
+ items, _errors, _stats = refindex.index_dir(src, kind="code_ref",
158
+ ledger=ledger)
159
+ ids, _sens = refindex.add_items(cg, items, kind="code_ref", root=src)
160
+ return ids
161
+
162
+
163
+ def _tamper_ref_hash(node_path, fake="deadbeefcafe"):
164
+ """把节点 .md 里 frontmatter 的 ref 哈希改成假值(源与水位都不动)。
165
+
166
+ `cg.get` 每次从盘上读节点,故改 .md 即改「节点自述」;水位条目不受影响——
167
+ 这正是「漂移只由节点 ref 判」要区分的场景(v1 拿水位当探测数据 ⇒ 静默)。
168
+ """
169
+ with open(node_path, encoding="utf-8") as f:
170
+ txt = f.read()
171
+ out = re.sub(r'"hash": "[0-9a-f]{12}"', '"hash": "%s"' % fake, txt, count=1)
172
+ if out == txt:
173
+ raise RuntimeError("夹具失效:节点 frontmatter 里没有可篡改的 ref.hash")
174
+ with open(node_path, "w", encoding="utf-8") as f:
175
+ f.write(out)
176
+
177
+
178
+ def _verdict(res):
179
+ """巡检结果 → 去重/归因视角的摘要(红基线断言只看这里)。"""
180
+ ids = [r.get("node_id") for r in list(res.get("stale") or [])
181
+ + list(res.get("dangling") or [])]
182
+ return {"ok": res["ok"], "ids": ids, "dup": len(ids) - len(set(ids)),
183
+ "mm": res.get("ledger_mismatch") or [],
184
+ "mm_count": res.get("ledger_mismatch_count", 0)}
185
+
186
+
187
+ def guard_verdict_owner(tmp):
188
+ """三组构造:水位与节点分别推到不同代,看快路径认谁(直调 refindex)。"""
189
+ # ①水位新、节点旧 —— 等价 mcp_server.py 的 index_dir→add_items 之间被杀
190
+ lib = _mk(tmp, "fo_lib")
191
+ src = _mk(tmp, "fo_src")
192
+ p = os.path.join(src, "probe_widget.py")
193
+ wr(p, PROBE_SRC)
194
+ cg = MdCG(lib)
195
+ ids = _index_into(cg, src, refindex.Ledger(lib))
196
+ cg.flush()
197
+ wr(p, PROBE_SRC2, bump=5.0)
198
+ refindex.index_dir(src, kind="code_ref", ledger=refindex.Ledger(lib))
199
+ res = refindex.check_refs(cg, ledger=refindex.Ledger(lib))
200
+ v = _verdict(res)
201
+ node_hashes = {refindex.ref_of(cg.get(n))[1]["hash"] for n in ids}
202
+ check("6①只推水位:ok=False(漂移由节点 ref 判,水位说了不算)",
203
+ res["ok"] is False, f"ok={res['ok']} stale={len(res['stale'])}", cid="A1")
204
+ check("6①stale 覆盖该文件全部节点,hash_expected 取节点 ref(不是水位新哈希)",
205
+ {r["node_id"] for r in res["stale"]} == set(ids)
206
+ and {r["hash_expected"] for r in res["stale"]} == node_hashes,
207
+ f"stale={sorted(r['node_id'] for r in res['stale'])}", cid="A1")
208
+ check("6①水位/节点不一致记归因行 watermark_behind(写序断裂可见,不进 ok)",
209
+ v["mm_count"] >= 1
210
+ and all(r["kind_of_gap"] == "watermark_behind" for r in v["mm"]),
211
+ str(v["mm"][:1]), cid="A2")
212
+
213
+ # ②节点新、水位旧 —— ledger 根本不参与,才算「只推节点」
214
+ lib = _mk(tmp, "fn_lib")
215
+ src = _mk(tmp, "fn_src")
216
+ p = os.path.join(src, "probe_widget.py")
217
+ wr(p, PROBE_SRC)
218
+ cg = MdCG(lib)
219
+ _index_into(cg, src, refindex.Ledger(lib))
220
+ cg.flush()
221
+ wr(p, PROBE_SRC2, bump=5.0)
222
+ _index_into(cg, src, ledger=None) # 全量重切,水位不动
223
+ cg.flush()
224
+ res = refindex.check_refs(cg, ledger=refindex.Ledger(lib))
225
+ v = _verdict(res)
226
+ check("6②只推节点:ok=True 且 stale/dangling 空(旧水位不得误报漂移)",
227
+ res["ok"] is True and not res["stale"] and not res["dangling"],
228
+ f"ok={res['ok']} stale={sorted(v['ids'])}", cid="B1")
229
+ # 归因面与「红不红」解耦:这里只断言「不一致被记下来了」(ok 由 B1 断),
230
+ # 免得同一次误报把两条断言一起打红、红项表失去「一项一行为」的对应。
231
+ check("6②旧水位与节点不一致仍归因(watermark_behind)",
232
+ v["mm_count"] >= 1
233
+ and all(r["kind_of_gap"] == "watermark_behind" for r in v["mm"]),
234
+ str(v["mm"][:1]), cid="B2")
235
+
236
+ # ③篡改节点 ref.hash(源与水位都不动)—— 只有节点自述变了
237
+ lib = _mk(tmp, "ft_lib")
238
+ src = _mk(tmp, "ft_src")
239
+ p = os.path.join(src, "probe_widget.py")
240
+ wr(p, PROBE_SRC)
241
+ cg = MdCG(lib)
242
+ ids = _index_into(cg, src, refindex.Ledger(lib))
243
+ cg.flush()
244
+ nid = sorted(ids)[0]
245
+ _tamper_ref_hash(os.path.join(lib, cg.index["nodes"][nid]["path"]))
246
+ res = refindex.check_refs(cg, ledger=refindex.Ledger(lib))
247
+ v = _verdict(res)
248
+ check("6③篡改节点 ref.hash(源与水位都不动):必须报 stale",
249
+ res["ok"] is False and any(r["node_id"] == nid for r in res["stale"]),
250
+ f"ok={res['ok']} stale={v['ids']}", cid="C1")
251
+ check("6③同一 node_id 至多一行(去重,不重复报红)", v["dup"] == 0,
252
+ str(v["ids"]), cid="C2")
253
+
254
+
255
+ def guard_collision(tmp):
256
+ """撞 id 现场(D1 状态):钉现状事实 + 去重 + 归因,不作期望语义断言。
257
+
258
+ `node_id` 不含 root ⇒ 两域同 rel path + 同 name 时撞 id、节点被**后写域**覆盖
259
+ (`add_items` 同 id upsert 整条替换)。本组断言这个**现状**(以及归因行可见、
260
+ 不重复报红);键一旦改为含 root,第一条会转红以提示同步迁移——那是迁移触发器,
261
+ 不是要实现的期望语义(见 docs/plans/真源索引_通用机制_v0.3.md「id 语义与设计债」)。
262
+ """
263
+ lib = _mk(tmp, "fc_lib")
264
+ domA = _mk(tmp, "fc_domA")
265
+ domB = _mk(tmp, "fc_domB")
266
+ for d in (domA, domB):
267
+ wr(os.path.join(d, "probe_widget.py"), PROBE_SRC)
268
+ cg = MdCG(lib)
269
+ ids_a = _index_into(cg, domA, refindex.Ledger(lib))
270
+ # #352-E(2026-10-15):id 键**仍未**含 root,但写入面加了冲突闸——B 的同名条目
271
+ # 不再覆写 A,改为「不写入 + 回执 conflicts 点名」。故本钉从「两域 id 集相等」
272
+ # (=后写域覆盖)改为「两域算出的 id 仍相同,但 B 一个都不写」。原钉的用途
273
+ # (键一旦改为含 root 即转红以提示同步迁移)保留:改 id 键后这里仍会转红。
274
+ _items_b, _eb, _sb = refindex.index_dir(domB, kind="code_ref",
275
+ ledger=refindex.Ledger(lib))
276
+ _cf = []
277
+ ids_b, _ = refindex.add_items(cg, _items_b, kind="code_ref", root=domB,
278
+ conflicts=_cf)
279
+ cg.flush()
280
+ nid = sorted(ids_a)[0]
281
+ ref = refindex.ref_of(cg.get(nid))[1] or {}
282
+ check("7①两域同 rel path + 同 name ⇒ 仍算出同一 id,但后写域被冲突闸拦下不写入",
283
+ bool(ids_a) and ids_b == []
284
+ and {c.get("id") for c in _cf} == set(ids_a),
285
+ f"A={ids_a} B={ids_b} conflicts={_cf}", cid="D1")
286
+ check("7①同一 id 只留一条节点,ref.root 仍属**先写域**(#352-E 后的钉子)",
287
+ len(cg.index["nodes"]) == len(ids_a)
288
+ and refindex._same_root(ref.get("root"), domA),
289
+ f"nodes={len(cg.index['nodes'])} root={ref.get('root')}", cid="D1")
290
+ _tamper_ref_hash(os.path.join(lib, cg.index["nodes"][nid]["path"]))
291
+ res = refindex.check_refs(cg, ledger=refindex.Ledger(lib))
292
+ v = _verdict(res)
293
+ check("7②该现场篡改节点 ref.hash:同一 node_id 至多一行(dup=0)",
294
+ v["dup"] == 0, str(v["ids"]), cid="D2")
295
+ check("7②水位/节点不一致记 ledger_mismatch 归因行(撞 id 不再静默)",
296
+ v["mm_count"] >= 1, str(v["mm_count"]), cid="D2")
297
+ # #352-E:撞 id 的归因从**巡检面**(watermark/node 两侧 root 不一致)前移到
298
+ # **写入面回执**(conflicts 带 existing_root/root)——写入面不再让撞 id 落盘,
299
+ # 巡检面自然不再有「节点归了后写域」这个事实可归因。归因字段本身(两侧 root)
300
+ # 一个没少,只是换了承担者。
301
+ check("7②撞 id 归因改由写入面回执承担(conflicts 带两侧 root,可归因到域)",
302
+ bool(_cf) and all(c.get("existing_root") and c.get("root")
303
+ and c.get("id") for c in _cf),
304
+ str(_cf[:1]), cid="D2")
305
+ # 该现场再推一步:**两域的源都改掉**(合法漂移,不是篡改)。v1 逐条水位条目各自
306
+ # 探测 ⇒ 同一个 node_id 被反复报(本夹具实测 HEAD 4 行 / dup=2 ⇒ v2 2 行 / dup=0,
307
+ # 两个数是该文件的两个节点各一行)。这一步才真正钉住「去重」,且不依赖篡改。
308
+ wr(os.path.join(domA, "probe_widget.py"), PROBE_SRC2, bump=5.0)
309
+ wr(os.path.join(domB, "probe_widget.py"), PROBE_SRC2, bump=5.0)
310
+ v = _verdict(refindex.check_refs(cg, ledger=refindex.Ledger(lib)))
311
+ check("7③两域源都改:同一 node_id 至多一行(v1 会按水位条目重复报)",
312
+ v["dup"] == 0 and bool(v["ids"]), str(v["ids"]), cid="D3")
313
+ _items_b2, _eb2, _sb2 = refindex.index_dir(domB, kind="code_ref",
314
+ ledger=refindex.Ledger(lib))
315
+ _cf2 = []
316
+ _ids_b2, _ = refindex.add_items(cg, _items_b2, kind="code_ref",
317
+ root=domB, conflicts=_cf2)
318
+ check("7③两域源都改后再索引 B:撞 id 仍由回执点名(带两侧 root,可归因到域)",
319
+ _ids_b2 == [] and bool(_cf2)
320
+ and all(c.get("existing_root") and c.get("root") for c in _cf2),
321
+ f"ids={_ids_b2} conflicts={_cf2[:1]}", cid="D3")
322
+
323
+
324
+ # ---------------- 12 自愈止血 + 水位保全(H1/H1b/H2/H3/H4/H5/H6 + 写序) ----------------
325
+
326
+ def guard_heal_bounded(tmp):
327
+ """水位不被写空 / 重建失败不许记成成功 / 同因不重试(直调 refindex|sustain)。
328
+
329
+ 全部用对称小夹具(每域 1 个源文件、2 个节点),断言的是**行为**而不是文案。
330
+ """
331
+ # ①水位保全(H1):目录仍在、零文件 ⇒ 空扫不对账,水位不被剪空
332
+ lib = _mk(tmp, "hb1_lib")
333
+ src = _mk(tmp, "hb1_src")
334
+ p = os.path.join(src, "hb1_widget.py")
335
+ wr(p, PROBE_SRC)
336
+ cg = MdCG(lib)
337
+ ids = _index_into(cg, src, refindex.Ledger(lib))
338
+ cg.flush()
339
+ os.remove(p) # 目录仍在、零文件
340
+ refindex.index_dir(src, kind="code_ref", ledger=refindex.Ledger(lib))
341
+ s = refindex.Ledger(lib).summary()
342
+ check("12①空扫不对账:水位条目与节点数保住([C1] 目录仍在零文件)",
343
+ s["files"] == 1 and s["nodes"] == len(ids),
344
+ f"files={s['files']} nodes={s['nodes']} ids={len(ids)}", cid="E1")
345
+ check("12①两面不打架:last_index 如实记 files=0 且 empty_scan=True",
346
+ (s.get("last_index") or {}).get("files") == 0
347
+ and (s.get("last_index") or {}).get("empty_scan") is True,
348
+ str(s.get("last_index"))[:120], cid="E1")
349
+
350
+ # ②悬空 root 不重建(H1b):不调 index_dir ⇒ 水位不被写成空
351
+ lib = _mk(tmp, "hb2_lib")
352
+ src = _mk(tmp, "hb2_src")
353
+ wr(os.path.join(src, "hb2_widget.py"), PROBE_SRC)
354
+ cg = MdCG(lib)
355
+ _index_into(cg, src, refindex.Ledger(lib))
356
+ cg.flush()
357
+ n_files = refindex.Ledger(lib).summary()["files"]
358
+ shutil.rmtree(src) # 源大域整个没了
359
+ r = refindex.rebuild(cg, ledger=refindex.Ledger(lib))
360
+ check("12②rebuild 遇悬空 root:不重建、ok=False、roots_missing 如实记",
361
+ r["ok"] is False and len(r["roots_missing"]) == 1 and r["indexed"] == 0
362
+ and any("已不是目录" in e for e in r["errors"]),
363
+ f"ok={r['ok']} missing={r['roots_missing']} err={r['errors'][:1]}", cid="E2")
364
+ check("12②悬空 root 不再把水位写成空([C3] files 保住 vs 现状 0)",
365
+ refindex.Ledger(lib).summary()["files"] == n_files,
366
+ f"files={refindex.Ledger(lib).summary()['files']} 期望 {n_files}", cid="E2")
367
+
368
+ # ③空扫组如实置 ok=False(H2)
369
+ lib = _mk(tmp, "hb3_lib")
370
+ src = _mk(tmp, "hb3_src")
371
+ p = os.path.join(src, "hb3_widget.py")
372
+ wr(p, PROBE_SRC)
373
+ cg = MdCG(lib)
374
+ _index_into(cg, src, refindex.Ledger(lib))
375
+ cg.flush()
376
+ os.remove(p)
377
+ r = refindex.rebuild(cg, ledger=refindex.Ledger(lib))
378
+ check("12③rebuild 空扫组:ok=False + errors + 逐组 empty_scan 如实回报",
379
+ r["ok"] is False and r["errors"] and r["rebuilt"]
380
+ and r["rebuilt"][0]["empty_scan"] is True
381
+ and r["rebuilt"][0]["files"] == 0,
382
+ str(r["rebuilt"])[:140], cid="E3")
383
+
384
+ # ④只有悬空 ⇒ heal 不重建(H3):止血点在「按 code 分派」
385
+ lib = _mk(tmp, "hb4_lib")
386
+ domA = _mk(tmp, "hb4_domA")
387
+ domB = _mk(tmp, "hb4_domB")
388
+ for d, nm in ((domA, "hb4_a.py"), (domB, "hb4_b.py")):
389
+ wr(os.path.join(d, nm), PROBE_SRC)
390
+ cg = MdCG(lib)
391
+ _index_into(cg, domA, refindex.Ledger(lib))
392
+ _index_into(cg, domB, refindex.Ledger(lib))
393
+ cg.flush()
394
+ shutil.rmtree(domB)
395
+ codes = {i["code"] for i in sustain.diagnose(cg)["issues"]}
396
+ acts = [a["code"] for a in sustain.heal(cg)["actions"]]
397
+ check("12④ref_dangling 不再触发重建(fix 字段无调度器消费,改它无效)",
398
+ "ref_dangling" in codes and "rebuild_refs" not in acts,
399
+ f"codes={sorted(codes)} actions={acts}", cid="E4")
400
+
401
+ # ⑤重建失败如实记 ok=False(H4)+ ⑥同因不重试(H6)
402
+ lib = _mk(tmp, "hb5_lib")
403
+ domA = _mk(tmp, "hb5_domA")
404
+ domB = _mk(tmp, "hb5_domB")
405
+ for d, nm in ((domA, "hb5_a.py"), (domB, "hb5_b.py")):
406
+ wr(os.path.join(d, nm), PROBE_SRC)
407
+ cg = MdCG(lib)
408
+ _index_into(cg, domA, refindex.Ledger(lib))
409
+ _index_into(cg, domB, refindex.Ledger(lib))
410
+ cg.flush()
411
+ key = (os.path.abspath(lib), "rebuild_refs")
412
+ sustain._HEAL_MEMO.pop(key, None)
413
+ p = os.path.join(domA, "hb5_a.py")
414
+ wr(p, PROBE_SRC2, bump=5.0) # 一个域漂移
415
+ shutil.rmtree(domB) # 另一个域悬空(重建必带 roots_missing)
416
+ h1 = sustain.heal(cg, repeat_guard=True, heal_interval=300.0)
417
+ row = next((a for a in h1["actions"] if a["code"] == "rebuild_refs"), None)
418
+ check("12⑤重建失败不得记成成功:applied=True 且 ok=False,并带结果摘要",
419
+ bool(row) and row["applied"] is True and row["ok"] is False
420
+ and bool((row.get("result") or {}).get("roots_missing")),
421
+ str(row)[:170], cid="E5")
422
+ wr(p, PROBE_SRC, bump=5.0) # 回到与上次相同的诊断信号
423
+ h2 = sustain.heal(cg, repeat_guard=True, heal_interval=300.0)
424
+ row2 = next((a for a in h2["actions"] if a["code"] == "rebuild_refs"), None)
425
+ check("12⑥同一信号重复失败:拦下并如实记 repeat_failure(不是静默、不重刷)",
426
+ bool(row2) and row2["applied"] is False
427
+ and row2.get("reason") == "repeat_failure"
428
+ and row2.get("streak") == 1 and row2.get("wait_s") == 300.0,
429
+ str(row2)[:170], cid="E6")
430
+ sustain._HEAL_MEMO[key]["ts"] = time.time() - 400 # 模拟退避窗已过
431
+ h3 = sustain.heal(cg, repeat_guard=True, heal_interval=300.0)
432
+ row3 = next((a for a in h3["actions"] if a["code"] == "rebuild_refs"), None)
433
+ check("12⑥退避窗过后放行重试(退避 ≠ 永久静默)",
434
+ bool(row3) and row3["applied"] is True, str(row3)[:170], cid="E6")
435
+ sustain._HEAL_MEMO.clear()
436
+
437
+ # ⑦sources 的水位面(H5):Ledger 没有 stat(),写错被 except 吞成 {}
438
+ st = sources.FileDispatcher(cg)._ledger_stat()
439
+ check("12⑦ingest stat 的水位面不再被 AttributeError 吞成空",
440
+ isinstance(st, dict) and "files" in st and "path" in st,
441
+ str(sorted(st))[:120], cid="E7")
442
+
443
+ # ⑧写序(#2):节点写入失败 ⇒ 水位不得先落盘
444
+ lib = _mk(tmp, "hb6_lib")
445
+ src = _mk(tmp, "hb6_src")
446
+ wr(os.path.join(src, "hb6_widget.py"), PROBE_SRC)
447
+ cg = MdCG(lib)
448
+ orig = refindex.add_items
449
+
450
+ # 生效条件:任何入参都抛 RuntimeError(把「节点写入失败」注入 op 面),用于验证水位未先落盘;
451
+ def _boom(*_a, **_k):
452
+ raise RuntimeError("夹具:节点写入失败")
453
+
454
+ refindex.add_items = _boom
455
+ raised = False
456
+ try:
457
+ index_read(cg, src)
458
+ except RuntimeError:
459
+ raised = True
460
+ finally:
461
+ refindex.add_items = orig
462
+ led = refindex.Ledger(lib)
463
+ check("12⑧节点写入失败时水位不落盘(残留方向=节点新/水位旧,不是反向)",
464
+ raised and not os.path.isfile(led.path),
465
+ f"raised={raised} ledger_exists={os.path.isfile(led.path)}", cid="E8")
466
+
467
+
468
+ # ---------------- 13 口径钉子:id 位置化 + 幂等判据由区间哈希承担(#5) ----------------
469
+
470
+ #: 口径真源(文档侧)。措辞与实现必须同源——改一边不改另一边即转红。
471
+ DOC_PLAN = os.path.join(_BASE, "docs", "plans", "真源索引_通用机制_v0.3.md")
472
+ #: 口径块定位:起=该标题行,止=下一个**二级**标题(`### ` 不算终止)。
473
+ DOC_SEC_HEAD = "### 3.1 "
474
+
475
+ #: §3.1(归一化后)的 sha1 前 16 位:**判据侧陈化硬失败**——口径一被改写(无论改真源
476
+ #: 还是改守卫)就必须同步本行,才谈得上「措辞与实现同源」。归一化=空白折叠 ⇒ 换行 /
477
+ #: 缩进 / 表格对齐**不**误红,只有措辞变化才红。失败信息里打印现算值,照抄即恢复。
478
+ DOC_SEC_FP = "f373ca6380e3debf"
479
+
480
+ #: §3.1 里必须**逐字**出现的口径字面(缺一条 ⇒ 措辞已从实现侧漂开)。
481
+ DOC_REQUIRED = (
482
+ "路径-标题寻址", # id 是位置化寻址(不是内容寻址)
483
+ "sha1(路径::符号名)", # code_ 的口径公式
484
+ "sha1(路径#标题路径)", # doc_ 的口径公式
485
+ "不含 root、不含正文", # 位置化的两个「不含」
486
+ "id 存在即已索引", # D2:明令禁止的判据(幂等不许看 id)
487
+ "hash + 行位全等", # 幂等判据的实际承担者
488
+ )
489
+
490
+ #: 行级引用 → 实现符号:(文档里必须出现的引用字面, 文件, 行号, 该行应以什么开头)。
491
+ #: 「文档说谁在哪一行」是措辞里最硬的一句——引用漂移(代码增删行)即红,失败信息附
492
+ #: 该符号**现址**,照抄回文档与本表即恢复(陈化可自愈到「人只需抄一次」)。
493
+ DOC_CITATIONS = (
494
+ ("refindex.py:577", "md_cg/refindex.py", 577, "def _code_ref("),
495
+ ("refindex.py:591", "md_cg/refindex.py", 591, "def _doc_ref("),
496
+ ("refindex.py:635", "md_cg/refindex.py", 635,
497
+ "got = codeindex.region_hash(lines, lineno, end)"),
498
+ ("refindex.py:652", "md_cg/refindex.py", 652, "def read_ref("),
499
+ ("refindex.py:720", "md_cg/refindex.py", 720, "def check_refs("),
500
+ ("refindex.py:931", "md_cg/refindex.py", 931, "def prune_orphans("),
501
+ ("srcindex.py:71", "md_cg/srcindex.py", 71, "def src_id("),
502
+ ("srcindex.py:86", "md_cg/srcindex.py", 86, "def file_hash("),
503
+ ("logref.py:119", "md_cg/logref.py", 119, "def node_state("),
504
+ ("docindex.py:294", "md_cg/docindex.py", 294, "def node_id("),
505
+ )
506
+
507
+
508
+ def _doc_block(path: str, head: str) -> str:
509
+ """取文档里 `head` 起、到下一个**二级**标题前的正文块,空白折叠后返回。
510
+
511
+ 起行不在文档里返回 ""(由调用方判红,不留静默分支)。归一化只折空白——
512
+ 重排 / 换行 / 表格对齐不影响指纹,只有**措辞**变化才改指纹。
513
+ """
514
+ try:
515
+ with open(path, encoding="utf-8") as f:
516
+ lines = f.read().replace("\r\n", "\n").split("\n")
517
+ except OSError:
518
+ return "" # 文档缺失=口径真源不可读,交由调用方判红
519
+ start = next((i for i, ln in enumerate(lines) if ln.startswith(head)), None)
520
+ if start is None:
521
+ return ""
522
+ i = start + 1
523
+ while i < len(lines) and not (lines[i].startswith("## ")
524
+ and not lines[i].startswith("### ")):
525
+ i += 1
526
+ return re.sub(r"\s+", " ", "\n".join(lines[start:i])).strip()
527
+
528
+
529
+ def guard_id_and_idempotence(tmp):
530
+ """#5 口径钉子:id 位置化(不含 root / 正文 / 行位)、幂等判据由区间哈希 + 行位承担。
531
+
532
+ 两侧都钉,谁单方面漂开都转红:①**实现侧**——id 用手算 sha1 独立复算(不借实现),
533
+ 对 root / 正文哈希 / 行位不敏感;水位与节点 ref 五项全等才是跳过的前提(「id 存在」
534
+ 不构成已索引);「同 id 而内容变」必须判 stale。②**文档侧**——`v0.3` 计划 §3.1 的
535
+ 口径字面齐全、整块指纹不陈化、行级引用仍指向所声称的符号。
536
+ """
537
+ from . import codeindex, docindex
538
+
539
+ # ---- ①id 口径:位置化,公式可手算复算 ----
540
+ it = {"path": "pkg/probe_widget.py", "name": "probe_fn", "kind": "function",
541
+ "lineno": 4, "end": 6, "hash": "aaaa1111bbbb", "lang": "py",
542
+ "precise": True}
543
+ want_code = "code_" + hashlib.sha1(
544
+ b"pkg/probe_widget.py::probe_fn").hexdigest()[:12]
545
+ check("13①code_ == sha1(路径::符号名)[:12](手算交叉核验,非复用实现)",
546
+ codeindex.node_id(it) == want_code
547
+ and refindex.node_id_of(it, "code_ref") == want_code,
548
+ f"{codeindex.node_id(it)} vs {want_code}", cid="F1")
549
+ dit = {"path": "docs/probe.md", "heading_path": ["一", "二"], "heading": "二",
550
+ "level": 2, "lineno": 10, "end": 20, "hash": "cccc2222", "dup": 1}
551
+ want_doc = "doc_" + hashlib.sha1(
552
+ "docs/probe.md#一/二".encode("utf-8")).hexdigest()[:12]
553
+ check("13①doc_ == sha1(路径#标题路径)[:12](手算交叉核验)",
554
+ docindex.node_id(dit) == want_doc
555
+ and refindex.node_id_of(dit, "doc_ref") == want_doc,
556
+ f"{docindex.node_id(dit)} vs {want_doc}", cid="F1")
557
+ check("13②id 不读 root(跨域同 rel path 必然同 id —— D1 的机械根)",
558
+ codeindex.node_id(dict(it, root="/domA")) == want_code
559
+ and codeindex.node_id(dict(it, root="/domB")) == want_code
560
+ and docindex.node_id(dict(dit, root="/domA")) == want_doc,
561
+ cid="F2")
562
+ check("13②id 不读正文哈希与行位(改内容 / 挪行位 id 不变 ⇒ 位置化而非内容寻址)",
563
+ codeindex.node_id(dict(it, hash="deadbeefdead", lineno=99, end=120)) == want_code
564
+ and docindex.node_id(dict(dit, hash="zzzz9999", lineno=1, end=2)) == want_doc,
565
+ cid="F2")
566
+
567
+ # ---- ②幂等判据:水位只记「能不能跳过」,判据键 = 节点 ref 的五项 + 源 size/mtime ----
568
+ lib = _mk(tmp, "fx_lib")
569
+ src = _mk(tmp, "fx_src")
570
+ p = os.path.join(src, "probe_widget.py")
571
+ wr(p, PROBE_SRC)
572
+ cg = MdCG(lib)
573
+ ids = _index_into(cg, src, refindex.Ledger(lib))
574
+ cg.flush()
575
+ nid = sorted(ids)[0]
576
+ _k0, ref0 = refindex.ref_of(cg.get(nid))
577
+ ents = [e for e in (refindex.Ledger(lib).load().get("files") or {}).values()
578
+ if (e.get("path") or "").endswith("probe_widget.py")]
579
+ e = ents[0] if ents else {}
580
+ n0 = (e.get("nodes") or [{}])[0]
581
+ check("13③水位键 = 源文件绝对路径且只有一条(跨域不撞名)",
582
+ len(ents) == 1, f"{len(ents)} 条", cid="F3")
583
+ check("13③水位条目记的正是判据要用的五项(root/path/lineno/end/hash)+ 源 size/mtime",
584
+ {"root", "path", "size", "mtime", "kind", "nodes"} <= set(e)
585
+ and {"id", "lineno", "end", "hash"} <= set(n0), sorted(e), cid="F3")
586
+ check("13③稳态:水位与节点 ref 五项全等(跳过的前提真的成立)",
587
+ refindex._same_root(e.get("root"), (ref0 or {}).get("root"))
588
+ and refindex._norm_rel(e.get("path")) == refindex._norm_rel((ref0 or {}).get("path"))
589
+ and n0.get("hash") == (ref0 or {}).get("hash")
590
+ and n0.get("lineno") == (ref0 or {}).get("lineno")
591
+ and n0.get("end") == (ref0 or {}).get("end"),
592
+ f"watermark={n0} ref={ref0}", cid="F3")
593
+ # 逐项:五项全等 + 源未变 ⇒ 跳过;任一项不等 ⇒ 一律进探测(「id 存在」不构成已索引)
594
+ n_rec = {"id": nid, "lineno": n0.get("lineno"), "end": n0.get("end"),
595
+ "hash": n0.get("hash")}
596
+ e_rec = {"root": e.get("root"), "path": e.get("path")}
597
+ rel = refindex._norm_rel(e.get("path"))
598
+ base = refindex._fast_path_verdict(cg, n_rec, e_rec, rel, e.get("root"), True)
599
+ check("13④五项全等 + 源 size/mtime 未变 ⇒ 跳过(probe=None,不读源)",
600
+ base["covered"] is True and base["probe"] is None, str(base), cid="F4")
601
+ holes = []
602
+ for k, v in (("hash", "000000000000"), ("lineno", 999), ("end", 1000)):
603
+ n2 = dict(n_rec)
604
+ n2[k] = v
605
+ if refindex._fast_path_verdict(cg, n2, e_rec, rel, e.get("root"),
606
+ True)["probe"] is None:
607
+ holes.append(k)
608
+ if refindex._fast_path_verdict(cg, n_rec, e_rec, "other_widget.py",
609
+ e.get("root"), True)["probe"] is None:
610
+ holes.append("path")
611
+ v_root = refindex._fast_path_verdict(cg, n_rec, e_rec, rel,
612
+ os.path.join(src, "elsewhere"), True)
613
+ if v_root["probe"] is None:
614
+ holes.append("root")
615
+ check("13④五项任一项不等 ⇒ 一律不跳过(「id 存在」不构成已索引)", not holes,
616
+ f"漏判项={holes}", cid="F4")
617
+ check("13④跨域不一致归因 collision_multi_root 且带两侧 root(不静默、不改判定)",
618
+ (v_root.get("mismatch") or {}).get("kind_of_gap") == "collision_multi_root"
619
+ and bool((v_root.get("mismatch") or {}).get("watermark_root"))
620
+ and bool((v_root.get("mismatch") or {}).get("node_root")),
621
+ str(v_root.get("mismatch")), cid="F4")
622
+
623
+ # ---- ③同 id 而内容变:必须判漂移(id 只做身份,不承担幂等) ----
624
+ wr(p, PROBE_SRC2, bump=5.0) # 行数不变,只有区间哈希变
625
+ res = refindex.check_refs(cg, ledger=refindex.Ledger(lib))
626
+ row = next((r for r in res["stale"] if r["node_id"] == nid), None)
627
+ check("13⑤同 id 而区间哈希变(源改一行)⇒ 必须判 stale,且 hash_expected 取节点 ref",
628
+ res["ok"] is False and row is not None
629
+ and row.get("hash_expected") == (ref0 or {}).get("hash"),
630
+ f"ok={res['ok']} stale={sorted(r['node_id'] for r in res['stale'])}",
631
+ cid="F5")
632
+
633
+ # ---- ④文档侧:措辞 / 指纹 / 行级引用都必须与实现同源 ----
634
+ sec = _doc_block(DOC_PLAN, DOC_SEC_HEAD)
635
+ check("13⑥口径真源可读:文档 §3.1 块非空", bool(sec), DOC_PLAN, cid="F6")
636
+ missing = [lit for lit in DOC_REQUIRED if lit not in sec]
637
+ check("13⑥文档 §3.1 逐字含口径字面(id 位置化 + 幂等判据不看 id)", not missing,
638
+ f"缺字面={missing}", cid="F6")
639
+ fp = hashlib.sha1(sec.encode("utf-8")).hexdigest()[:16]
640
+ check("13⑥文档 §3.1 指纹未陈化(措辞一改即红,须同步 DOC_SEC_FP)",
641
+ fp == DOC_SEC_FP, f"现={fp} 判据侧={DOC_SEC_FP}(照抄现算值即恢复)",
642
+ cid="F6")
643
+ doc_text = ""
644
+ try:
645
+ with open(DOC_PLAN, encoding="utf-8") as f:
646
+ doc_text = f.read()
647
+ except OSError:
648
+ pass # 缺文档 ⇒ 每条引用都记「未出现」,逐条报红
649
+ bad = []
650
+ for literal, fpath, lineno, claim in DOC_CITATIONS:
651
+ if literal not in doc_text:
652
+ bad.append(f"{literal} 未出现在文档")
653
+ continue
654
+ lines = rd(os.path.join(_BASE, fpath)).replace("\r\n", "\n").split("\n")
655
+ got = lines[lineno - 1] if 0 < lineno <= len(lines) else ""
656
+ # 判据取 strip 后前缀:行内缩进不算「引用漂移」(只钉「这一行是不是这句」)
657
+ if not got.strip().startswith(claim):
658
+ now = next((i for i, ln in enumerate(lines, 1)
659
+ if ln.strip().startswith(claim)), "?")
660
+ bad.append(f"{fpath}:{lineno} 现为 {got.strip()[:36]!r}"
661
+ f"({claim.strip()} 现在 {fpath}:{now})")
662
+ check("13⑦文档行级引用仍指向所声称的符号(引用漂移即红,失败信息附现址)",
663
+ not bad, ";".join(bad[:3]), cid="F7")
664
+
665
+
666
+ # ---------------- 14 常驻循环止血(heal 在役面) ----------------
667
+
668
+ def guard_loop_tick_heal(tmp):
669
+ """`SustainLoop._tick_heal`:空转被止住 + 判据未被削弱(不静默、不假装修好)。
670
+
671
+ 直调 `_tick_heal()`(**不起线程、不打心跳戳**),`d` 显式落到临时目录——
672
+ 绝不写 ~/.mdcg。落点取循环自己的账(`heals` / `last_diagnose`)与 heal 的
673
+ 返回值,不断言内部实现细节。
674
+ """
675
+ lib = _mk(tmp, "lt_lib")
676
+ domA = _mk(tmp, "lt_domA")
677
+ domB = _mk(tmp, "lt_domB")
678
+ for d, nm in ((domA, "lt_a.py"), (domB, "lt_b.py")):
679
+ wr(os.path.join(d, nm), PROBE_SRC)
680
+ cg = MdCG(lib)
681
+ _index_into(cg, domA, refindex.Ledger(lib))
682
+ _index_into(cg, domB, refindex.Ledger(lib))
683
+ cg.flush()
684
+ sustain._HEAL_MEMO.pop((os.path.abspath(lib), "rebuild_refs"), None)
685
+ pa = os.path.join(domA, "lt_a.py")
686
+ wr(pa, PROBE_SRC2, bump=5.0) # 一域漂移
687
+ shutil.rmtree(domB) # 一域悬空 ⇒ 重建必带 roots_missing
688
+ loop = sustain.SustainLoop(cg, "p28_guard_loop", heal_interval=300.0,
689
+ auto_heal=True, d=os.path.join(tmp, "lt_sustain"))
690
+ orig = sustain.heal
691
+ seen = {}
692
+
693
+ def _spy(cg_, **kw):
694
+ res = orig(cg_, **kw)
695
+ seen["kw"] = kw
696
+ seen["actions"] = res.get("actions")
697
+ return res
698
+
699
+ sustain.heal = _spy
700
+ try:
701
+ loop._tick_heal()
702
+ finally:
703
+ sustain.heal = orig
704
+ check("14①常驻循环按 repeat_guard=True + heal_interval 调 heal(空转止在这一面)",
705
+ (seen.get("kw") or {}).get("repeat_guard") is True
706
+ and (seen.get("kw") or {}).get("heal_interval") == 300.0,
707
+ str(sorted((seen.get("kw") or {}))), cid="G1")
708
+ h1 = loop.heals[-1]
709
+ check("14②首轮真跑并如实记没修好:actions 含 rebuild_refs 且 failed 单列该码",
710
+ "rebuild_refs" in h1["actions"] and "rebuild_refs" in h1["failed"],
711
+ str(h1), cid="G2")
712
+ check("14③判据未被削弱:诊断面照旧报出病灶(ref_stale 仍在 last_diagnose)",
713
+ "ref_stale" in ((loop.last_diagnose or {}).get("issues") or []),
714
+ str(loop.last_diagnose), cid="G2")
715
+ # 同信号第二轮:先让信号与上次一致(重建已把索引推到新代 ⇒ 写回旧正文再漂一次)
716
+ wr(pa, PROBE_SRC, bump=5.0)
717
+ sustain.heal = _spy
718
+ try:
719
+ loop._tick_heal()
720
+ finally:
721
+ sustain.heal = orig
722
+ row2 = next((a for a in (seen.get("actions") or [])
723
+ if a.get("code") == "rebuild_refs"), None)
724
+ check("14④同一信号重复失败 ⇒ 被拦下并如实记 repeat_failure(不是又重跑一次)",
725
+ bool(row2) and row2.get("applied") is False
726
+ and row2.get("reason") == "repeat_failure" and row2.get("streak") == 1,
727
+ str(row2), cid="G3")
728
+ h2 = loop.heals[-1]
729
+ check("14④拦下也照样进账:heals 记该动作 + failed 单列,诊断面仍报病灶(都不静默)",
730
+ "rebuild_refs" in h2["actions"] and "rebuild_refs" in h2["failed"]
731
+ and "ref_stale" in ((loop.last_diagnose or {}).get("issues") or []),
732
+ str(h2), cid="G3")
733
+ sustain._HEAL_MEMO.clear()
734
+
735
+
736
+ # ---------------- 红基线(判别力自证) ----------------
737
+
738
+ #: 定点变异表:(标签, 现实现锚点, 改回后的旧行为)。
739
+ #: **源 = 工作区 md_cg/refindex.py 的文本**(不取 git HEAD:HEAD 在改动提交那一刻就等于
740
+ #: 实现本身,判别力静默失效——同款教训见 md_cg/test_logref.py 的 _MUTATIONS 头注)。
741
+ #: 锚点必须**恰好命中 1 次**:0 次=漂移、≥2 次=歧义,皆 ANCHOR-MISS → fail-closed。
742
+ _MUTATIONS = (
743
+ ("快路径探测数据:节点自身 frontmatter ref(现行)→ 水位条目(v1 旧行为)",
744
+ ' out["probe"] = {**ref, "path": ref.get("path") or rel,\n'
745
+ ' "root": ref.get("root") or src_root}',
746
+ ' out["probe"] = {**n, "path": rel, "root": src_root}'),
747
+ )
748
+
749
+ #: 变异态下**必须**转红的项:恰好命中这个集合(多一项少一项都报红):
750
+ #: A1 6①只推水位 ⇒ v1 拿水位当探测数据 ⇒ 哈希「匹配」而静默 ok=True(漏报)
751
+ #: B1 6②只推节点 ⇒ v1 拿旧水位当探测数据 ⇒ 旧哈希对不上新源(误报 stale)
752
+ #: C1 6③篡改节点 ref.hash ⇒ v1 根本不看节点自述 ⇒ 静默 ok=True
753
+ EXPECTED_RED = ("A1", "B1", "C1")
754
+
755
+
756
+ # 生效条件:锚点命中次数不是恰好 1 次时由 _mutated_verdict 抛出(fail-closed)。
757
+ class AnchorMiss(RuntimeError):
758
+ pass
759
+
760
+
761
+ #: 导入时的实现体字节码(红基线模式会把变异体挂回该属性,故不能与属性现值比)
762
+ _IMPL_CODEC = refindex._fast_path_verdict.__code__.co_code
763
+
764
+
765
+ def _anchor_selfcheck() -> None:
766
+ """锚点仍在、且 EXPECTED_RED 都是本文件里真实存在的 cid → 静默通过。"""
767
+ _mutated_verdict() # 锚点漂移/歧义 → AnchorMiss
768
+ with open(os.path.abspath(__file__), encoding="utf-8") as f:
769
+ src = f.read()
770
+ gone = [i for i in EXPECTED_RED if ('cid="%s"' % i) not in src]
771
+ if gone:
772
+ raise AnchorMiss("EXPECTED_RED 含本文件里不存在的 check id(红项表已陈化):%s"
773
+ % "、".join(gone))
774
+
775
+
776
+ def _mutated_verdict():
777
+ """读工作区 `md_cg/refindex.py`(**只读,不落盘**)按 _MUTATIONS 逐处文本替换,
778
+ 把变异后的模块体 exec 进真模块 globals 的副本,返回变异后的 _fast_path_verdict。
779
+ 锚点漂移/歧义时抛 AnchorMiss。"""
780
+ p = os.path.join(_BASE, "md_cg", "refindex.py")
781
+ with open(p, encoding="utf-8") as f:
782
+ text = f.read()
783
+ for label, anchor, repl in _MUTATIONS:
784
+ n = text.count(anchor)
785
+ if n != 1:
786
+ raise AnchorMiss(
787
+ "ANCHOR-MISS:定点变异锚点漂移/歧义 —— %s 命中 %d 次(应恰好 1 次)"
788
+ "|锚点首行:%s" % (label, n, anchor.splitlines()[0].strip()))
789
+ text = text.replace(anchor, repl)
790
+ g = dict(vars(refindex)) # __name__ / __package__ 保留 ⇒ 相对 import 照旧解析
791
+ exec(compile(text, "<mutated md_cg/refindex.py>", "exec"), g)
792
+ fn = g["_fast_path_verdict"]
793
+ # 与**导入时**的实现体比(不能比当前属性:红基线模式会把变异体挂回该属性)
794
+ if fn.__code__.co_code == _IMPL_CODEC:
795
+ raise AnchorMiss("ANCHOR-MISS:变异后字节码与实现相同(替换未生效)")
796
+ return fn
797
+
798
+
799
+ def main(argv=None):
800
+ ap = argparse.ArgumentParser(
801
+ description="P28 守卫:ref 增量索引 + 漂移/悬空巡检(R3)")
802
+ ap.add_argument("--head-baseline", action="store_true",
803
+ help="红基线:对工作区 md_cg/refindex.py 做定点变异(快路径探测"
804
+ "数据改回水位条目=v1 旧行为),断言红项集合 == 预期红项")
805
+ args = ap.parse_args(argv)
806
+ # 锚点自证:**两态都查**(绿态也查)——锚点漂移后若只查红基线,绿线会继续绿着
807
+ # 而判别力已死(本守卫的判别力全靠这张表)。漂移即 fail-closed。
808
+ try:
809
+ _anchor_selfcheck()
810
+ except AnchorMiss as exc:
811
+ print("!! %s" % exc)
812
+ print("结果:锚点自证未过(fail-closed)——定点变异表须同步实现")
813
+ return 2
814
+ if args.head_baseline:
815
+ # 变异体注入(只在本进程内,不落盘)
816
+ refindex._fast_path_verdict = _mutated_verdict()
817
+ print("!! 红基线模式:md_cg/refindex.py 经定点变异 —— %s ⇒ 快路径改用水位条目"
818
+ % ";".join(lbl for lbl, _a, _r in _MUTATIONS))
819
+
820
+ print("=" * 68)
821
+ print("md 认知图 P28 验收 · ref 增量索引 + 漂移/悬空巡检(R3)")
822
+ print("=" * 68)
823
+
824
+ tmp = tempfile.mkdtemp(prefix="mdcg_refcheck_")
825
+ # 部署声明:本用例的读写白名单根(2026-10-10 设计者裁定 dsh #85 选 A 后,
826
+ # MDCG_INGEST_ROOT / MDCG_EXPORT_ROOT 未配置**不再放开**,会回落 mdcg 记忆库根
827
+ # 与工作区;本用例沙箱在两者之外,故必须显式声明——这正是新语义要求的
828
+ # 「部署用环境变量声明可读写的根」。
829
+ os.environ["MDCG_INGEST_ROOT"] = tmp
830
+ os.environ["MDCG_EXPORT_ROOT"] = tmp
831
+ src = os.path.join(tmp, "pkg")
832
+ os.makedirs(src)
833
+ alpha = os.path.join(src, "alpha.py")
834
+ beta = os.path.join(src, "beta.py")
835
+ wr(alpha, ALPHA)
836
+ wr(beta, BETA)
837
+
838
+ corpus.reset_root(ROOT)
839
+ cg = MdCG(ROOT)
840
+
841
+ try:
842
+ # ============================================ ① 统一 ref 协议 + 注册表
843
+ print("\n【1】统一 ref 协议与提取器注册表")
844
+ check("tokens.ALL_OPS 仍收录 ref / index_code",
845
+ "ref" in tokens.ALL_OPS and "index_code" in tokens.ALL_OPS)
846
+ check("REF_KEYS = (code_ref, doc_ref)",
847
+ refindex.REF_KEYS == ("code_ref", "doc_ref"), str(refindex.REF_KEYS))
848
+ reg = refindex.registry()
849
+ check("注册表覆盖 code_ref / doc_ref",
850
+ "code_ref" in reg and "doc_ref" in reg, str(sorted(reg)))
851
+ check("kind_of_path 按后缀判 kind",
852
+ refindex.kind_of_path("a.py") == "code_ref"
853
+ and refindex.kind_of_path("a.md") == "doc_ref"
854
+ and refindex.kind_of_path("a.txt") == "",
855
+ f"py={refindex.kind_of_path('a.py')} md={refindex.kind_of_path('a.md')}")
856
+ try:
857
+ refindex.extract("x = 1", "foo.xyz")
858
+ raised = False
859
+ except ValueError as exc:
860
+ raised = "无提取器" in str(exc) or "无索引提取器" in str(exc)
861
+ check("无提取器后缀显式报错(不静默)", raised)
862
+
863
+ # ============================================ ② 全量索引 + 水位
864
+ print("\n【2】index_code 全量 + _refindex.json 水位")
865
+ out1 = index_read(cg, src)
866
+ check("op=index_code ok", out1.get("ok"), str(out1)[:120])
867
+ check("files == 2(命中后缀)", out1.get("files") == 2, str(out1.get("files")))
868
+ check("indexed > 0 且 error_count == 0",
869
+ out1.get("indexed", 0) > 0 and out1.get("error_count") == 0)
870
+ cg.flush()
871
+ led = refindex.Ledger(ROOT)
872
+ summ = led.summary()
873
+ check("_refindex.json 落盘", summ["exists"], summ["path"])
874
+ check("水位 files == 2", summ["files"] == 2, str(summ["files"]))
875
+ check("水位 nodes == 本轮 indexed",
876
+ summ["nodes"] == out1.get("indexed"), f"{summ['nodes']} == {out1.get('indexed')}")
877
+ n_before = len((cg.index.get("nodes") or {}))
878
+ ids1 = set(out1.get("ids") or [])
879
+
880
+ # ============================================ ③ 增量:未变则跳过
881
+ print("\n【3】增量索引:未变文件不重切(修 F)")
882
+ out2 = index_read(cg, src, incremental=True)
883
+ check("incremental 跳过未变文件",
884
+ out2.get("skipped_unchanged") == 2, str(out2.get("skipped_unchanged")))
885
+ check("incremental 下 indexed == 0(没有重复条目)",
886
+ out2.get("indexed") == 0, str(out2.get("indexed")))
887
+ cg.flush()
888
+ check("增量后节点数不变(幂等)",
889
+ len((cg.index.get("nodes") or {})) == n_before,
890
+ f"{len((cg.index.get('nodes') or {}))} == {n_before}")
891
+ out3 = index_read(cg, src)
892
+ check("默认(不传 incremental)仍是全量:不跳过",
893
+ out3.get("skipped_unchanged") == 0, str(out3.get("skipped_unchanged")))
894
+ check("重跑 id 集合稳定", set(out3.get("ids") or []) == ids1,
895
+ f"{len(set(out3.get('ids') or []))} vs {len(ids1)}")
896
+
897
+ # ============================================ ④ 回读 + 检索带 ref
898
+ print("\n【4】op=ref 回读 + 检索结果带 ref / ref_kind")
899
+ rr = ref_read(cg, node_id=(out1.get("ids") or [""])[0])
900
+ check("op=ref read ok", rr.get("ok") is True, str(rr)[:120])
901
+ check("ref_kind == code_ref", rr.get("ref_kind") == "code_ref",
902
+ str(rr.get("ref_kind")))
903
+ check("回读带 hash_match(源未动 → True)",
904
+ rr.get("hash_match") is True, str(rr.get("hash_match")))
905
+ read = call_tool(cg, "cg", {"op": "read", "query": "compute_energy", "k": 10})
906
+ hits = [r for r in read.get("results", []) if r.get("ref_kind") == "code_ref"]
907
+ check("检索结果带 ref / ref_kind 字段", bool(hits), f"hits={len(hits)}")
908
+ check("检索的 ref 指回源文件",
909
+ bool(hits) and isinstance(hits[0].get("ref"), dict)
910
+ and hits[0]["ref"].get("path") == "alpha.py",
911
+ str(hits[0].get("ref") if hits else None))
912
+
913
+ # ============================================ ⑤ 漂移:stale
914
+ print("\n【5】漂移检测:改源一行 → stale(只读、不抛)")
915
+ wr(alpha, ALPHA_DRIFT, bump=5.0)
916
+ mt_before = {p: os.path.getmtime(p) for p in (alpha, beta)}
917
+ chk = ref_read(cg, action="check")
918
+ mt_after = {p: os.path.getmtime(p) for p in (alpha, beta)}
919
+ check("action=check 返回 ok", isinstance(chk, dict), str(type(chk)))
920
+ check("check 判定 ok=False(有漂移)", chk.get("ok") is False, str(chk.get("ok")))
921
+ check("报出 stale 且指向 alpha.py",
922
+ any((r.get("path") or "").endswith("alpha.py") for r in chk.get("stale", [])),
923
+ str([r.get("path") for r in chk.get("stale", [])]))
924
+ check("巡检只读:源文件 mtime 不变", mt_before == mt_after)
925
+ stale_rr = ref_read(cg, ref={"path": "alpha.py", "lineno": 5, "end": 7,
926
+ "hash": "deadbeef", "root": src, "precise": True})
927
+ check("回读陈旧 ref → hash_match=False / stale=True",
928
+ stale_rr.get("hash_match") is False and stale_rr.get("stale") is True,
929
+ str(stale_rr.get("hash_match")))
930
+ index_read(cg, src) # 重建:漂移消除
931
+ chk2 = ref_read(cg, action="check")
932
+ check("重跑 index_code 后 stale 消除", chk2.get("ok") is True,
933
+ str([r.get("path") for r in chk2.get("stale", [])]))
934
+
935
+ # ============================================ 6 判据分工(#4)
936
+ print("\n【6】判据分工:漂移只由节点 ref 判,水位只判「能不能跳过」")
937
+ guard_verdict_owner(tmp)
938
+
939
+ # ============================================ 7 撞 id 现场(D1 现状钉子)
940
+ print("\n【7】撞 id 现场:归因可见 + 不重复报红(现状钉子 / 迁移触发器)")
941
+ guard_collision(tmp)
942
+
943
+ # ============================================ 8 悬空:dangling
944
+ print("\n【8】悬空检测:删源文件 → dangling")
945
+ os.remove(beta)
946
+ chk3 = ref_read(cg, action="check")
947
+ check("报出 dangling 且指向 beta.py",
948
+ any((r.get("path") or "").endswith("beta.py")
949
+ for r in chk3.get("dangling", [])),
950
+ str([r.get("path") for r in chk3.get("dangling", [])]))
951
+ check("dangling 时 ok=False", chk3.get("ok") is False)
952
+ wr(beta, BETA) # 恢复源 + 重建
953
+ index_read(cg, src)
954
+ chk4 = ref_read(cg, action="check")
955
+ check("恢复源并重建后 dangling 消除",
956
+ chk4.get("ok") is True and not chk4.get("dangling"),
957
+ str([r.get("path") for r in chk4.get("dangling", [])]))
958
+
959
+ # ============================================ 9 截断不静默 + 水位留痕
960
+ print("\n【9】截断显式上报 + 写进水位")
961
+ tr = index_read(cg, src, max_files=1)
962
+ check("max_files=1 → truncated=True", tr.get("truncated") is True)
963
+ check("truncated_reason 指出 max_files",
964
+ "max_files" in (tr.get("truncated_reason") or ""),
965
+ str(tr.get("truncated_reason")))
966
+ stt = ref_read(cg, action="stat")
967
+ check("action=stat 返回水位与 last_index",
968
+ isinstance(stt.get("ledger"), dict)
969
+ and "last_index" in stt["ledger"], str(stt.get("ledger"))[:120])
970
+ check("水位记下 last_index.truncated=True",
971
+ (stt["ledger"].get("last_index") or {}).get("truncated") is True)
972
+ index_read(cg, src)
973
+ check("重跑全量后 last_index.truncated 回 False",
974
+ (ref_read(cg, action="stat")["ledger"].get("last_index") or {})
975
+ .get("truncated") is False)
976
+
977
+ # ============================================ 10 接入 sustain.diagnose / heal
978
+ print("\n【10】接入 sustain:diagnose 报 ref_stale,heal 的 rebuild_refs 修复")
979
+ wr(alpha, ALPHA_DRIFT2, bump=5.0)
980
+ diag = call_tool(cg, "cg", {"op": "sustain", "action": "diagnose"})
981
+ codes = {i["code"] for i in diag.get("issues", [])}
982
+ check("diagnose issues 含 ref_stale", "ref_stale" in codes, str(sorted(codes)))
983
+ check("diagnose stats 报 ref_stale / ref_checked",
984
+ diag.get("stats", {}).get("ref_stale", 0) >= 1
985
+ and "ref_checked" in diag.get("stats", {}),
986
+ str(diag.get("stats", {}).get("ref_stale")))
987
+ check("巡检不改源文件(改动仍在)", ALPHA_MARK2 in rd(alpha))
988
+ heal = call_tool(cg, "cg", {"op": "sustain", "action": "heal"})
989
+ check("heal 动作含 rebuild_refs",
990
+ any(a.get("code") == "rebuild_refs" for a in heal.get("actions", [])),
991
+ str([a.get("code") for a in heal.get("actions", [])]))
992
+ diag2 = call_tool(cg, "cg", {"op": "sustain", "action": "diagnose"})
993
+ codes2 = {i["code"] for i in diag2.get("issues", [])}
994
+ check("heal 后 ref_stale 消除", "ref_stale" not in codes2, str(sorted(codes2)))
995
+ check("heal 不修改源文件", ALPHA_MARK2 in rd(alpha))
996
+
997
+ # ============================================ 11 派生物不膨胀 + 未知 action
998
+ print("\n【11】_refindex.json 不膨胀 / 未知 action 不静默")
999
+ led2 = refindex.Ledger(ROOT)
1000
+ p = led2.path
1001
+ size1 = os.path.getsize(p)
1002
+ for _ in range(3):
1003
+ index_read(cg, src)
1004
+ size2 = os.path.getsize(p)
1005
+ check("_refindex.json 体积有界(< 20KB)", size2 < 20000, f"{size2}B")
1006
+ check("重跑 3 次不膨胀", size2 == size1, f"{size1} -> {size2}")
1007
+ check("水位 files 仍 == 2(无幽灵条目)", led2.summary()["files"] == 2,
1008
+ str(led2.summary()["files"]))
1009
+ try:
1010
+ r = ref_read(cg, action="nope")
1011
+ silent = (isinstance(r, dict) and r.get("ok") is True)
1012
+ except Exception: # noqa: BLE001
1013
+ silent = False
1014
+ check("未知 action 不静默成功", silent is False)
1015
+
1016
+ # ============================================ 12 自愈止血 + 水位保全
1017
+ print("\n【12】自愈止血与水位保全(水位不被写空 / 失败不记成成功 / 同因不重试)")
1018
+ guard_heal_bounded(tmp)
1019
+
1020
+ # ============================================ 13 口径钉子(#5)
1021
+ print("\n【13】口径钉子:id 位置化 + 幂等判据由区间哈希承担(文档↔实现同源)")
1022
+ guard_id_and_idempotence(tmp)
1023
+
1024
+ # ============================================ 14 常驻循环止血
1025
+ print("\n【14】常驻循环止血(repeat_guard / 不静默 / 不假装修好)")
1026
+ guard_loop_tick_heal(tmp)
1027
+
1028
+ finally:
1029
+ shutil.rmtree(tmp, ignore_errors=True)
1030
+
1031
+ print("\n" + "=" * 68)
1032
+ print(f"通过 {PASS} / 失败 {FAIL}")
1033
+ if FAILS:
1034
+ print("失败项:\n - " + "\n - ".join(FAILS))
1035
+ print("=" * 68)
1036
+ if args.head_baseline:
1037
+ # 短码集合比对:变异后**恰好**命中预期红项(多一项少一项都报红)。无 cid 的
1038
+ # 失败在 check 里记了哨兵 "?<name>",故「波及到没登记的断言」也在此报红。
1039
+ got, exp = sorted(set(RED)), sorted(EXPECTED_RED)
1040
+ if got == exp:
1041
+ print("红基线符合预期:变异后恰好命中 %d 项 —— %s"
1042
+ % (len(exp), "、".join(exp)))
1043
+ return 0
1044
+ print("红基线与预期不符(fail-closed)——预期红项:%s;实得红项:%s"
1045
+ % ("、".join(exp) or "(无)", "、".join(got) or "(无)"))
1046
+ return 1
1047
+ return 1 if FAIL else 0
1048
+
1049
+
1050
+ if __name__ == "__main__":
1051
+ if hasattr(sys.stdout, "reconfigure"):
1052
+ sys.stdout.reconfigure(encoding="utf-8", errors="replace")
1053
+ sys.exit(main())