@modusensus/dsh-mneme 0.8.13 → 0.8.15

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 (271) hide show
  1. package/LICENSE +21 -21
  2. package/README.md +655 -626
  3. package/assets/icon.svg +3 -3
  4. package/bin/cli.mjs +689 -689
  5. package/bin/dsh-mneme-mcp.mjs +510 -510
  6. package/bin/dsh-mneme-serve.mjs +151 -0
  7. package/cordis.patch.yml +21 -21
  8. package/docs/CONFIGURATION.md +263 -0
  9. package/docs/CONTINUITY.md +124 -0
  10. package/docs/DAEMON.md +46 -0
  11. package/docs/ENTITIES.md +245 -0
  12. package/docs/LOCAL_MODEL.md +222 -0
  13. package/docs/SEMANTIC.md +256 -0
  14. package/docs/SLEEP.md +164 -0
  15. package/docs/STORAGE.md +89 -0
  16. package/lib/api-standalone.js +673 -572
  17. package/lib/api.js +1437 -1433
  18. package/lib/bootstrap.js +211 -211
  19. package/lib/client.js +5444 -5295
  20. package/lib/commands.js +95 -95
  21. package/lib/config.js +880 -836
  22. package/lib/content-hash.js +54 -54
  23. package/lib/continuity.js +220 -220
  24. package/lib/document-index.js +94 -94
  25. package/lib/document.js +303 -303
  26. package/lib/dream/clustering.js +118 -118
  27. package/lib/dream/decisions.js +674 -674
  28. package/lib/dream/narratives.js +44 -44
  29. package/lib/dream/sleep.js +819 -819
  30. package/lib/dream.js +1501 -1501
  31. package/lib/embedding.js +177 -177
  32. package/lib/entities/extractor.js +252 -252
  33. package/lib/graph/anchoring.js +179 -179
  34. package/lib/guide.js +74 -74
  35. package/lib/heat.js +144 -144
  36. package/lib/hot-memory.js +46 -46
  37. package/lib/index.js +535 -671
  38. package/lib/inject.js +464 -450
  39. package/lib/lang.js +248 -229
  40. package/lib/local-embedder.js +358 -358
  41. package/lib/maintenance.js +142 -142
  42. package/lib/mirror.js +330 -330
  43. package/lib/organize.js +450 -450
  44. package/lib/peak-hours.js +117 -117
  45. package/lib/placement.js +13 -0
  46. package/lib/pre-inject-gate.js +340 -0
  47. package/lib/quality-filter.js +126 -126
  48. package/lib/recall-stats.js +175 -175
  49. package/lib/reranker.js +269 -269
  50. package/lib/runtime/adopt.js +176 -176
  51. package/lib/runtime/closure.js +172 -172
  52. package/lib/runtime/download.js +247 -247
  53. package/lib/runtime/layout.js +265 -265
  54. package/lib/runtime/loader.js +252 -252
  55. package/lib/runtime/model-fetch.js +361 -361
  56. package/lib/runtime/provision.js +237 -237
  57. package/lib/runtime/tarball.js +130 -130
  58. package/lib/runtime/verify.js +223 -204
  59. package/lib/scope.js +146 -146
  60. package/lib/search/adaptive.js +60 -60
  61. package/lib/search/bm25.js +96 -96
  62. package/lib/semantic.js +229 -0
  63. package/lib/sensitive-scan.js +209 -209
  64. package/lib/serve.js +215 -0
  65. package/lib/service.js +2663 -2596
  66. package/lib/settings.js +517 -509
  67. package/lib/store.js +3007 -3007
  68. package/lib/summarize.js +891 -856
  69. package/lib/tools.js +743 -737
  70. package/lib/vector-index.js +110 -110
  71. package/lib/version-check.js +122 -122
  72. package/lib/write-admission.js +402 -402
  73. package/locale/en.json +6 -6
  74. package/locale/zh.json +6 -6
  75. package/package.json +8 -4
  76. package/runtime-manifest.json +526 -526
  77. package/scripts/benchmark-embed.js +201 -201
  78. package/scripts/benchmark-recall.js +249 -249
  79. package/scripts/benchmark-rerank.js +166 -166
  80. package/scripts/build-runtime-manifest.mjs +224 -224
  81. package/scripts/changelog-prep.mjs +36 -0
  82. package/scripts/check-sync.js +42 -42
  83. package/scripts/e2e-dsh.js +220 -220
  84. package/scripts/mneme-runtime.mjs +277 -277
  85. package/scripts/stress-dsh.js +255 -255
  86. package/scripts/sync-lib.js +47 -47
  87. package/scripts/sync-test-badge.mjs +69 -69
  88. package/scripts/test-count-sync.mjs +44 -44
  89. package/src/api-standalone.js +673 -572
  90. package/src/api.js +1437 -1433
  91. package/src/bootstrap.js +211 -211
  92. package/src/commands.js +95 -95
  93. package/src/config.js +880 -836
  94. package/src/content-hash.js +54 -54
  95. package/src/continuity.js +220 -220
  96. package/src/document-index.js +94 -94
  97. package/src/document.js +303 -303
  98. package/src/dream/clustering.js +118 -118
  99. package/src/dream/decisions.js +674 -674
  100. package/src/dream/narratives.js +44 -44
  101. package/src/dream/sleep.js +819 -819
  102. package/src/dream.js +1501 -1501
  103. package/src/embedding.js +177 -177
  104. package/src/entities/extractor.js +252 -252
  105. package/src/graph/anchoring.js +179 -179
  106. package/src/guide.js +74 -74
  107. package/src/heat.js +144 -144
  108. package/src/hot-memory.js +46 -46
  109. package/src/index.js +535 -671
  110. package/src/inject.js +464 -450
  111. package/src/lang.js +248 -229
  112. package/src/local-embedder.js +358 -358
  113. package/src/maintenance.js +142 -142
  114. package/src/mirror.js +330 -330
  115. package/src/organize.js +450 -450
  116. package/src/peak-hours.js +117 -117
  117. package/src/placement.js +13 -0
  118. package/src/pre-inject-gate.js +340 -0
  119. package/src/quality-filter.js +126 -126
  120. package/src/recall-stats.js +175 -175
  121. package/src/reranker.js +269 -269
  122. package/src/runtime/adopt.js +176 -176
  123. package/src/runtime/closure.js +172 -172
  124. package/src/runtime/download.js +247 -247
  125. package/src/runtime/layout.js +265 -265
  126. package/src/runtime/loader.js +252 -252
  127. package/src/runtime/model-fetch.js +361 -361
  128. package/src/runtime/provision.js +237 -237
  129. package/src/runtime/tarball.js +130 -130
  130. package/src/runtime/verify.js +223 -204
  131. package/src/scope.js +146 -146
  132. package/src/search/adaptive.js +60 -60
  133. package/src/search/bm25.js +96 -96
  134. package/src/semantic.js +229 -0
  135. package/src/sensitive-scan.js +209 -209
  136. package/src/serve.js +215 -0
  137. package/src/service.js +2663 -2596
  138. package/src/settings.js +517 -509
  139. package/src/store.js +3007 -3007
  140. package/src/summarize.js +891 -856
  141. package/src/tools.js +743 -737
  142. package/src/vector-index.js +110 -110
  143. package/src/version-check.js +122 -122
  144. package/src/write-admission.js +402 -402
  145. package/test/adaptive-inject-budget.test.js +85 -85
  146. package/test/api-routes-gaps.test.js +218 -218
  147. package/test/api.test.js +1702 -1688
  148. package/test/audit.test.js +448 -448
  149. package/test/baseline-persist.test.js +248 -248
  150. package/test/benchmark.test.js +94 -94
  151. package/test/bootstrap.test.js +220 -220
  152. package/test/changelog-prep.test.js +60 -0
  153. package/test/client.test.js +1230 -1169
  154. package/test/clustering.test.js +100 -100
  155. package/test/commands.test.js +106 -106
  156. package/test/config.test.js +66 -66
  157. package/test/conflict-freeze.test.js +300 -300
  158. package/test/content-hash.test.js +204 -204
  159. package/test/continuity.test.js +333 -333
  160. package/test/document-dir.test.js +233 -233
  161. package/test/document.test.js +671 -671
  162. package/test/dream-archive-guardrail.test.js +116 -116
  163. package/test/dream-candidate-hybrid.test.js +281 -281
  164. package/test/dream-coverage-degrade.test.js +86 -86
  165. package/test/dream-failure-backoff.test.js +173 -173
  166. package/test/dream-peak-hours.test.js +194 -194
  167. package/test/dream.test.js +1385 -1383
  168. package/test/embedding-ready.test.js +196 -196
  169. package/test/entities.test.js +627 -547
  170. package/test/entity-recall.test.js +179 -179
  171. package/test/epistemic.test.js +325 -325
  172. package/test/fnew-0112.test.js +311 -311
  173. package/test/fnew-03.test.js +422 -422
  174. package/test/graph-anchoring.test.js +184 -184
  175. package/test/graph-api.test.js +175 -175
  176. package/test/graph-inject-budget.test.js +103 -103
  177. package/test/graph-passive-confirm.test.js +91 -91
  178. package/test/graph-weight-evolution.test.js +104 -104
  179. package/test/heat.test.js +175 -175
  180. package/test/helpers/dream-mock.js +82 -82
  181. package/test/helpers/peer-worker.mjs +25 -25
  182. package/test/helpers/runtime-source.js +60 -60
  183. package/test/helpers/tar-builder.js +80 -80
  184. package/test/helpers/write-admission-samples.js +308 -308
  185. package/test/hot-memory.test.js +145 -145
  186. package/test/inject-content-cap.test.js +91 -91
  187. package/test/inject-first-round-bm25.test.js +89 -89
  188. package/test/inject-heat.test.js +77 -61
  189. package/test/inject-parent-gate.test.js +106 -106
  190. package/test/inject-pools.test.js +194 -194
  191. package/test/inject-recall.test.js +105 -105
  192. package/test/inject-rotation.test.js +161 -161
  193. package/test/inject.test.js +420 -420
  194. package/test/lang.test.js +87 -87
  195. package/test/lastrunat-persist.test.js +101 -101
  196. package/test/lib-smoke.test.js +96 -96
  197. package/test/llm-audit.test.js +598 -598
  198. package/test/llm-message-source.test.js +207 -207
  199. package/test/local-embedder.test.js +292 -292
  200. package/test/maintenance.test.js +391 -391
  201. package/test/mcp-stdio.test.js +385 -385
  202. package/test/mirror-dirty.test.js +424 -424
  203. package/test/mirror-edit-digest.test.js +187 -187
  204. package/test/mirror-generation.test.js +499 -499
  205. package/test/mirror-types.test.js +208 -208
  206. package/test/mirror.test.js +249 -249
  207. package/test/narrative-bars.test.js +162 -162
  208. package/test/organize.test.js +418 -418
  209. package/test/peer-blockers.test.js +193 -193
  210. package/test/placement.test.js +36 -0
  211. package/test/policy-epoch.test.js +259 -259
  212. package/test/pre-inject-gate.test.js +434 -0
  213. package/test/quality-filter.test.js +186 -186
  214. package/test/reasoning-effort.test.js +752 -752
  215. package/test/recall-anti-update.test.js +82 -82
  216. package/test/recall-evals.test.js +253 -253
  217. package/test/recall-layer.test.js +326 -326
  218. package/test/recall-stats.test.js +189 -189
  219. package/test/receipt-chain.test.js +451 -451
  220. package/test/reflection.test.js +226 -226
  221. package/test/reindex-backfill.test.js +62 -62
  222. package/test/reranker-single-logit.test.js +73 -73
  223. package/test/reranker.test.js +197 -197
  224. package/test/resident-status.test.js +83 -83
  225. package/test/runtime-adopt.test.js +206 -206
  226. package/test/runtime-cli.test.js +89 -89
  227. package/test/runtime-closure.test.js +193 -193
  228. package/test/runtime-download.test.js +302 -302
  229. package/test/runtime-layout.test.js +163 -163
  230. package/test/runtime-loader.test.js +255 -255
  231. package/test/runtime-manifest.test.js +159 -159
  232. package/test/runtime-model-fetch.test.js +348 -348
  233. package/test/runtime-provision.test.js +101 -101
  234. package/test/runtime-tarball.test.js +81 -81
  235. package/test/runtime-verify.test.js +319 -279
  236. package/test/scope-candidates.test.js +239 -239
  237. package/test/scope-explicit.test.js +452 -452
  238. package/test/scope-inject-soft.test.js +106 -106
  239. package/test/scope-retrieval.test.js +197 -197
  240. package/test/scope-storage.test.js +255 -255
  241. package/test/scope-strict.test.js +367 -343
  242. package/test/search-fusion.test.js +90 -90
  243. package/test/semantic.test.js +124 -124
  244. package/test/sensitive-scan.test.js +207 -207
  245. package/test/serve-bin.test.js +121 -0
  246. package/test/serve.test.js +196 -0
  247. package/test/service-search.test.js +174 -174
  248. package/test/service.test.js +405 -405
  249. package/test/settings.test.js +261 -261
  250. package/test/skip-invalid-restore.test.js +262 -262
  251. package/test/sleep-action-set.test.js +422 -422
  252. package/test/sleep-heat.test.js +126 -126
  253. package/test/sleep-scheduler-rearm.test.js +89 -89
  254. package/test/sleep.test.js +527 -527
  255. package/test/standalone-api.test.js +751 -544
  256. package/test/store-vector-cache.test.js +98 -98
  257. package/test/store.test.js +396 -396
  258. package/test/stress.test.js +209 -209
  259. package/test/summarize-occurred-at.test.js +127 -0
  260. package/test/summarize-reasoning-effort.test.js +189 -189
  261. package/test/summarize.test.js +1465 -1465
  262. package/test/supersede-state-tracking.test.js +394 -0
  263. package/test/test-count-sync.test.js +60 -60
  264. package/test/tools-dto-schema-matrix.test.js +325 -325
  265. package/test/tools-memory-get-schema.test.js +65 -65
  266. package/test/tools.test.js +428 -374
  267. package/test/updated-at-semantics.test.js +113 -113
  268. package/test/vector-index.test.js +221 -221
  269. package/test/version-check.test.js +152 -152
  270. package/test/webServer-optional.test.js +82 -82
  271. package/test/write-admission.test.js +668 -668
@@ -1,279 +1,319 @@
1
- import test from "node:test";
2
- import assert from "node:assert/strict";
3
- import { mkdirSync, mkdtempSync, writeFileSync } from "node:fs";
4
- import { tmpdir } from "node:os";
5
- import { join } from "node:path";
6
- import { DEFAULT_PROBE_TEXTS, defaultEngine, verifyFunctional, verifyPayload } from "../src/runtime/verify.js";
7
- import { TRANSFORMERS_ENTRY } from "../src/runtime/layout.js";
8
- import { pathToFileURL } from "node:url";
9
-
10
- // 验证三件套(issue #131 / PR-A)。重点不是「能跑通」,而是「坏的那几种能被抓住」:
11
- // 结构对但原生二进制与本机不匹配、返回非归一化向量、向量退化、加载直接抛——
12
- // 这些都必须变成一条可读的失败结论,而不是让调用方拿到一个假的 ok。
13
-
14
- /** 造一个结构上合法的 payload(只有形状,没有真东西)。 */
15
- function makePayload({ platform = "win32", arch = "x64", withEntry = true } = {}) {
16
- const dir = mkdtempSync(join(tmpdir(), "mneme-verify-"));
17
- const write = (rel, body) => {
18
- const full = join(dir, rel);
19
- mkdirSync(join(full, ".."), { recursive: true });
20
- writeFileSync(full, body);
21
- };
22
- for (const name of ["@huggingface/transformers", "onnxruntime-node", "sharp"]) {
23
- write(join("node_modules", name, "package.json"), JSON.stringify({ name, version: name === "@huggingface/transformers" ? "4.2.0" : "1.0.0" }));
24
- }
25
- if (withEntry) write(TRANSFORMERS_ENTRY, "export const x = 1;\n");
26
- mkdirSync(join(dir, "node_modules", "onnxruntime-node", "bin", "napi-v6", platform, arch), { recursive: true });
27
- return dir;
28
- }
29
-
30
- /** 归一化的假向量:`rows` 行,每行 dim 维,方向由 seed 决定。 */
31
- function fakeRows(rows, dim, seed = 0) {
32
- return Array.from({ length: rows }, (_, r) => {
33
- const v = Array.from({ length: dim }, (_, i) => Math.sin((i + 1) * (r + 1 + seed)));
34
- const length = Math.hypot(...v);
35
- return v.map((x) => x / length);
36
- });
37
- }
38
-
39
- /** 注入一个「什么都能返回」的 engine。 */
40
- function engineReturning(value) {
41
- return async () => async () => value;
42
- }
43
-
44
- test("功能验证:正常的向量通过,并报出维度与耗时", async () => {
45
- const result = await verifyFunctional("/whatever", { engine: engineReturning(fakeRows(2, 512)) });
46
- assert.equal(result.ok, true, result.reason);
47
- assert.equal(result.dim, 512);
48
- assert.equal(result.rows, 2);
49
- assert.ok(result.elapsedMs >= 0);
50
- });
51
-
52
- test("功能验证:行数不符要报出来(而不是只取第一行当作通过)", async () => {
53
- const result = await verifyFunctional("/whatever", { engine: engineReturning(fakeRows(1, 512)) });
54
- assert.equal(result.ok, false);
55
- assert.match(result.reason, /返回行数不符/);
56
- });
57
-
58
- test("功能验证:非归一化向量要报出来(说明 normalize 没生效)", async () => {
59
- const rows = fakeRows(2, 8).map((row) => row.map((x) => x * 3));
60
- const result = await verifyFunctional("/whatever", { engine: engineReturning(rows) });
61
- assert.equal(result.ok, false);
62
- assert.match(result.reason, /不是 L2 归一化/);
63
- });
64
-
65
- test("功能验证:向量退化成同一方向要报出来", async () => {
66
- const one = fakeRows(1, 16)[0];
67
- const result = await verifyFunctional("/whatever", { engine: engineReturning([one, one]) });
68
- assert.equal(result.ok, false);
69
- assert.match(result.reason, /嵌入退化/);
70
- });
71
-
72
- test("功能验证:加载失败与推理失败分别给出不同措辞", async () => {
73
- const loadFail = await verifyFunctional("/whatever", {
74
- engine: async () => {
75
- throw new Error("Cannot find package 'onnxruntime-node'");
76
- }
77
- });
78
- assert.equal(loadFail.ok, false);
79
- assert.match(loadFail.reason, /加载运行时失败/);
80
- assert.match(loadFail.reason, /onnxruntime-node/);
81
-
82
- const embedFail = await verifyFunctional("/whatever", {
83
- engine: async () => async () => {
84
- throw new Error("boom");
85
- }
86
- });
87
- assert.equal(embedFail.ok, false);
88
- assert.match(embedFail.reason, /推理失败/);
89
- });
90
-
91
- test("功能验证:返回空向量要报出来,而不是当成 dim=0 通过", async () => {
92
- const result = await verifyFunctional("/whatever", { engine: engineReturning([[], []]) });
93
- assert.equal(result.ok, false);
94
- assert.match(result.reason, /空向量/);
95
- });
96
-
97
- test("三件套:结构不过就不再做功能验证(省掉一次无谓的加载)", async () => {
98
- const dir = makePayload({ withEntry: false });
99
- let engineCalled = false;
100
- const result = await verifyPayload(dir, {
101
- platform: "win32",
102
- arch: "x64",
103
- engine: async () => {
104
- engineCalled = true;
105
- return async () => fakeRows(2, 8);
106
- }
107
- });
108
- assert.equal(result.ok, false);
109
- assert.equal(result.functional.ok, false);
110
- assert.match(result.functional.reason, /跳过功能验证/);
111
- assert.equal(engineCalled, false);
112
- });
113
-
114
- test("三件套:结构 + 功能都过、完整性缺失时如实标 unverified 且不判定失败", async () => {
115
- const dir = makePayload();
116
- const result = await verifyPayload(dir, {
117
- platform: "win32",
118
- arch: "x64",
119
- engine: engineReturning(fakeRows(2, 8))
120
- });
121
- assert.equal(result.ok, true);
122
- assert.equal(result.integrity.status, "unverified");
123
- assert.equal(result.integrity.ok, true);
124
- assert.equal(result.version, "4.2.0");
125
- });
126
-
127
- test("三件套:--strict 场景下哈希不匹配即整体不通过", async () => {
128
- const dir = makePayload();
129
- const matched = await verifyPayload(dir, {
130
- platform: "win32",
131
- arch: "x64",
132
- engine: engineReturning(fakeRows(2, 8)),
133
- integrity: { status: "sha512-matched", detail: "tarball 比对通过" }
134
- });
135
- assert.equal(matched.ok, true);
136
- assert.equal(matched.integrity.status, "sha512-matched");
137
-
138
- const mismatch = await verifyPayload(dir, {
139
- platform: "win32",
140
- arch: "x64",
141
- engine: engineReturning(fakeRows(2, 8)),
142
- integrity: { status: "sha512-mismatch", detail: "与钉死的哈希不一致" }
143
- });
144
- assert.equal(mismatch.ok, false);
145
- assert.equal(mismatch.structural.ok, true, "结构仍然是对的,失败来自完整性");
146
- assert.equal(mismatch.functional.ok, true);
147
- });
148
-
149
- test("三件套:清单里记下的来源完整性被认账,不因状态词不同就判失败(issue #268)", async () => {
150
- const dir = makePayload();
151
- // download.js 在落盘前逐个 tarball 比对过 sha512,把这个结论写进清单的形状与词就是下面这样。
152
- writeFileSync(
153
- join(dir, "mneme-runtime.json"),
154
- JSON.stringify({
155
- procedure: "downloaded",
156
- version: "4.2.0",
157
- integrity: { status: "verified", checked: 3, detail: "每个 tarball 的 sha512 与 runtime-manifest.json 一致" }
158
- })
159
- );
160
- const result = await verifyPayload(dir, {
161
- platform: "win32",
162
- arch: "x64",
163
- engine: engineReturning(fakeRows(2, 8))
164
- });
165
- assert.equal(result.integrity.status, "verified");
166
- assert.equal(result.integrity.ok, true, "下载档记下的 verified 是可用结论,不是失败");
167
- assert.match(result.integrity.detail, /每个 tarball/);
168
- assert.equal(result.ok, true);
169
- });
170
-
171
- test("三件套:清单里记下的不一致照样判失败(别把「读到一份结论」变成「一律放过」)", async () => {
172
- const dir = makePayload();
173
- writeFileSync(
174
- join(dir, "mneme-runtime.json"),
175
- JSON.stringify({ procedure: "downloaded", integrity: { status: "sha512-mismatch", detail: "与钉死的哈希不一致" } })
176
- );
177
- const result = await verifyPayload(dir, {
178
- platform: "win32",
179
- arch: "x64",
180
- engine: engineReturning(fakeRows(2, 8))
181
- });
182
- assert.equal(result.integrity.ok, false);
183
- assert.equal(result.integrity.status, "sha512-mismatch");
184
- assert.equal(result.ok, false);
185
- });
186
-
187
- test("三件套:清单里的 integrity 形状非法时 fail-closed,不读成「没验过」(#269 评审)", async () => {
188
- const dir = makePayload();
189
- writeFileSync(join(dir, "mneme-runtime.json"), JSON.stringify({ procedure: "downloaded", integrity: false }));
190
- const result = await verifyPayload(dir, {
191
- platform: "win32",
192
- arch: "x64",
193
- engine: engineReturning(fakeRows(2, 8))
194
- });
195
- assert.equal(result.integrity.status, "malformed");
196
- assert.equal(result.integrity.ok, false, "形状错误不是「验过了」,也不能当成「没记录」放过");
197
- assert.equal(result.ok, false);
198
- assert.match(result.integrity.detail, /形状非法/);
199
- });
200
-
201
- test("三件套:显式报 unverified 与「没传」同判,不因为说出口就变失败", async () => {
202
- const dir = makePayload();
203
- const result = await verifyPayload(dir, {
204
- platform: "win32",
205
- arch: "x64",
206
- engine: engineReturning(fakeRows(2, 8)),
207
- integrity: { status: "unverified" }
208
- });
209
- assert.equal(result.ok, true);
210
- assert.equal(result.integrity.status, "unverified");
211
- assert.match(result.integrity.detail, /没有可比对的原始产物/);
212
- });
213
-
214
- test("假通过防线 ①:行不是数组时返回失败,而不是抛异常", async () => {
215
- // 契约是「任何失败都返回结果对象,不抛」;非数组行会让 .length/.some 直接抛。
216
- const dir = makePayload();
217
- const result = await verifyPayload(dir, {
218
- platform: "win32",
219
- arch: "x64",
220
- engine: engineReturning([null, null])
221
- });
222
- assert.equal(result.ok, false);
223
- assert.match(result.functional.reason, /不是数组/);
224
- });
225
-
226
- test("假通过防线 ②:非有限数值必须判失败(NaN 的比较恒为 false,会静默通过)", async () => {
227
- const dir = makePayload();
228
- for (const bad of [[NaN, 0], [Infinity, 0], ["0.5", "0.5"]]) {
229
- const result = await verifyPayload(dir, {
230
- platform: "win32",
231
- arch: "x64",
232
- engine: engineReturning([bad, [0, 0]])
233
- });
234
- assert.equal(result.ok, false, `含 ${JSON.stringify(bad)} 的向量不该被判通过`);
235
- assert.match(result.functional.reason, /非有限数值/);
236
- }
237
- });
238
-
239
- test("默认 engine 是导出的函数,且探针文本固定(改它等于改验收标准)", () => {
240
- assert.equal(typeof defaultEngine, "function");
241
- assert.deepEqual(DEFAULT_PROBE_TEXTS, ["猫咪喜欢晒太阳", "the quick brown fox"]);
242
- });
243
-
244
- // ── defaultEngine:注入假 transformers 模块,钉住「真 import + 行切分」的契约 ──
245
-
246
- /** 造一个假 transformers 入口:export env + pipeline,把收到的参数记到 globalThis。 */
247
- function fakeEntry(dims) {
248
- const dir = mkdtempSync(join(tmpdir(), "mneme-engine-"));
249
- const entry = join(dir, "entry.mjs");
250
- writeFileSync(entry, `
251
- export const env = {};
252
- export async function pipeline(kind, model, opts) {
253
- globalThis.__engineSeen = { kind, model, opts, cacheDir: env.cacheDir, remote: env.allowRemoteModels };
254
- const d = ${JSON.stringify(dims)};
255
- const data = Float32Array.from({ length: d[0] * d[1] }, (_, i) => i + 1);
256
- return async () => ({ dims: d, data });
257
- }
258
- `);
259
- return pathToFileURL(entry).href;
260
- }
261
-
262
- test("defaultEngine: 假 transformers 入口可驱动嵌入,并按输出维度做行切分", async () => {
263
- const entry = fakeEntry([2, 4]);
264
- const embed = await defaultEngine(entry, { cacheDir: "/tmp/mneme-cache" });
265
- const rows = await embed(["猫咪", "fox"]);
266
- assert.equal(rows.length, 2);
267
- assert.deepEqual(rows[0], [1, 2, 3, 4]);
268
- assert.deepEqual(rows[1], [5, 6, 7, 8]);
269
- // 缓存目录与离线开关必须落到 env:验证过程绝不能偷偷下载模型。
270
- assert.equal(globalThis.__engineSeen.cacheDir, "/tmp/mneme-cache");
271
- assert.equal(globalThis.__engineSeen.remote, false);
272
- assert.equal(globalThis.__engineSeen.kind, "feature-extraction");
273
- });
274
-
275
- test("defaultEngine: 输出宽度为 0 时明确抛错(防止调用方死循环)", async () => {
276
- const entry = fakeEntry([2, 0]);
277
- const embed = await defaultEngine(entry, {});
278
- await assert.rejects(() => embed(["x"]), /维度不可用/);
279
- });
1
+ import test from "node:test";
2
+ import assert from "node:assert/strict";
3
+ import { mkdirSync, mkdtempSync, writeFileSync } from "node:fs";
4
+ import { tmpdir } from "node:os";
5
+ import { join } from "node:path";
6
+ import { DEFAULT_PROBE_TEXTS, defaultEngine, resolveModelCacheDir, verifyFunctional, verifyPayload } from "../src/runtime/verify.js";
7
+ import { TRANSFORMERS_ENTRY, defaultModelCacheDir } from "../src/runtime/layout.js";
8
+ import { pathToFileURL } from "node:url";
9
+
10
+ // 验证三件套(issue #131 / PR-A)。重点不是「能跑通」,而是「坏的那几种能被抓住」:
11
+ // 结构对但原生二进制与本机不匹配、返回非归一化向量、向量退化、加载直接抛——
12
+ // 这些都必须变成一条可读的失败结论,而不是让调用方拿到一个假的 ok。
13
+
14
+ /** 造一个结构上合法的 payload(只有形状,没有真东西)。 */
15
+ function makePayload({ platform = "win32", arch = "x64", withEntry = true } = {}) {
16
+ const dir = mkdtempSync(join(tmpdir(), "mneme-verify-"));
17
+ const write = (rel, body) => {
18
+ const full = join(dir, rel);
19
+ mkdirSync(join(full, ".."), { recursive: true });
20
+ writeFileSync(full, body);
21
+ };
22
+ for (const name of ["@huggingface/transformers", "onnxruntime-node", "sharp"]) {
23
+ write(join("node_modules", name, "package.json"), JSON.stringify({ name, version: name === "@huggingface/transformers" ? "4.2.0" : "1.0.0" }));
24
+ }
25
+ if (withEntry) write(TRANSFORMERS_ENTRY, "export const x = 1;\n");
26
+ mkdirSync(join(dir, "node_modules", "onnxruntime-node", "bin", "napi-v6", platform, arch), { recursive: true });
27
+ return dir;
28
+ }
29
+
30
+ /** 归一化的假向量:`rows` 行,每行 dim 维,方向由 seed 决定。 */
31
+ function fakeRows(rows, dim, seed = 0) {
32
+ return Array.from({ length: rows }, (_, r) => {
33
+ const v = Array.from({ length: dim }, (_, i) => Math.sin((i + 1) * (r + 1 + seed)));
34
+ const length = Math.hypot(...v);
35
+ return v.map((x) => x / length);
36
+ });
37
+ }
38
+
39
+ /** 注入一个「什么都能返回」的 engine。 */
40
+ function engineReturning(value) {
41
+ return async () => async () => value;
42
+ }
43
+
44
+ test("功能验证:正常的向量通过,并报出维度与耗时", async () => {
45
+ const result = await verifyFunctional("/whatever", { engine: engineReturning(fakeRows(2, 512)) });
46
+ assert.equal(result.ok, true, result.reason);
47
+ assert.equal(result.dim, 512);
48
+ assert.equal(result.rows, 2);
49
+ assert.ok(result.elapsedMs >= 0);
50
+ });
51
+
52
+ test("功能验证:行数不符要报出来(而不是只取第一行当作通过)", async () => {
53
+ const result = await verifyFunctional("/whatever", { engine: engineReturning(fakeRows(1, 512)) });
54
+ assert.equal(result.ok, false);
55
+ assert.match(result.reason, /返回行数不符/);
56
+ });
57
+
58
+ test("功能验证:非归一化向量要报出来(说明 normalize 没生效)", async () => {
59
+ const rows = fakeRows(2, 8).map((row) => row.map((x) => x * 3));
60
+ const result = await verifyFunctional("/whatever", { engine: engineReturning(rows) });
61
+ assert.equal(result.ok, false);
62
+ assert.match(result.reason, /不是 L2 归一化/);
63
+ });
64
+
65
+ test("功能验证:向量退化成同一方向要报出来", async () => {
66
+ const one = fakeRows(1, 16)[0];
67
+ const result = await verifyFunctional("/whatever", { engine: engineReturning([one, one]) });
68
+ assert.equal(result.ok, false);
69
+ assert.match(result.reason, /嵌入退化/);
70
+ });
71
+
72
+ test("功能验证:加载失败与推理失败分别给出不同措辞", async () => {
73
+ const loadFail = await verifyFunctional("/whatever", {
74
+ engine: async () => {
75
+ throw new Error("Cannot find package 'onnxruntime-node'");
76
+ }
77
+ });
78
+ assert.equal(loadFail.ok, false);
79
+ assert.match(loadFail.reason, /加载运行时失败/);
80
+ assert.match(loadFail.reason, /onnxruntime-node/);
81
+
82
+ const embedFail = await verifyFunctional("/whatever", {
83
+ engine: async () => async () => {
84
+ throw new Error("boom");
85
+ }
86
+ });
87
+ assert.equal(embedFail.ok, false);
88
+ assert.match(embedFail.reason, /推理失败/);
89
+ });
90
+
91
+ test("功能验证:返回空向量要报出来,而不是当成 dim=0 通过", async () => {
92
+ const result = await verifyFunctional("/whatever", { engine: engineReturning([[], []]) });
93
+ assert.equal(result.ok, false);
94
+ assert.match(result.reason, /空向量/);
95
+ });
96
+
97
+ test("功能验证:cacheDir 与 model 都透传给 engine(面板必须验配置里那一份,不是探针默认值)", async () => {
98
+ // 锁的是调用链上的接线,不是解析规则本身:tools.js 的 verify 分支把 config 的
99
+ // embedModelCacheDir / localEmbedModel 交给 verifyPayload,中间任何一段丢参数,
100
+ // 验的都变成「另一份模型」,结论就不再是用户实际使用情况的证据(issue #387)。
101
+ let seen = null;
102
+ const result = await verifyFunctional("/whatever", {
103
+ engine: async (entryUrl, opts) => {
104
+ seen = opts;
105
+ return async () => fakeRows(2, 8);
106
+ },
107
+ cacheDir: "",
108
+ model: "Xenova/bge-m3"
109
+ });
110
+ assert.equal(result.ok, true, result.reason);
111
+ assert.equal(seen.model, "Xenova/bge-m3");
112
+ assert.equal(seen.cacheDir, "");
113
+ });
114
+
115
+ test("三件套:结构不过就不再做功能验证(省掉一次无谓的加载)", async () => {
116
+ const dir = makePayload({ withEntry: false });
117
+ let engineCalled = false;
118
+ const result = await verifyPayload(dir, {
119
+ platform: "win32",
120
+ arch: "x64",
121
+ engine: async () => {
122
+ engineCalled = true;
123
+ return async () => fakeRows(2, 8);
124
+ }
125
+ });
126
+ assert.equal(result.ok, false);
127
+ assert.equal(result.functional.ok, false);
128
+ assert.match(result.functional.reason, /跳过功能验证/);
129
+ assert.equal(engineCalled, false);
130
+ });
131
+
132
+ test("三件套:结构 + 功能都过、完整性缺失时如实标 unverified 且不判定失败", async () => {
133
+ const dir = makePayload();
134
+ const result = await verifyPayload(dir, {
135
+ platform: "win32",
136
+ arch: "x64",
137
+ engine: engineReturning(fakeRows(2, 8))
138
+ });
139
+ assert.equal(result.ok, true);
140
+ assert.equal(result.integrity.status, "unverified");
141
+ assert.equal(result.integrity.ok, true);
142
+ assert.equal(result.version, "4.2.0");
143
+ });
144
+
145
+ test("三件套:--strict 场景下哈希不匹配即整体不通过", async () => {
146
+ const dir = makePayload();
147
+ const matched = await verifyPayload(dir, {
148
+ platform: "win32",
149
+ arch: "x64",
150
+ engine: engineReturning(fakeRows(2, 8)),
151
+ integrity: { status: "sha512-matched", detail: "tarball 比对通过" }
152
+ });
153
+ assert.equal(matched.ok, true);
154
+ assert.equal(matched.integrity.status, "sha512-matched");
155
+
156
+ const mismatch = await verifyPayload(dir, {
157
+ platform: "win32",
158
+ arch: "x64",
159
+ engine: engineReturning(fakeRows(2, 8)),
160
+ integrity: { status: "sha512-mismatch", detail: "与钉死的哈希不一致" }
161
+ });
162
+ assert.equal(mismatch.ok, false);
163
+ assert.equal(mismatch.structural.ok, true, "结构仍然是对的,失败来自完整性");
164
+ assert.equal(mismatch.functional.ok, true);
165
+ });
166
+
167
+ test("三件套:清单里记下的来源完整性被认账,不因状态词不同就判失败(issue #268)", async () => {
168
+ const dir = makePayload();
169
+ // download.js 在落盘前逐个 tarball 比对过 sha512,把这个结论写进清单的形状与词就是下面这样。
170
+ writeFileSync(
171
+ join(dir, "mneme-runtime.json"),
172
+ JSON.stringify({
173
+ procedure: "downloaded",
174
+ version: "4.2.0",
175
+ integrity: { status: "verified", checked: 3, detail: "每个 tarball 的 sha512 与 runtime-manifest.json 一致" }
176
+ })
177
+ );
178
+ const result = await verifyPayload(dir, {
179
+ platform: "win32",
180
+ arch: "x64",
181
+ engine: engineReturning(fakeRows(2, 8))
182
+ });
183
+ assert.equal(result.integrity.status, "verified");
184
+ assert.equal(result.integrity.ok, true, "下载档记下的 verified 是可用结论,不是失败");
185
+ assert.match(result.integrity.detail, /每个 tarball/);
186
+ assert.equal(result.ok, true);
187
+ });
188
+
189
+ test("三件套:清单里记下的不一致照样判失败(别把「读到一份结论」变成「一律放过」)", async () => {
190
+ const dir = makePayload();
191
+ writeFileSync(
192
+ join(dir, "mneme-runtime.json"),
193
+ JSON.stringify({ procedure: "downloaded", integrity: { status: "sha512-mismatch", detail: "与钉死的哈希不一致" } })
194
+ );
195
+ const result = await verifyPayload(dir, {
196
+ platform: "win32",
197
+ arch: "x64",
198
+ engine: engineReturning(fakeRows(2, 8))
199
+ });
200
+ assert.equal(result.integrity.ok, false);
201
+ assert.equal(result.integrity.status, "sha512-mismatch");
202
+ assert.equal(result.ok, false);
203
+ });
204
+
205
+ test("三件套:清单里的 integrity 形状非法时 fail-closed,不读成「没验过」(#269 评审)", async () => {
206
+ const dir = makePayload();
207
+ writeFileSync(join(dir, "mneme-runtime.json"), JSON.stringify({ procedure: "downloaded", integrity: false }));
208
+ const result = await verifyPayload(dir, {
209
+ platform: "win32",
210
+ arch: "x64",
211
+ engine: engineReturning(fakeRows(2, 8))
212
+ });
213
+ assert.equal(result.integrity.status, "malformed");
214
+ assert.equal(result.integrity.ok, false, "形状错误不是「验过了」,也不能当成「没记录」放过");
215
+ assert.equal(result.ok, false);
216
+ assert.match(result.integrity.detail, /形状非法/);
217
+ });
218
+
219
+ test("三件套:显式报 unverified 与「没传」同判,不因为说出口就变失败", async () => {
220
+ const dir = makePayload();
221
+ const result = await verifyPayload(dir, {
222
+ platform: "win32",
223
+ arch: "x64",
224
+ engine: engineReturning(fakeRows(2, 8)),
225
+ integrity: { status: "unverified" }
226
+ });
227
+ assert.equal(result.ok, true);
228
+ assert.equal(result.integrity.status, "unverified");
229
+ assert.match(result.integrity.detail, /没有可比对的原始产物/);
230
+ });
231
+
232
+ test("假通过防线 ①:行不是数组时返回失败,而不是抛异常", async () => {
233
+ // 契约是「任何失败都返回结果对象,不抛」;非数组行会让 .length/.some 直接抛。
234
+ const dir = makePayload();
235
+ const result = await verifyPayload(dir, {
236
+ platform: "win32",
237
+ arch: "x64",
238
+ engine: engineReturning([null, null])
239
+ });
240
+ assert.equal(result.ok, false);
241
+ assert.match(result.functional.reason, /不是数组/);
242
+ });
243
+
244
+ test("假通过防线 ②:非有限数值必须判失败(NaN 的比较恒为 false,会静默通过)", async () => {
245
+ const dir = makePayload();
246
+ for (const bad of [[NaN, 0], [Infinity, 0], ["0.5", "0.5"]]) {
247
+ const result = await verifyPayload(dir, {
248
+ platform: "win32",
249
+ arch: "x64",
250
+ engine: engineReturning([bad, [0, 0]])
251
+ });
252
+ assert.equal(result.ok, false, `含 ${JSON.stringify(bad)} 的向量不该被判通过`);
253
+ assert.match(result.functional.reason, /非有限数值/);
254
+ }
255
+ });
256
+
257
+ test("默认 engine 是导出的函数,且探针文本固定(改它等于改验收标准)", () => {
258
+ assert.equal(typeof defaultEngine, "function");
259
+ assert.deepEqual(DEFAULT_PROBE_TEXTS, ["猫咪喜欢晒太阳", "the quick brown fox"]);
260
+ });
261
+
262
+ // ── defaultEngine:注入假 transformers 模块,钉住「真 import + 行切分」的契约 ──
263
+
264
+ /** 造一个假 transformers 入口:export env + pipeline,把收到的参数记到 globalThis。 */
265
+ function fakeEntry(dims) {
266
+ const dir = mkdtempSync(join(tmpdir(), "mneme-engine-"));
267
+ const entry = join(dir, "entry.mjs");
268
+ writeFileSync(entry, `
269
+ export const env = {};
270
+ export async function pipeline(kind, model, opts) {
271
+ globalThis.__engineSeen = { kind, model, opts, cacheDir: env.cacheDir, remote: env.allowRemoteModels };
272
+ const d = ${JSON.stringify(dims)};
273
+ const data = Float32Array.from({ length: d[0] * d[1] }, (_, i) => i + 1);
274
+ return async () => ({ dims: d, data });
275
+ }
276
+ `);
277
+ return pathToFileURL(entry).href;
278
+ }
279
+
280
+ test("defaultEngine: 假 transformers 入口可驱动嵌入,并按输出维度做行切分;显式 cacheDir 按原样用", async () => {
281
+ const entry = fakeEntry([2, 4]);
282
+ const embed = await defaultEngine(entry, { cacheDir: "/tmp/mneme-cache" });
283
+ const rows = await embed(["猫咪", "fox"]);
284
+ assert.equal(rows.length, 2);
285
+ assert.deepEqual(rows[0], [1, 2, 3, 4]);
286
+ assert.deepEqual(rows[1], [5, 6, 7, 8]);
287
+ // 缓存目录与离线开关必须落到 env:验证过程绝不能偷偷下载模型。
288
+ assert.equal(globalThis.__engineSeen.cacheDir, "/tmp/mneme-cache");
289
+ assert.equal(globalThis.__engineSeen.opts.cache_dir, "/tmp/mneme-cache", "显式配的目录不许被默认值顶掉");
290
+ assert.equal(globalThis.__engineSeen.remote, false);
291
+ assert.equal(globalThis.__engineSeen.kind, "feature-extraction");
292
+ });
293
+
294
+ test("defaultEngine: cacheDir 省略或为空串时落到用户级默认目录(issue #387)", async () => {
295
+ // #387 的现场:embedModelCacheDir 的 schema 默认值就是空串,而 transformers.js 收到
296
+ // 空/undefined 等于「不设置 env.cacheDir」——它转去找包内相对路径,于是真实嵌入用的是
297
+ // 用户目录那份模型、面板 verify 却在包内找,报出 functional: false 的假失败。
298
+ // 空值必须在这里补成与嵌入侧同一个目录,「验证用哪份模型」才不再取决于调用方有没有填配置。
299
+ for (const given of [undefined, "", " "]) {
300
+ const entry = fakeEntry([2, 4]);
301
+ await defaultEngine(entry, { cacheDir: given });
302
+ const expected = defaultModelCacheDir();
303
+ assert.equal(globalThis.__engineSeen.cacheDir, expected, `env.cacheDir 没落到默认目录(cacheDir=${JSON.stringify(given)})`);
304
+ assert.equal(globalThis.__engineSeen.opts.cache_dir, expected, `pipeline 也没拿到默认目录(cacheDir=${JSON.stringify(given)})`);
305
+ }
306
+ });
307
+
308
+ test("resolveModelCacheDir:只有空值才补默认", () => {
309
+ assert.equal(resolveModelCacheDir("/custom/models"), "/custom/models");
310
+ assert.equal(resolveModelCacheDir(" /custom/models "), "/custom/models");
311
+ assert.equal(resolveModelCacheDir(undefined), defaultModelCacheDir());
312
+ assert.equal(resolveModelCacheDir(null), defaultModelCacheDir());
313
+ });
314
+
315
+ test("defaultEngine: 输出宽度为 0 时明确抛错(防止调用方死循环)", async () => {
316
+ const entry = fakeEntry([2, 0]);
317
+ const embed = await defaultEngine(entry, {});
318
+ await assert.rejects(() => embed(["x"]), /维度不可用/);
319
+ });