sillyspec 3.33.8 → 3.34.0

package/eval/knowledge/datasets/organic.jsonl

@@ -0,0 +1,79 @@
+{"id":"s-tel-01","repo":"sillyspec","face":"decisions","query":"2026-09-28-kgd 方案探索：用枚举词表扫描开放世界形态的需求并自动分类","expected":["decisions/cli-entry.md#D-unmapped-150@v1"],"note":"防复潮：查询应召回已否决的「枚举开放世界」决策（epoch 迁移后落 cli-entry 域）"}
+{"id":"s-tel-02","repo":"sillyspec","face":"decisions","query":"route-hindsight 模块（指标计算/阈值/落库/读取）＋单测 brainstorm Step4/5 指引模板加机制词检索固定动作","expected":["decisions/core-engine.md"],"note":"防复潮注入挂 brainstorm Step2 的决策在 core-engine 域"}
+{"id":"s-tel-03","repo":"sillyspec","face":"decisions","query":"flow.js 前门盘问渲染＋hindsight 提示注入＋收口指标接线＋单测 complete.js 方案步 --done 门检索回显＋config-schema 逃生阀","expected":["decisions/cli-entry.md"],"note":"flow 前门/命令面决策在 cli-entry 域"}
+{"id":"s-tel-04","repo":"sillyspec","face":"fr","query":"标题/摘要提取对段头行免疫：标签行跳过取次行+复合词白名单+模块索引摘要同源+模板首行改概括句","expected":["fr/cli-entry.md"],"subset":"approx","note":"changes.title 三段链路等 FR 在 cli-entry 域"}
+{"id":"s-tel-05","repo":"sillyspec","face":"decisions","query":"用枚举词表扫描开放世界形态的需求并自动分类","expected":["decisions/cli-entry.md#D-unmapped-150@v1"],"note":"同一意图的短查询变体"}
+{"id":"s-man-01","repo":"sillyspec","face":"entries","query":"worktree 清理 node_modules junction 删除","expected":["sillyspec-gotchas.md"],"scope_files":["src/worktree.js"],"subset":"frozen","note":"junction 解链坑在 gotchas 手工条目"}
+{"id":"s-man-02","repo":"sillyspec","face":"decisions","query":"worktree 清理 node_modules junction 删除","expected":["decisions/worktree.md#D-worktree-001@v1"],"scope_files":["src/worktree.js"],"subset":"frozen","note":"两次事故实录的决策条目（hybrid 面已命中 raw_rank=1）"}
+{"id":"s-man-03","repo":"sillyspec","face":"entries","query":"worktree doctor --fix 标记 installed 但 node_modules junction 实际未建","expected":["sillyspec-gotchas.md#worktree-doctor---fix-标记-installed-但-node_modules-junction-实际未建"],"scope_files":["src/worktree.js"],"subset":"frozen","note":"精确锚点命中验证"}
+{"id":"s-man-04","repo":"sillyspec","face":"entries","query":"daemon mcp-server 子进程不继承 claude.exe 环境","expected":["patterns.md#mcp-server-子进程不继承-claudeexe-完整环境env-必须放-mcpserversenv"],"scope_files":["src/mcp-server.js"],"subset":"frozen"}
+{"id":"s-man-05","repo":"sillyspec","face":"entries","query":"execute worktree 里没有 node_modules 子代理 cwd 要显式 worktree 路径","expected":["sillyspec-gotchas.md#execute-worktree-无-node_modules--子代理-cwd-需显式-worktree-路径"],"scope_files":["src/stages/execute.js"],"subset":"frozen"}
+{"id":"s-man-06","repo":"sillyspec","face":"entries","query":"reopen --from step n 后 --done 把后续未执行步骤回填 completed","expected":["sillyspec-gotchas.md#reopen---from-step-n-后---done-会把后续未执行步骤回填-completed"]}
+{"id":"s-man-07","repo":"sillyspec","face":"entries","query":"sqlite pragma foreign_keys 默认关闭 cascade 失效 孤儿变更清理","expected":["sillyspec-gotchas.md#sqlite-pragma-foreign_keys-默认关闭致-cascade-失效清理孤儿变更"]}
+{"id":"s-man-08","repo":"sillyspec","face":"entries","query":"plan execute 子代理 cwd 设到变更目录 嵌套 sillyspec 副作用","expected":["sillyspec-gotchas.md#planexecute-子代理可能把-cwd-设到变更目录产生嵌套-sillyspec-副作用"]}
+{"id":"s-man-09","repo":"sillyspec","face":"fr","query":"sillyspec 版本发布 package.json 同步 R5 锚","expected":["fr/release.md"],"subset":"approx"}
+{"id":"s-man-10","repo":"sillyspec","face":"fr","query":"docs gate 基线棘轮 339 替换 379","expected":["fr/docs.md"],"subset":"approx"}
+{"id":"p-tel-01","repo":"platform","face":"entries","query":"工具输出捕获链路按 Windows 码页探测解码,修复 TOOL_RESULT 中文乱码落库问题","expected":["known-issues.md"],"note":"码页乱码条目在 known-issues"}
+{"id":"p-tel-02","repo":"platform","face":"entries","query":"修复 daemon 码页探测解码器 GBK 流式回退死代码（StringDecoder.write 不抛错）并顺修两处注释漂移","expected":["known-issues.md"]}
+{"id":"p-tel-03","repo":"platform","face":"entries","query":"buildSpawnEnv 注入 PYTHONIOENCODING/PYTHONUTF8 缺省让 claude Bash 子进程源端吐 UTF-8","expected":["known-issues.md"]}
+{"id":"p-tel-04","repo":"platform","face":"entries","query":"修复 codex driver JSON-RPC id 空间碰撞：nextRpcId 与 nextJsonRpcId 统一单计数器分配","expected":["patterns.md"]}
+{"id":"p-tel-05","repo":"platform","face":"entries","query":"后端事件表 ORM + alembic 建表迁移 + conftest 建表","expected":["known-issues.md"],"note":"alembic 并行变更撞 revision 多 head 坑"}
+{"id":"p-tel-06","repo":"platform","face":"entries","query":"后端 schema + service（append_events 去重上限 / list_events 正序增量）","expected":["conventions.md"]}
+{"id":"p-tel-07","repo":"platform","face":"entries","query":"蒸馏任务条点击跳转会话页——distill-task-bar 增查看会话入口，经 /sessions?session=<agent_session_id> 深链打开对应蒸馏会话","expected":["known-issues.md"],"note":"AgentMission.session_id 列不可信判别口径坑"}
+{"id":"p-tel-08","repo":"platform","face":"decisions","query":"claude 会话 one_m 勾选未作用于主模型——ANTHROPIC_MODEL 与 interactive create model 参数按角色映射 one_m 补 [1m] 后缀","expected":["decisions/interactive.md"]}
+{"id":"p-tel-09","repo":"platform","face":"entries","query":"会话页空闲降频：消息队列兜底轮询 5s→30s（SSE queue_changed 事件驱动即时刷新为主链，轮询纯兜底）","expected":["conventions.md#前端-sse-消费统一-fetch-ssetoken-走-authorization-header禁用-eventsource"]}
+{"id":"p-tel-10","repo":"platform","face":"fr","query":"手机端会话页样式优化：降噪（用量条/配置条/占位文案）+ 会话内容放大（气泡字号宽度），仅 mobile 变体生效","expected":["fr/app-mobile-pages.md"],"subset":"approx"}
+{"id":"p-tel-11","repo":"platform","face":"fr","query":"手机端会话页功能入口收纳（用户约束：不砍功能）：用量条收进头部更多菜单会话用量区，配置条 mobile 增设设置展开钮","expected":["fr/app-mobile-pages.md"],"subset":"approx"}
+{"id":"p-tel-12","repo":"platform","face":"fr","query":"手机端会话页二轮样式收敛：正文 16 回调 14px，底部配置区一行化，打断钮 mobile 图标化","expected":["fr/app-mobile-pages.md"],"subset":"approx"}
+{"id":"p-tel-13","repo":"platform","face":"entries","query":"沉淀弹层三修：快速修复多选失效（只能选一个）弹层样式溢出超出视口 新建 agent 配置补供应商/模型选择","expected":["conventions.md#tailwind-md-是视口断点非容器断点侧栏内嵌组件禁用响应式前缀"]}
+{"id":"p-tel-14","repo":"platform","face":"fr","query":"修复 frontend-ci 连续失败：page.test.tsx user_input 断言未随 ql-20260918-003 行为变更同步","expected":["fr/api-types.md"],"subset":"approx"}
+{"id":"p-tel-15","repo":"platform","face":"fr","query":"registry 注册两 format + host-fs-handler 透传新字段与 totalUsage 平台 messages schema 可选新字段 + openapi/gen:types","expected":["fr/host-fs-handler.md"],"subset":"approx"}
+{"id":"p-dec-01","repo":"platform","face":"decisions","query":"backend 落选 方案 B 扩展 CLI 向量端点至全语料 放开 JWT","expected":["decisions/backend.md#D-backend-009@v1"]}
+{"id":"p-dec-02","repo":"platform","face":"decisions","query":"backend 落选 方案 C 基于 DB 表混合检索 ScanDocument 新增 knowledge 入库","expected":["decisions/backend.md#D-backend-010@v1"]}
+{"id":"p-dec-03","repo":"platform","face":"decisions","query":"backend 落选 本地文件缓存 向量序列化 runtime 缓存文件","expected":["decisions/backend.md#D-backend-020@v1"]}
+{"id":"p-dec-04","repo":"platform","face":"decisions","query":"backend 落选 仅做启动预热 不改存储","expected":["decisions/backend.md#D-backend-021@v1"]}
+{"id":"p-dec-05","repo":"platform","face":"decisions","query":"backend 落选 方案 B 后台常驻同步协程","expected":["decisions/backend.md#D-backend-023@v1"]}
+{"id":"p-dec-06","repo":"platform","face":"decisions","query":"backend 落选 平台行携带记录名 方案 B","expected":["decisions/backend.md#D-backend-033@v1"]}
+{"id":"p-dec-07","repo":"platform","face":"decisions","query":"backend 落选 新增 daemon 列记录 RPC 方案 C","expected":["decisions/backend.md#D-sillyhub-daemon-001@v1"]}
+{"id":"p-dec-08","repo":"platform","face":"decisions","query":"backend 落选 方案 C 首查触发 周期刷新混合","expected":["decisions/backend.md#D-backend-024@v1"]}
+{"id":"p-scope-01","repo":"platform","face":"scope","query":"frontend SSE fetch-sse Authorization header","expected":["conventions.md#前端-sse-消费统一-fetch-ssetoken-走-authorization-header禁用-eventsource"],"scope_files":["frontend/src/lib/fetch-sse.ts"],"subset":"added"}
+{"id":"p-scope-02","repo":"platform","face":"scope","query":"backend auth_deps monkeypatch has_permission","expected":["testing-gotchas.md#后端monkeypatch-has_permission-必须双-patchauth_deps-导入时绑定--rbac-模块属性"],"scope_files":["backend/app/core/auth_deps.py"],"subset":"added"}
+{"id":"p-scope-03","repo":"platform","face":"scope","query":"backend audit hooks lifespan 审计日志","expected":["known-issues.md#-audit_hooks-只在测试-lifespan-注册生产审计要业务代码显式写-auditlog"],"scope_files":["backend/app/main.py"],"subset":"added"}
+{"id":"s-scope-01","repo":"sillyspec","face":"scope","query":"cli flag 白名单死路报错","expected":["known-issues.md#-cli-flag-白名单死路报错指引指向未登记-flag双实例"],"scope_files":["test/flag-contract.test.mjs"],"subset":"added"}
+{"id":"s-scope-02","repo":"sillyspec","face":"scope","query":"知识匹配引擎 枚举开放世界 否决","expected":["decisions/core-engine.md#D-unmapped-197@v1"],"scope_files":["src/knowledge-match.js"],"subset":"added"}
+{"id":"s-scope-03","repo":"sillyspec","face":"scope","query":"flow 前门 决策记录","expected":["decisions/cli-entry.md#D-unmapped-433@v1"],"scope_files":["src/flow.js"],"subset":"added"}
+{"id":"s-scope-04","repo":"sillyspec","face":"scope","query":"execute 阶段 风险死路","expected":["decisions/stages.md#D-unmapped-153@v1"],"scope_files":["src/stages/execute.js"],"subset":"added"}
+{"id":"s-scope-05","repo":"sillyspec","face":"scope","query":"CLI 入口 决策","expected":["decisions/cli-entry.md#D-unmapped-161@v1"],"scope_files":["src/index.js"],"subset":"added"}
+{"id":"p-scope-04","repo":"platform","face":"scope","query":"frontend SSE agent-stream 流式消费","expected":["conventions.md#前端-sse-消费统一-fetch-ssetoken-走-authorization-header禁用-eventsource"],"scope_files":["frontend/src/lib/agent-stream.ts"],"subset":"added"}
+{"id":"p-scope-05","repo":"platform","face":"scope","query":"login_enabled get_current_user 检查","expected":["known-issues.md#-2026-06-17--login_enabled-必须在-get_current_user-检查不能只在-login-入口commit-d62ec975"],"scope_files":["backend/app/core/auth_deps.py"],"subset":"added"}
+{"id":"p-scope-06","repo":"platform","face":"scope","query":"tailwind md 视口断点 侧栏","expected":["conventions.md#tailwind-md-是视口断点非容器断点侧栏内嵌组件禁用响应式前缀"],"scope_files":["frontend/src/components/changes/detail/change-sessions-card.tsx"],"subset":"added"}
+{"id":"s-fusion-01","repo":"sillyspec","face":"entries","query":"worktree 清理流程说明","expected":["known-issues.md#-cli-flag-白名单死路报错指引指向未登记-flag双实例"],"scope_files":["test/flag-contract.test.mjs"],"subset":"fusion","note":"路由命中 worktree 相关但手工条目仅 scope doc-refs 可达（融合增量）"}
+{"id":"s-fusion-02","repo":"sillyspec","face":"decisions","query":"cli flag 白名单 死路 报错","expected":["decisions/core-engine.md#D-unmapped-197@v1"],"scope_files":["src/knowledge-match.js"],"subset":"fusion","note":"路由命中 known-issues 但 rejected 决策仅 scope 两跳可达（融合增量）"}
+// organic 评测集 — 遥测 query 抽样 + 人工 anchor 级标签（冻结：提交后不随遥测漂移）
+// 字段：id / repo(sillyspec|platform) / face(entries|decisions|search|scope) / query / expected[]
+// expected 含 "#" = 精确 file#anchor；不含 "#" = 文件级（匹配该文件任意命中）
+// scope_files（可选）= scope 面图遍历的机器键；无此字段的 case 不进 scope 面
+// 标注口径：查询的真实技术意图应对口的知识条目；当前未命中即为红基线，供后续变更对比
+
+{"id":"s-fr-prod-01","repo":"sillyspec","face":"fr","query":"worktree 工作树 清理 隔离","scope_files":["src/worktree.js"],"expected":["fr/worktree.md"],"subset":"prod","note":"生产同源：src/worktree.js → worktree 域"}
+{"id":"p-fr-prod-01","repo":"platform","face":"fr","query":"SSE 流式消费 token authorization","scope_files":["frontend/src/lib/fetch-sse.ts"],"expected":["fr/frontend.md"],"subset":"prod","note":"生产同源：fetch-sse.ts → frontend 域"}
+{"id":"auto-sillyspec-01","repo":"sillyspec","face":"decisions","query":"cli-entry 收口机器门拦 decisions.md 缺席","expected":["decisions/cli-entry.md#D-unmapped-433@v1"]}
+{"id":"auto-sillyspec-02","repo":"sillyspec","face":"decisions","query":"core-engine 否决入库侧路由词表派生 同窗词片按序截断=抽签","expected":["decisions/core-engine.md#D-unmapped-197@v1"]}
+{"id":"auto-sillyspec-03","repo":"sillyspec","face":"decisions","query":"core-engine 错位目录归一否决「marker 值改为完整目录名」","expected":["decisions/core-engine.md#D-unmapped-375@v1"]}
+{"id":"auto-sillyspec-04","repo":"sillyspec","face":"decisions","query":"core-engine archive 复合真相源方案否决","expected":["decisions/core-engine.md#D-unmapped-377@v1"]}
+{"id":"auto-sillyspec-05","repo":"sillyspec","face":"decisions","query":"docs 收割器机器自动拆 rejected 条目","expected":["decisions/docs.md#D-unmapped-299@v1"]}
+{"id":"auto-sillyspec-06","repo":"sillyspec","face":"decisions","query":"docs distill 侧从 confirmed","expected":["decisions/docs.md#D-unmapped-300@v1"]}
+{"id":"auto-sillyspec-07","repo":"sillyspec","face":"decisions","query":"docs 只修厚道不动轻量道","expected":["decisions/docs.md#D-unmapped-301@v1"]}
+{"id":"auto-sillyspec-08","repo":"sillyspec","face":"decisions","query":"docs 存量 40+ 种散值","expected":["decisions/docs.md#D-unmapped-436@v1"]}
+{"id":"auto-platform-01","repo":"platform","face":"decisions","query":"app-admin-pages 落选 并入设置页 Tab（方案","expected":["decisions/app-admin-pages.md#D-unmapped-004@v1"]}
+{"id":"auto-platform-02","repo":"platform","face":"decisions","query":"app-admin-pages 落选 工作区首页挂卡片（方案 C）","expected":["decisions/app-admin-pages.md#D-unmapped-005@v1"]}
+{"id":"auto-platform-03","repo":"platform","face":"decisions","query":"backend 落选 方案 B（扩展","expected":["decisions/backend.md#D-backend-009@v1"]}
+{"id":"auto-platform-04","repo":"platform","face":"decisions","query":"backend 落选 方案 C（基于","expected":["decisions/backend.md#D-backend-010@v1"]}
+{"id":"auto-platform-05","repo":"platform","face":"decisions","query":"backend 落选 本地文件缓存（向量序列化到 .runtime","expected":["decisions/backend.md#D-backend-020@v1"]}
+{"id":"auto-platform-06","repo":"platform","face":"decisions","query":"backend 落选 仅做启动预热（不改存储）","expected":["decisions/backend.md#D-backend-021@v1"]}
+{"id":"auto-platform-07","repo":"platform","face":"decisions","query":"backend 落选 方案 B（后台常驻同步协程）","expected":["decisions/backend.md#D-backend-023@v1"]}
+{"id":"auto-platform-08","repo":"platform","face":"decisions","query":"backend 落选 方案 C（首查触发","expected":["decisions/backend.md#D-backend-024@v1"]}
+{"id":"auto-platform-09","repo":"platform","face":"decisions","query":"backend 落选 平台行携带记录名（方案 B）","expected":["decisions/backend.md#D-backend-033@v1"]}
+{"id":"auto-platform-10","repo":"platform","face":"decisions","query":"backend 落选 新增 daemon「列记录」RPC（方案","expected":["decisions/backend.md#D-sillyhub-daemon-001@v1"]}
+{"id":"auto-platform-11","repo":"platform","face":"decisions","query":"backend 落选 方案 C（进程内卸载","expected":["decisions/backend.md#D-unmapped-057@v1"]}
+{"id":"auto-platform-12","repo":"platform","face":"decisions","query":"backend 方案 B（只加大宽限窗 不加族系列）","expected":["decisions/backend.md#D-unmapped-065@v1"]}

package/eval/knowledge/fixtures/index.mjs

@@ -0,0 +1,132 @@
+/**
+ * eval/knowledge/fixtures/index.mjs — 六类缺陷探针 fixture（合成最小知识库 + 正反例断言）
+ *
+ * 每个 fixture 的 files 相对 .sillyspec/knowledge/ 物化到 mkdtemp 仓根后跑真实引擎。
+ * 断言语义（evaluateCase）：contains 必须全部命中、excludes 必须全部不命中、empty 指定空性、
+ * score_contract 要求 search 视图 score === matchedTags.length、health 逐键对账 catalog。
+ * 探针的 pass 语义是「修复后的期望」：缺陷存在时 pass=false（executor 检出缺陷），修复后转 true。
+ */
+export const PROBE_FIXTURES = [
+  {
+    defect: 'score-contract',
+    files: {
+      'INDEX.md': '# Knowledge Index\n\n## Handbook\n- alpha|beta|gamma|delta → [Alpha 手册](handbook.md#alpha)\n',
+      'handbook.md': '# Handbook\n\n## Alpha\n\nalpha 的说明文字。\n',
+    },
+    cases: [
+      {
+        name: 'search score 必须等于实际命中标签数',
+        kind: 'search',
+        query: 'alpha 用法',
+        expect: { score_contract: true, contains: ['handbook.md'] },
+      },
+    ],
+  },
+  {
+    defect: 'word-boundary',
+    files: {
+      'INDEX.md': [
+        '# Knowledge Index', '',
+        '## Handbook',
+        '- node → [Node 环境](env.md#node)',
+        '- worktree → [Worktree](wt.md#wt)', '',
+      ].join('\n'),
+      'env.md': '# 环境\n\n## node\n\nnode 版本要求说明。\n',
+      'wt.md': '# Worktree\n\n## wt\n\nworktree 隔离说明。\n',
+    },
+    cases: [
+      { name: 'node 不得命中 node_modules（_ 是词字符）', kind: 'entries', query: '修复 node_modules junction', expect: { excludes: ['env.md'] } },
+      { name: 'node 独立词仍命中', kind: 'entries', query: 'node 版本升级', expect: { contains: ['env.md'] } },
+      { name: 'worktree 不得命中 worktree_apply（_ 是词字符）', kind: 'entries', query: 'worktree_apply 冲突', expect: { excludes: ['wt.md'] } },
+    ],
+  },
+  {
+    defect: 'route-rows',
+    files: {
+      'INDEX.md': [
+        '# Knowledge Index', '',
+        '## Manual',
+        '- alpha|beta → [Alpha 手册](handbook.md#alpha)', '',
+        '## Decisions',
+        '- alpha|beta → [决策域](decisions/demo.md)', '',
+        '## FR',
+        '- alpha|beta → [FR 域](fr/demo.md)', '',
+      ].join('\n'),
+      'handbook.md': '# Handbook\n\n## Alpha\n\nalpha 手册内容。\n',
+      'decisions/demo.md': '## D-001@v1 alpha 决策\n状态：implemented\n变更：fix-demo\n',
+      'fr/demo.md': '## FR-demo-001 alpha 需求\n变更：fix-demo\n状态：active\n',
+    },
+    cases: [
+      {
+        name: 'entries 不得含 fr/decisions 路由行',
+        kind: 'entries',
+        query: 'alpha 功能',
+        expect: { contains: ['handbook.md#alpha'], excludes: ['decisions/demo.md', 'fr/demo.md'] },
+      },
+      { name: 'decisionHits 通道不受影响', kind: 'decisions', query: 'alpha 功能', expect: { contains: ['decisions/demo.md#D-001@v1'] } },
+    ],
+  },
+  {
+    defect: 'scope-manual',
+    files: {
+      'INDEX.md': [
+        '# Knowledge Index', '',
+        '## Decisions',
+        '- foo → [决策域](decisions/demo.md)', '',
+        '## Gotchas',
+        '- foo|junction → [Foo 坑](gotchas.md#foo-junction)',
+        '- other → [Other 区](gotchas.md#other-section)', '',
+      ].join('\n'),
+      'gotchas.md': [
+        '# Gotchas', '',
+        '## foo junction 清理', '',
+        '修 src/foo.js 时先解链 node_modules junction。', '',
+        '## other section', '',
+        '另一个模块维护 src/bar.js 的说明。', '',
+      ].join('\n'),
+      'decisions/demo.md': '## D-001@v1 foo 决策\n状态：implemented\n锚点：src/foo.js:10\n变更：fix-demo\n',
+    },
+    cases: [
+      { name: '决策锚点强边可达（基础链路）', kind: 'scope', scope_files: ['src/foo.js'], expect: { contains: ['decisions/demo.md#D-001@v1'] } },
+      { name: '手工条目正文路径可达（手工知识不短路）', kind: 'scope', scope_files: ['src/foo.js'], expect: { contains: ['gotchas.md#foo-junction'] } },
+      { name: 'entries-only 正例（无决策锚点仍返回）', kind: 'scope', scope_files: ['src/bar.js'], expect: { contains: ['gotchas.md#other-section'] } },
+      { name: 'section 隔离负例（异路径不串节）', kind: 'scope', scope_files: ['src/bar.js'], expect: { excludes: ['gotchas.md#foo-junction'] } },
+    ],
+  },
+  {
+    defect: 'fixed-entries',
+    files: {
+      'INDEX.md': [
+        '# Knowledge Index', '',
+        '## Known Issues',
+        '- alpha → [alpha 已修复](known-issues.md#alpha-已修复)',
+        '- beta → [beta 坑](known-issues.md#beta-坑)', '',
+      ].join('\n'),
+      'known-issues.md': [
+        '# Known Issues', '',
+        '## 🟢 alpha 问题已修复（commit abc）', '',
+        '历史坑，已修复。', '',
+        '## 🔴 beta 问题', '',
+        '当前仍然存在的坑。', '',
+      ].join('\n'),
+    },
+    cases: [
+      { name: '已修复条目不进默认 entries', kind: 'entries', query: 'alpha 问题', expect: { excludes: ['known-issues.md#alpha-已修复'] } },
+      { name: '未修复条目正常命中', kind: 'entries', query: 'beta 问题', expect: { contains: ['known-issues.md#beta-坑'] } },
+    ],
+  },
+  {
+    defect: 'linkage',
+    files: {
+      'INDEX.md': '# Knowledge Index\n\n## Decisions\n- demo|linkage → [决策域](decisions/demo.md)\n',
+      'decisions/demo.md': [
+        '## D-001@v1 有锚点', '状态：implemented', '锚点：src/foo.js:10', '',
+        '## D-002@v1 有文件字段', '状态：implemented', '文件：src/bar.js', '',
+        '## D-003@v1 无关联', '状态：implemented', '理由：纯决策记录', '',
+      ].join('\n'),
+    },
+    cases: [
+      { name: '语料健康报告关联率', kind: 'catalog', expect: { health: { total: 3, anchored: 1, files: 1, linked: 2 } } },
+    ],
+  },
+]

package/eval/knowledge/lib/dataset.mjs

@@ -0,0 +1,171 @@
+/**
+ * eval/knowledge/lib/dataset.mjs — 三源数据集加载/校验/现场构建
+ *
+ * organic：提交冻结的 JSONL（遥测 query 抽样 + 人工 anchor 级标签），加载即校验。
+ * self：现场从归档 flow-state.input 构建，期望集 = 该变更蒸馏出的决策（按 变更+id/原编号 映射），
+ *       仅用于「蒸馏→检索链路」smoke，禁止进质量指标（自产自测偏差，design 已论证）。
+ * probes：fixtures 目录枚举（每个子目录一个缺陷探针，runner 负责执行）。
+ */
+import { existsSync, readFileSync, readdirSync } from 'node:fs'
+import { join } from 'node:path'
+import { parseDecisionEntries, normalizeAnchorForCompare } from '../../../src/knowledge-match.js'
+import { readFlowState } from '../../../src/flow-state.js'
+
+const FACES = new Set(['entries', 'decisions', 'search', 'scope', 'fr'])
+
+/** organic case 字段校验；返回 null = 合法，否则错误原因。
+ *  repo 是数据集标签（对应 --repo 的 label 或路径 basename），非项目枚举——任意仓可用。 */
+export function validateOrganicCase(c) {
+  if (!c || typeof c !== 'object' || Array.isArray(c)) return 'not-object'
+  if (!c.id || typeof c.id !== 'string') return 'id'
+  if (!c.repo || typeof c.repo !== 'string') return 'repo'
+  if (!FACES.has(c.face)) return 'face'
+  if (!c.query || typeof c.query !== 'string' || !c.query.trim()) return 'query'
+  if (!Array.isArray(c.expected) || c.expected.length === 0) return 'expected-empty'
+  if (c.expected.some((e) => typeof e !== 'string' || !e.trim())) return 'expected-item'
+  if (c.scope_files !== undefined) {
+    if (!Array.isArray(c.scope_files) || c.scope_files.length === 0) return 'scope_files'
+    if (c.scope_files.some((f) => typeof f !== 'string' || !f.trim())) return 'scope_files-item'
+  }
+  return null
+}
+
+/**
+ * 加载 organic JSONL：空行/`//` 注释跳过；坏 JSON、字段缺失、expected 空 → skipped（禁止静默当零命中）。
+ * @returns {{cases: Array, skipped: Array<{line:number, reason:string, id?:string}>, missing: boolean}}
+ */
+export function loadOrganic(filePath) {
+  const cases = []
+  const skipped = []
+  if (!existsSync(filePath)) return { cases, skipped, missing: true }
+  const lines = readFileSync(filePath, 'utf8').split('\n')
+  lines.forEach((raw, i) => {
+    const line = raw.trim()
+    if (!line || line.startsWith('//')) return
+    let c
+    try { c = JSON.parse(line) } catch { skipped.push({ line: i + 1, reason: 'bad-json' }); return }
+    const err = validateOrganicCase(c)
+    if (err) { skipped.push({ line: i + 1, reason: err, id: c && c.id }); return }
+    cases.push({ source: 'organic', ...c })
+  })
+  return { cases, skipped, missing: false }
+}
+
+/**
+ * 现场构建 self 集：归档 flow-state.input → 该变更蒸馏决策的映射（变更+id 或 变更+原编号）。
+ * 映射不到任何决策的归档 → skipped（无期望集可测，不造 case）。
+ * @returns {{cases: Array, skipped: Array<{archive:string, reason:string}>}}
+ */
+export function buildSelfCases(repoRoot, repo) {
+  const archive = join(repoRoot, '.sillyspec', 'changes', 'archive')
+  const knowledgeDir = join(repoRoot, '.sillyspec', 'knowledge')
+  const cases = []
+  const skipped = []
+  if (!existsSync(archive)) return { cases, skipped }
+  const decisions = parseDecisionEntries(knowledgeDir)
+  for (const name of readdirSync(archive).sort()) {
+    const dir = join(archive, name)
+    const st = readFlowState(dir)
+    if (!st || !st.input || typeof st.input !== 'string' || !st.input.trim()) {
+      skipped.push({ archive: name, reason: 'no-input' })
+      continue
+    }
+    const dmd = join(dir, 'decisions.md')
+    if (!existsSync(dmd)) { skipped.push({ archive: name, reason: 'no-decisions-md' }); continue }
+    const localIds = [...readFileSync(dmd, 'utf8').matchAll(/^##\s+(D-(?:[A-Za-z0-9_-]+-)?\d+@v\d+)/gm)].map((m) => m[1])
+    const own = []
+    for (const id of localIds) {
+      const hit = decisions.find((h) => h.change === name && (h.id === id || h.originId === id))
+      if (hit) own.push(`${hit.file}#${hit.id}`)
+    }
+    if (own.length === 0) { skipped.push({ archive: name, reason: 'no-mapped-decision' }); continue }
+    cases.push({
+      source: 'self', id: `self-${name}`, repo, face: 'self',
+      query: st.input.trim(), expected: [...new Set(own)], own: [...new Set(own)],
+    })
+  }
+  return { cases, skipped }
+}
+
+/**
+ * 遥测 inject 事件 → organic 候选池（去重 change+query，单变更限流）。
+ * @returns {Array<{change:string, query:string}>}
+ */
+export function extractTelemetryPool(telemetryPath, { limitPerChange = 3 } = {}) {
+  if (!existsSync(telemetryPath)) return []
+  const seen = new Set()
+  const byChange = new Map()
+  for (const line of readFileSync(telemetryPath, 'utf8').split('\n')) {
+    const t = line.trim()
+    if (!t) continue
+    let r
+    try { r = JSON.parse(t) } catch { continue }
+    if (!r || r.type !== 'inject' || typeof r.query !== 'string' || typeof r.change !== 'string') continue
+    const q = r.query.replace(/\r\n/g, '\n').trim()
+    if (!q) continue
+    const key = `${r.change}\u0000${q}`
+    if (seen.has(key)) continue
+    seen.add(key)
+    const arr = byChange.get(r.change) || []
+    if (arr.length < limitPerChange) arr.push(q)
+    byChange.set(r.change, arr)
+  }
+  const out = []
+  for (const [change, queries] of byChange) for (const q of queries) out.push({ change, query: q })
+  // 稳定排序（不依赖遥测事件到达顺序——乱序/迟到无关，变量见 design 盲维①）
+  out.sort((a, b) => a.change.localeCompare(b.change) || a.query.localeCompare(b.query))
+  return out
+}
+
+/** 枚举探针 fixture 目录（每子目录一个；无目录 = 空数组）。 */
+export function listProbes(fixturesDir) {
+  try {
+    return readdirSync(fixturesDir, { withFileTypes: true })
+      .filter((d) => d.isDirectory())
+      .map((d) => d.name)
+      .sort()
+  } catch {
+    return []
+  }
+}
+
+/**
+ * 标签有效性自检（2026-11-02 复核 P1-①）：expected 的 file 必须存在于目标仓 knowledge 下、
+ * 含 # 时 anchor 必须能以归一化形态在目标文件正文中找到（不再回退 INDEX——防止「anchor 只在
+ * INDEX、heading 已漂移」的旧契约通过自检）；scope_files 必须是目标仓真实存在的文件。
+ * reject 结构性不可命中标签（如 epoch 迁移后 D-id 换了域文件）。
+ * @param {Array} cases organic cases
+ * @param {Map<string,string>} repoRootByTag 仓标签 → 仓根
+ * @returns {Array<{id:string, expected?:string, reason:string}>}
+ */
+export function validateLabels(cases, repoRootByTag) {
+  const issues = []
+  for (const c of cases) {
+    const root = repoRootByTag.get(c.repo)
+    if (!root) continue // 未参与本次运行的仓标签不算标签错误
+    const kdir = join(root, '.sillyspec', 'knowledge')
+    for (const sf of c.scope_files || []) {
+      const p = join(root, String(sf).replace(/\\/g, '/'))
+      if (!existsSync(p)) issues.push({ id: c.id, expected: sf, reason: 'scope-file-missing' })
+    }
+    for (const exp of c.expected) {
+      const [file, anchor] = String(exp).split('#')
+      const fp = join(kdir, file)
+      if (!existsSync(fp)) {
+        issues.push({ id: c.id, expected: exp, reason: 'file-missing' })
+        continue
+      }
+      if (!anchor) continue
+      let text = ''
+      try { text = readFileSync(fp, 'utf8') } catch {
+        issues.push({ id: c.id, expected: exp, reason: 'file-unreadable' })
+        continue
+      }
+      const normAnchor = normalizeAnchorForCompare(anchor)
+      if (!normAnchor || !normalizeAnchorForCompare(text).includes(normAnchor)) {
+        issues.push({ id: c.id, expected: exp, reason: 'anchor-heading-missing', file })
+      }
+    }
+  }
+  return issues
+}

package/eval/knowledge/lib/engine.mjs

@@ -0,0 +1,136 @@
+/**
+ * eval/knowledge/lib/engine.mjs — 只读适配生产检索引擎（评测面）
+ *
+ * 统一命中键：entries/search/scope-entry = `file#anchor`（anchor 缺省为 file）；
+ * decisions/scope-decision = `file#id`；fr = `fr/<域>.md#<FR-id>`。
+ * 期望侧不含 "#" 时按文件级前缀匹配（runner 展开）。
+ *
+ * 三个检索通道与生产同源：
+ * - entries/decisions 走 matchKnowledgeHybrid（路由→向量→词片回退，生产 flow/prompt/execute 同入口）；
+ * - search 走 matchKnowledgeHybrid + buildSearchMatches（cmdSearch 生产构造单一源）；
+ * - fr 走 readActiveFrDigest（FR 索引 digest 生产通道）+ 标题重叠排序。
+ * 零网络零遥测：向量层由调用方 env 关闭；只调纯读函数。图构建缓存写 <specRoot>/.runtime/cache
+ * （可再生派生，fixture 场景落 mkdtemp，真实仓仅缓存，不触测试/遥测面）。
+ */
+import { join } from 'node:path'
+import { existsSync, readdirSync, readFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { parseDecisionEntries, keywordMatchesContext } from '../../../src/knowledge-match.js'
+import { buildKnowledgeGraph, scopeRecall } from '../../../src/knowledge-graph.js'
+import { matchKnowledgeHybrid } from '../../../src/knowledge-vector.js'
+import { buildSearchMatches } from '../../../src/stages/knowledge.js'
+import { readActiveFrDigest, frTitleOverlap, resolveTouchedDomains, rankFrDigestForInjection } from '../../../src/fr-index.js'
+import { parseModuleMapSimple } from '../../../src/modules.js'
+
+const graphCache = new Map()
+
+function specRootOf(repoRoot) { return join(repoRoot, '.sillyspec') }
+function knowledgeDirOf(repoRoot) { return join(specRootOf(repoRoot), 'knowledge') }
+
+function entryKey(e) { return e.anchor ? `${e.file}#${e.anchor}` : e.file }
+
+/** entries 面：生产 hybrid 入口的 entries（只手工条目；路由行已由变更 1 移入 routeHits）。
+ *  opts.scopeFiles（仅 fusion 子集传）：融合增量 scopeEntries 追加在路由 entries 之后。 */
+export async function runEntries(repoRoot, query, opts = {}) {
+  const r = await matchKnowledgeHybrid(knowledgeDirOf(repoRoot), query, { cwd: repoRoot, scopeFiles: opts.scopeFiles || [] })
+  const hits = [...(r.entries || []).map(entryKey), ...(r.scopeEntries || []).map(entryKey)]
+  return { matched: r.matched, hits, raw: r.entries || [], routeHits: r.routeHits || [], scopeEntries: r.scopeEntries || [] }
+}
+
+/** decisions 面：生产 hybrid 入口的 decisionHits（rejected/死路优先序）。
+ *  opts.scopeFiles（仅 fusion 子集传）：融合增量 scopeDecisionHits 追加在路由 decisionHits 之后。 */
+export async function runDecisions(repoRoot, query, opts = {}) {
+  const r = await matchKnowledgeHybrid(knowledgeDirOf(repoRoot), query, { cwd: repoRoot, scopeFiles: opts.scopeFiles || [] })
+  const hits = [
+    ...(r.decisionHits || []).map((h) => `${h.file}#${h.id}`),
+    ...(r.scopeDecisionHits || []).map((h) => `${h.file}#${h.id}`),
+  ]
+  return { hits, raw: r.decisionHits || [], scopeDecisionHits: r.scopeDecisionHits || [] }
+}
+
+/** 合并 docs/<项目>/modules/_module-map.yaml 模块索引（生产触达域解析输入）。 */
+function moduleIndexFor(specRoot) {
+  const out = {}
+  try {
+    for (const p of readdirSync(join(specRoot, 'docs'), { withFileTypes: true })) {
+      if (!p.isDirectory()) continue
+      const f = join(specRoot, 'docs', p.name, 'modules', '_module-map.yaml')
+      if (!existsSync(f)) continue
+      Object.assign(out, parseModuleMapSimple(readFileSync(f, 'utf8')))
+    }
+  } catch { /* 无 docs = 空索引 */ }
+  return out
+}
+
+/** fr 面：带 scope_files 的 case 走生产同源通道（resolveTouchedDomains 触达域 → 域内
+ *  readActiveFrDigest → rankFrDigestForInjection TierA/TierB）；纯文本查询无机器键，走近似面
+ *  （全域 digest + 标题重叠，channel=approx）并单列、不进对比门槛。 */
+export function runFr(repoRoot, query, opts = {}) {
+  const specRoot = specRootOf(repoRoot)
+  const kdir = knowledgeDirOf(repoRoot)
+  const scopeFiles = Array.isArray(opts.scopeFiles) ? opts.scopeFiles.filter(Boolean) : []
+  if (scopeFiles.length > 0) {
+    try {
+      const domains = [...resolveTouchedDomains(tmpdir(), moduleIndexFor(specRoot), scopeFiles, kdir)]
+      const frs = readActiveFrDigest(kdir, domains)
+      const ranked = rankFrDigestForInjection({ archiveRoot: join(specRoot, 'changes', 'archive'), frs, changed: scopeFiles })
+      const ordered = ranked.ranked || []
+      return { hits: ordered.map((f) => `fr/${f.domain}.md#${f.id}`), raw: ordered, channel: 'prod' }
+    } catch { /* 生产路由失败 → 近似面兜底（如实标 channel） */ }
+  }
+  let domains = []
+  try {
+    domains = readdirSync(join(kdir, 'fr')).filter((f) => f.endsWith('.md')).map((f) => f.replace(/\.md$/, ''))
+  } catch { /* 无 fr 目录 = 空面 */ }
+  const entries = readActiveFrDigest(kdir, domains)
+  const q = String(query || '')
+  const scored = entries.map((e) => ({ e, s: frTitleOverlap(q, `${e.id} ${e.title || ''}`) }))
+  scored.sort((a, b) => (b.s - a.s) || a.e.id.localeCompare(b.e.id))
+  return { hits: scored.map((x) => `fr/${x.e.domain}.md#${x.e.id}`), raw: scored.map((x) => x.e), channel: 'approx' }
+}
+
+/**
+ * search 面（cmdSearch 等价）：真实 hybrid 检索 + 生产构造 buildSearchMatches。
+ * limit 缺省 Infinity = 全量（raw 口径）；rendered 由 faces 层按消费面 limit 截断。
+ * 返回行带 matchedTags（与 score 契约对照用）。
+ */
+export async function searchView(repoRoot, query, limit = Infinity) {
+  const knowledgeDir = knowledgeDirOf(repoRoot)
+  const result = await matchKnowledgeHybrid(knowledgeDir, query, { cwd: repoRoot })
+  if (!result || !result.matched) return []
+  const ctx = String(query || '').toLowerCase()
+  return buildSearchMatches(result, knowledgeDir, limit, query).map((row, i) => ({
+    ...row,
+    file: result.entries[i] ? result.entries[i].file : row.path.replace(/^knowledge[\\/]/, ''),
+    matchedTags: (row.tags || []).filter((k) => keywordMatchesContext(k, ctx)),
+  }))
+}
+
+/** scope 面：图遍历召回（引擎内建 cap 保持——entry 3 + decision 20，与生产 scopeRecall 一致）。 */
+export function runScope(repoRoot, scopeFiles) {
+  const root = specRootOf(repoRoot)
+  if (!graphCache.has(root)) graphCache.set(root, buildKnowledgeGraph(root))
+  const recall = scopeRecall(graphCache.get(root), scopeFiles || [])
+  if (!recall) return { entryHits: [], decisionHits: [], merged: [] }
+  const entryHits = recall.entries.map(entryKey)
+  const decisionHits = recall.decisionHits.map((h) => `${h.file}#${h.id}`)
+  return { entryHits, decisionHits, merged: [...entryHits, ...decisionHits] }
+}
+
+/** 语料健康：决策条目总数 / 有锚点 / 有文件字段 / 任一关联（linkage 探针与 baseline 用）。 */
+export function catalogHealth(repoRoot) {
+  const kdir = knowledgeDirOf(repoRoot)
+  if (!existsSync(join(kdir, 'INDEX.md'))) return { total: 0, anchored: 0, files: 0, linked: 0 }
+  const hits = parseDecisionEntries(kdir)
+  let anchored = 0
+  let files = 0
+  let linked = 0
+  for (const h of hits) {
+    const hasAnchor = Boolean(h.anchor && h.anchor.trim() && h.anchor.trim() !== '未记录')
+    const hasFiles = Array.isArray(h.files) && h.files.length > 0
+    if (hasAnchor) anchored++
+    if (hasFiles) files++
+    if (hasAnchor || hasFiles) linked++
+  }
+  return { total: hits.length, anchored, files, linked }
+}

package/eval/knowledge/lib/faces.mjs

@@ -0,0 +1,88 @@
+/**
+ * eval/knowledge/lib/faces.mjs — 分面执行（raw/rendered 双口径 + 截断归因）
+ *
+ * raw = 引擎全量命中序（不预先截断）；rendered = 各消费面实际切片（flow 📚 行 top-3、
+ * rejected top-5、search limit 10、scope entry 3 + decision 20）。raw 命中而 rendered 不中
+ * → truncated；每 case 的 raw_rank/rendered_rank 进 details（FR-03 的 rank 记录面）。
+ */
+import { runEntries, runDecisions, runFr, searchView, runScope } from './engine.mjs'
+import { hitMatches } from './probes.mjs'
+import { summarize } from './metrics.mjs'
+
+export const FACES = ['entries', 'decisions', 'search', 'scope', 'fr']
+
+/** 期望谓词排名（file 级前缀语义；1-based，未命中 = Infinity）。 */
+export function rankWithMatches(hits, expected) {
+  let best = Infinity
+  hits.forEach((h, i) => {
+    if (expected.some((e) => hitMatches(h, e))) best = Math.min(best, i + 1)
+  })
+  return best
+}
+
+/**
+ * scope 面按 subset 分组指标（frozen/added——FR-05 子集单列，禁止用新增 case 掩盖冻结子集）。
+ * reachable = rank 非 null 比例：融合子集（2026-11-02-knowledge-fusion）的价值是「从不可达变可达」，
+ * 在「不重排路由」契约下期望可能落在路由尾部（top3=0 但 reachable=1）——两者都要如实可见。
+ */
+export function subsetSummaries(faceResult, cases) {
+  const groups = {}
+  for (const c of cases) {
+    const d = faceResult.details.find((x) => x.id === c.id)
+    if (!d) continue
+    const key = c.subset || 'default'
+    ;(groups[key] = groups[key] || []).push(d.raw_rank === null ? Infinity : d.raw_rank)
+  }
+  const r4 = (x) => Math.round(x * 10000) / 10000
+  return Object.fromEntries(Object.entries(groups).map(([k, ranks]) => [
+    k,
+    { ...summarize(ranks), reachable: ranks.length > 0 ? r4(ranks.filter((r) => Number.isFinite(r)).length / ranks.length) : 0 },
+  ]))
+}
+
+/** 单仓单面：raw/rendered 双口径 + 截断清单 + 每 case 名次明细。 */
+export async function runFace(repoPath, face, cases) {
+  const ranksRaw = []
+  const ranksRendered = []
+  const truncated = []
+  const details = []
+  for (const c of cases) {
+    let rawHits = []
+    let renderedHits = []
+    if (face === 'entries') {
+      const r = await runEntries(repoPath, c.query, { scopeFiles: c.subset === 'fusion' ? (c.scope_files || []) : [] })
+      rawHits = r.hits
+      renderedHits = r.hits.slice(0, 3)
+    } else if (face === 'decisions') {
+      const r = await runDecisions(repoPath, c.query, { scopeFiles: c.subset === 'fusion' ? (c.scope_files || []) : [] })
+      rawHits = r.hits
+      renderedHits = r.hits.slice(0, 5)
+    } else if (face === 'fr') {
+      const r = runFr(repoPath, c.query, { scopeFiles: c.scope_files || [] })
+      rawHits = r.hits
+      renderedHits = r.hits.slice(0, 8)
+    } else if (face === 'search') {
+      rawHits = (await searchView(repoPath, c.query, Infinity)).map((r) => r.file)
+      renderedHits = rawHits.slice(0, 10)
+    } else if (face === 'scope') {
+      const r = runScope(repoPath, c.scope_files || [])
+      rawHits = r.merged
+      renderedHits = [...r.entryHits.slice(0, 3), ...r.decisionHits.slice(0, 20)]
+    }
+    // search 面单位是 file：expected 的 anchor 后缀投影掉
+    const expected = face === 'search' ? c.expected.map((e) => String(e).split('#')[0]) : c.expected
+    const rr = rankWithMatches(rawHits, expected)
+    const rn = rankWithMatches(renderedHits, expected)
+    ranksRaw.push(rr)
+    ranksRendered.push(rn)
+    const isTruncated = Number.isFinite(rr) && !Number.isFinite(rn)
+    if (isTruncated) truncated.push(c.id)
+    details.push({
+      id: c.id,
+      raw_rank: Number.isFinite(rr) ? rr : null,
+      rendered_rank: Number.isFinite(rn) ? rn : null,
+      truncated: isTruncated,
+    })
+  }
+  return { raw: summarize(ranksRaw), rendered: summarize(ranksRendered), truncated, details }
+}