治理: Skill 测试治理第一阶段——harness 控制平面 + 实现测试迁出运行时目录

范围（不含 design-story-foundation、docs/、humanization/README.md 等进行中改动）: 1. 新增 harness/ 控制平面 - skill_harness.py 静态审计：32 个运行时 Skill 的 frontmatter/manifest/文档污染，当前 0 问题 - run_selected.py 选择性执行器：manifest 与磁盘一一对账、依赖阻断、空跑与 skip-only 失败关闭、AST 测试形状门 - manifests/skills.json：32 个 Skill 的合同责任方与协作领域登记 - manifests/test-inventory.json：81 个测试资产登记 - specs/skill-testing.md 与 README.md：测试分层、证据边界与 harness 职责 2. 实现测试从 .claude/skills/*/scripts/ 迁至 tests/skills/<skill>/ - 71 个测试文件迁移并修复项目根与临时目录运行导入 - 数据库触发器测试宽泛异常收窄为 psycopg.errors.RaiseException - 抽取离线大测试拆出真实 PG smoke（默认阻断，不计入离线通过） - 抽取 presence 去重边界拆出独立测试：493 + 78 = 571 项检查不变 3. 运行时文档清理 - 13 个 SKILL.md 移除自测/离线验证段落、测试命令与测试文件事实源表述，只保留运行时合同；业务运行合同、额度、授权与离线模式均保留 4. SoT 同步 - AGENTS.md：新增 Skill 领域索引（7 个合同责任方分组，覆盖 32 个运行时 Skill） - 领域 07：测试入口改由 harness/manifests/ 登记，SKILL.md 不承载测试命令 - humanization 覆盖矩阵：活动测试路径同步迁移验证证据: harness 自测 15 项 + runner 自测 13 项通过；静态审计 32 Skill / 0 问题； 73 个非数据库测试通过；8 个集成条目中 6 个 PostgreSQL 项被依赖门明确阻断； py_compile 与 git diff --check 通过。未连接 PostgreSQL、网络、真实模型或额度。已知边界: 真正 skill_behavior_eval 仍为 0，尚未验证任何 Skill 自然语言行为； evaluate-frozen-replay 的 raw 存储边界冲突留待单独治理。
2026-08-19 01:50:20 +08:00 · 2026-08-19 01:50:20 +08:00 · c9f69d9d6d
commit c9f69d9d6d
parent 476bb71da4
97 changed files with 5247 additions and 821 deletions
--- a/.agent/docs/architecture/domains/07-Agent与Skill领域.md
+++ b/.agent/docs/architecture/domains/07-Agent与Skill领域.md
@ -31,7 +31,7 @@ Agent 与 Skill 领域拥有角色职责、可调用能力合同、确定性工
 6. 输入产出落库：哪些输入和产出必须进库可见，见索引 §3。
 7. 稳定错误码和失败恢复。
 8. raw、授权、预算和审计边界。
-9. 离线测试与验收命令。
+9. 开发验证与行为评测由 `harness/manifests/` 登记；`SKILL.md` 不承载测试命令、测试文件路径或测试结论。Skill 运行时需要执行的业务 dry-run/评测操作仍属于运行合同。

 数据库是权威：Skill 对自己读写哪些表负责，声明失败时如何关闭，并确保经手的输入和产出都落库。没落库的输入产出，在系统视角里等于不存在。只读看板只查库渲染，不替 Skill 写任何数据。

@ -49,7 +49,7 @@ Agent 与 Skill 领域拥有角色职责、可调用能力合同、确定性工
 - 默认从仓库任意工作目录调用，必须自行解析项目根和输入绝对路径，不能依赖调用者先 `cd` 到特定目录。
 - 机械事实必须结构化输出稳定状态和错误码；人读日志是补充，不是唯一接口。
 - Tool 不调用模型，除非所属 Skill 明确声明该步骤本质需要模型。
- Tool 变更必须有离线测试、`py_compile` 和 `git diff --check` 证据。
+- Tool 变更必须有登记在 `harness/manifests/` 的相关实现测试、`py_compile` 和 `git diff --check` 证据；测试结果只证明机械合同，不升级为 Skill 行为或内容质量结论。

 ## 5. ReAct 工作方式

--- a/.claude/skills/access-database/SKILL.md
+++ b/.claude/skills/access-database/SKILL.md
@ -24,9 +24,6 @@ description: 通过唯一受控入口查询或修改 muse-example PostgreSQL，
 # 大对象（raw 全文）经 stdin 传 JSON 数组（避开 shell 转义 / ARG_MAX）：
 .venv/bin/python .claude/skills/access-database/scripts/db.py execparams "INSERT ... VALUES (%s,%s)" --stdin < params.json    # params.json = ["<sha>", "<完整全文>"]

-# execparams 参数装载逻辑离线自测（不连库）
-.venv/bin/python .claude/skills/access-database/scripts/test_db_params.py
-
 # SQL 文件应用：整文件一个事务，失败全回滚
 .venv/bin/python .claude/skills/access-database/scripts/db.py apply db/ddl/91-example实验私货.sql

--- a/.claude/skills/call-content-model/SKILL.md
+++ b/.claude/skills/call-content-model/SKILL.md
@ -7,7 +7,7 @@ description: 通过 New-API 的统一治理入口调用内容模型，执行额

 创始人拍板（2026-07-13）：清洗与拆书的内容生产 LLM **全部走 New-API 的 MiniMax-M3**；主会话（Fable5）只固化 agent/提示词/skill 与发起调用。本 skill 是唯一出口。

-管线内容生产调用的**标准入口是 `chat_governed`**（受 5 小时额度窗 + 全局降级链治理）；`chat`/`chat` CLI 是不受治理的直连，仅供调试。治理政策的机械事实源是 `.claude/skills/call-content-model/scripts/llm.py` + `test_quota.py`（AGENTS.md §6 点名），模型链切换必须由该 skill 治理并留下日志。
+管线内容生产调用的**标准入口是 `chat_governed`**（受 5 小时额度窗 + 全局降级链治理）；`chat`/`chat` CLI 是不受治理的直连，仅供调试。治理政策的机械事实源是运行配置、共享额度账本 `example_llm_quota` 和模型运行适配器（`llm.py`），模型链切换必须由该 skill 治理并留下日志。

 ## 用法

--- a/.claude/skills/capture-ai-flavor-cases/SKILL.md
+++ b/.claude/skills/capture-ai-flavor-cases/SKILL.md
@ -121,13 +121,3 @@ disable-model-invocation: true
 `dashboard/server.py` 的 `/ai-flavor` 默认查这三张表，数据库不可用时才明确标注离线回退；页面不会因打开而重新读取原文。

 状态语义：案例卡 `shadow` 只供复核，`canonical` 仅表示获授权且完成评审，`rejected`/`archived` 不进入生成上下文；重验证 `verified` 才能确认、投影样例或消费规则，`stale`（全文哈希变化）、`unavailable`（来源不可得）和 `card_mismatch`（锚点变化）都使当前卡在这些动作上失效，但历史回执保留。
-
-## 机械验收
-
-```bash
-.venv/bin/python .claude/skills/capture-ai-flavor-cases/scripts/test_capture_cases.py
-.venv/bin/python .claude/skills/access-database/scripts/test_skill_catalog.py
-git diff --check
-```
-
-测试必须覆盖：来源 hash、重验证 verified/stale/unavailable/card_mismatch、未授权 hash-only、重复 ID、live feedback 来源绑定、shadow 不能投影样例、缺重验证回执不能确认、跨作品/反例门，以及候选评测不足时不能 active。
--- a/.claude/skills/diagnose-ai-flavor/SKILL.md
+++ b/.claude/skills/diagnose-ai-flavor/SKILL.md
@ -31,10 +31,3 @@ disable-model-invocation: true
 - 诊断不得改正文、不得产 patch。
 - `decision_proposal` 只是机械层初步建议；候选/语义命中必须经过功能仲裁，不能当执行指令。
 - 命中不等于修改命令：发现清单是给仲裁的输入，不是执行指令。
-
-## 自测
-
-```bash
-cd agent-example
-.venv/bin/python .claude/skills/diagnose-ai-flavor/scripts/test_diagnose_ai_flavor.py
-```
--- a/.claude/skills/embed-knowledge/SKILL.md
+++ b/.claude/skills/embed-knowledge/SKILL.md
@ -32,16 +32,6 @@ description: 使用固定 Qwen3 嵌入模型将知识草稿或实体批量写入
 - **落库字段**：`example_knowledge_embedding(draft_id, content_hash, embed_text, model, dimensions=1024, embedding)`；draft 确认落 entity 后由 confirm 流程把 owner 迁到 `entity_id` 并清空 `draft_id`，关系草稿确认后关闭无 canonical owner 的临时向量。
 - 汇报：新嵌 N、跳过 M、失败 K；幂等、失活和冲突原因均输出可追踪明细。

-## 离线验证
-
-```bash
-.venv/bin/python .claude/skills/embed-knowledge/scripts/test_embed_drafts_offline.py
-.venv/bin/python -m py_compile .claude/skills/embed-knowledge/scripts/embed_drafts.py \
-  .claude/skills/embed-knowledge/scripts/test_embed_drafts_offline.py
-```
-
-离线测试只使用 fake connection 检查并发顺序、SQL 条件和 owner 反例，不连接真实数据库，不调用 embedding 或 reset。
-
 ## 红线

 - 调用必须 `trust_env=False`（系统代理会假 502）；令牌用 `MUSE_AI_NEW_API_TOKEN`（勿用管理令牌，打 /v1 报无效）。
--- a/.claude/skills/establish-voice-baseline/SKILL.md
+++ b/.claude/skills/establish-voice-baseline/SKILL.md
@ -36,11 +36,3 @@ disable-model-invocation: true
 - 只读已确认正文与作者样张；不得从草稿或未确认候选归纳基线。
 - 基线必须人工确认（`--reviewer` 必填，数据库 CHECK 兜底）。
 - 本技能不产一个字正文。
-
-## 自测
-
-```bash
-cd agent-example
-.venv/bin/python .claude/skills/establish-voice-baseline/scripts/test_establish_voice_baseline.py
-.venv/bin/python humanization/tests/test_humanization_v2.py
-```
--- a/.claude/skills/evaluate-frozen-replay/SKILL.md
+++ b/.claude/skills/evaluate-frozen-replay/SKILL.md
@ -48,14 +48,14 @@ disable-model-invocation: true

 ## 编排入口

- `scripts/run_replay.py --mode dry_run`：只执行授权、来源、冻结和三臂 manifest 预检，不调用模型；这是首个机制 smoke 入口。
+- `scripts/run_replay.py --mode dry_run`：只执行授权、来源、冻结和三臂 manifest 预检，不调用模型；用于确认评测执行的前置条件。
 - `freeze-context/scripts/load_reference_work.py`：从 PostgreSQL 只读事务组装仓库外临时配置；读取 `example_reference_authorization_snapshot` 当前原文件版本的最新快照，组装 authorization 外层与 snapshot。缺授权记录仍生成可审计配置，但送入 `run_replay` 后必须保持 `blocked_authorization`。
 - `scripts/run_replay.py --mode execute`：在全部前置门通过后，依次执行三臂 planner、整组 schema、逐臂盲 detector、两个独立盲 judge、rubric 校验、稳定性门和去盲汇总；`--output-dir` 必须位于仓库外的临时目录。
 - detector 输入输出均为 JSON；输入只有匿名候选 ID、候选和公共冻结到 `as_of` 的规划上下文，不含 arm 名、`cardInjection`、`cardManifest`、任何臂特有卡内容、目标章 proxy 或其他评委结果。卡注入合法性只由确定性预检负责。报告不合约属于系统失败，须在 judge 前失败关闭；合同合法的 `failed` / `needs_evidence` 属于候选质量信号，三臂均须保留候选与报告并继续盲评和 Gate，Gate A 只由 C 臂高严重度残留与硬约束覆盖率裁决质量失败。
 - 合法的非 `passed` 报告在安全 manifest 与 CAS 中只保存按臂分组的 `semantic-diagnostic-v1`：固定结果枚举、固定原因/区段枚举、调用次数和有界阻断计数。完整报告只留受控 raw 和编排器内存中的 Gate builder 输入；安全输出不得保存错误消息、正文、引文、事实文本、补证查询/原因、任何检测项 ID/字段路径、纠错草稿或模型原始字段。检测器不合约时同样只保存固定闭集诊断并失败关闭。
 - 任一样本发生系统失败、检测器不合约、盲评无效/不稳定、预算或 raw/CAS 失败后，整轮已无法构建完整 Gate 输入，必须立即停止后续样本并进入统一迁移；不得继续调用模型消耗预算，也不得因失败删除本轮诊断 raw。
 - 两个 judge 使用不同 `judgeId` 和独立无会话进程。第二个 judge 的匿名候选顺序必须与第一个完全相反；任何 rubric 不合约标记 `judge_invalid`，任一同维差值大于 `0.5` 标记 `judge_unstable`，两者都不得标记 `completed`。
- 只有三臂 schema 合法、detector 均产出合同合法报告、双 judge rubric 与稳定性门通过，才去盲生成逐维 `B-A` / `C-A` 差值矩阵并标记 `completed`；detector 报告合同合法不等于其质量终态必须为 `passed`。`--detector-bin`、`--judge-primary-bin`、`--judge-secondary-bin` 可分别指定本地 runner；未指定时复用 `--planner-bin`，测试只能使用 fake binary。
+- 只有三臂 schema 合法、detector 均产出合同合法报告、双 judge rubric 与稳定性门通过，才去盲生成逐维 `B-A` / `C-A` 差值矩阵并标记 `completed`；detector 报告合同合法不等于其质量终态必须为 `passed`。`--detector-bin`、`--judge-primary-bin`、`--judge-secondary-bin` 可分别指定本地 runner；未指定时复用 `--planner-bin`；离线评测可使用 fake adapter，真实评测执行使用配置绑定的 runner。
 - planner、detector、judge 子进程统一受 `--timeout-seconds` 限制，默认 300 秒；任一超时分别落盘 `planner_timeout`、`detector_timeout`、`judge_timeout`，不得继续进入后续阶段或标记 `completed`。
 - `scripts/write_report.py`：从 `run_result.json` 生成独立严格 schema 的安全摘要，只接受受限标识符、枚举、数字、短安全摘要和 SHA-256；不会读取候选正文，也不会把候选路径以外的原始响应写入报告。

@ -67,7 +67,7 @@ disable-model-invocation: true

 真实五章装配还必须在 `commonControls` 预注册选择器版本、选择器规范 JSON 的原始字节 SHA-256、`inputProvenance=oracle_reference_scaffold` 和统一 `maxContextChars=140000`。五个样本的 A/C 补充原文预算统一预注册为 2000 Unicode code point；loader 在数据库读取前机械核对这些值，不得根据实际文本降低或抬高预算。连续四章基线装不下、A/C 任一臂补充原文不足 2000，或两臂 WriterCreativeInput hash 相同/允许差异为空时均失败关闭。

-每个样本必须提供 `writerContextInput`。仓内基础配置只允许 `contentMode=sanitized_contract_fixture`，并为 A/C 各提供恰好 2000 code point 的明确脱敏合成补充原文，用于机械证明 WriterContext 合同、章号冻结和创作输入差异；这些文本不得来自原书。loader 装配真实临时配置后必须改为 `canonical_frozen_prose`。生产 `--execute` 在创建 vault 或 runner 前拒绝脱敏夹具；脱敏夹具只可由离线测试适配器执行。任何模式都不得把原书全文、完整目标细纲或标准答案写入 Git。
+每个样本必须提供 `writerContextInput`。仓内基础配置只允许 `contentMode=sanitized_contract_fixture`，并为 A/C 各提供恰好 2000 code point 的明确脱敏合成补充原文，用于机械证明 WriterContext 合同、章号冻结和创作输入差异；这些文本不得来自原书。loader 装配真实临时配置后必须改为 `canonical_frozen_prose`。生产 `--execute` 在创建 vault 或 runner 前拒绝脱敏夹具；脱敏夹具只可由离线评测适配器执行。任何模式都不得把原书全文、完整目标细纲或标准答案写入 Git。

 三臂都必须构造并校验完整 `WriterContext v1`，且固定 `mode=diagnostic_only`、`purpose=evaluation`、`acceptanceEligible=false`。adapter 随后投影 `WriterCreativeInput v2`；writer 模型只看到创作投影，不看到运行身份、manifest、hash、实验臂或验收状态：

@ -99,8 +99,6 @@ dry-run 只输出计划、manifest 和上下文摘要，不调用 writer、seman

 正式 execute 还必须显式传 `--raw-archive-dir <仓外绝对目录>`；缺失、相对路径、位于本轮输出目录内或与临时 vault 跨文件系统时，均在建立 vault 和调用模型前失败关闭。运行结束以 `rawDisposition.status=migrated` 和 `raw-vault-migration-receipt-v1` 证明产物已迁移，不再以删除后的 `closed` 作为完成条件。

-离线完整链冒烟由 `scripts/test_run_writer_replay.py` 的单样本三臂 production fake 用例承担：它真实经过 writer 投影、机械门、semantic detector、双评委、raw vault、CAS 和 GateInputBuilder，但所有模型均为本地确定性 fake，不产生费用、不得作为 Gate 样本结果。真实一次调用能力只由 `refresh_runtime_probe.py` 的合成 writer 探针验证；它不代表 detector/judge 或五样本 Gate 已通过。
-
 预算合同分为不可变的 `plannedCalls` 和独立安全上限 `maxCalls`。Gate A 当前预注册 writer 60 次（15 次基础写作 + 最多 45 次篇幅修订）、semantic detector 24 次（15 次基础检测 + 9 次纠错/API 重试余量）、blind judge 45 次（最多三位评委，每位基础调用后最多两个格式纠错/API 重试槽位）；各 `maxCalls=150`、单次 cap `$5`、总预算 `$2250`。启动预留只按 `plannedCalls * maxBudgetUsdPerCall`，即 `$645`；`$2250` 是三角色各 150 次安全容量对应的有限执行上限，不是预计消费。本预算授权不等于正式 `--execute` 授权。每次调用前账本同时检查角色计划槽位、`maxCalls`、累计实际成本和未执行计划的最坏预留；调用后只能使用可信 `ExecutionReceipt.totalCostUsd` 结算。缺回执成本、重复/错序结算、单次 cap 或总预算越界均 fail closed，未触发的修订、纠错与第三评计划必须保留在 `remainingPlannedCalls`。

 ## 稳定运行纪律
@ -124,12 +122,12 @@ Gate 裁决器 `writer_gate.py`（归属 `score-content-quality`，本 Skill 只
 `scripts/refresh_runtime_probe.py` 是刷新 `executionAuthorization.runtimeProbe` 的唯一通道。正式执行门要求探针的 `executionProfileSha256` 必须等于当前 writer profile 的身份哈希；writer 合同（schema/prompt/模型/CLI/预算）升级后旧探针必然失配，门以 `EXECUTE_PROBE_CONTRACT_MISMATCH` 失败关闭，此时必须用当前合同重测探针，不得手动改探针值绕过。

 - 输入合同：只读 `--config` 指定的 base 配置，只用 `executionProfiles.writer`（探针只验证 writer 合同），不读 semantic/judge，不读 `oracleTruthPacks`/`samples`/raw 物料。探针输入是固定、极小、全合成的能力探针 prompt（一段自包含的合成微任务），绝不使用原书全文、真实正文或 raw。
- 调用合同：Claude 调用经可注入 invoker 发起，默认 invoker = `claude_runtime.run_claude`（真实调用层是薄薄一层）；`--dry-run` 注入固定假产出，只走通「构建 profile + 重签 + 写文件」链路供冒烟，不发起真实调用。
+- 调用合同：Claude 调用经可注入 invoker 发起，默认 invoker = `claude_runtime.run_claude`（真实调用层是薄薄一层）；`--dry-run` 注入固定假产出，只走通「构建 profile + 重签 + 写文件」链路供离线预览，不发起真实调用。
 - 输出合同：把绑定新合同的探针替换 `executionAuthorization.runtimeProbe`，并按与门校验逐字节同源的规范自哈希算法重签该探针的 `receiptSha256`（`budget`/`rawRetention` 原样保留、各自自哈希不变），把刷新后的完整配置写到 `--output` 指定的【新文件】，不就地覆盖原配置，便于 review diff 后再替换。新探针字段集与现有 runtimeProbe 完全一致，不缺不多。
 - 失败关闭：调用失败 / 结构化输出 schema 不过 / 单次成本超冻结 cap / 超 deadline / 回执不可信（模型不匹配、标记错误、非零退出、异常终止、API 错误、身份哈希指向旧合同、输出哈希与产出不一致）→ 不写 successful 探针、不产出刷新配置、退出非零，原配置保持不动。
 - 审计合同：标准输出只记录做了什么（探针身份哈希、结构化输出哈希、回执哈希、成本、是否成功、输出文件），绝不打印完整 prompt/response、raw 路径或供应商原始响应。

-离线用法（冒烟，不调模型）：
+离线用法（不调用模型）：

 ```bash
 .venv/bin/python .claude/skills/evaluate-frozen-replay/scripts/refresh_runtime_probe.py \
--- a/.claude/skills/execute-claude-task/SKILL.md
+++ b/.claude/skills/execute-claude-task/SKILL.md
@ -25,9 +25,3 @@ disable-model-invocation: true
 - 不从全局 Claude 会话继承业务状态，不静默换模型或供应商。
 - 不自行决定候选是否通过，不写 Candidate、Quality 或 Canonical 数据。
 - 不拥有 CAS、raw vault、运行登记或回执修复实现。
-
-## 离线验证
-
-```bash
-.venv/bin/python .claude/skills/execute-claude-task/scripts/test_claude_runtime.py
-```
--- a/.claude/skills/extract-work-knowledge/SKILL.md
+++ b/.claude/skills/extract-work-knowledge/SKILL.md
@ -43,10 +43,3 @@ disable-model-invocation: true
 - 写入 `muse_knowledge_draft`、`example_upgrade_window`、`example_upgrade_alias`、`example_upgrade_presence`、`example_upgrade_card_state`、`example_upgrade_audit` 和 `example_knowledge_embedding`。
 - `SOURCE_TYPE="upgrade_book"`、`upgrade-reset` 等是数据库历史兼容值，不随 Skill 改名。
 - 不提供修复、重置、备份、恢复或存量迁移入口；这些动作统一由 `maintain-work-extraction` 承担。
-
-## 离线验证
-
-```bash
-.venv/bin/python .claude/skills/extract-work-knowledge/scripts/test_upgrade_work_lock_offline.py
-.venv/bin/python .claude/skills/extract-work-knowledge/scripts/test_parse_upgrade_offline.py
-```
--- a/.claude/skills/maintain-work-extraction/SKILL.md
+++ b/.claude/skills/maintain-work-extraction/SKILL.md
@ -69,10 +69,3 @@ reset 软删目标 `upgrade_book` 草稿和无 Canonical owner 的活向量，
 - 备份与 preview 默认只读；execute 只修改目标租户、目标作品、`upgrade_book` 边界内的派生状态。
 - 不生成新抽取内容，不提供正常 `run/windows/status`。
 - `migrate_upgrade_windows.py` 当前只输出迁移计划，不落库；真实写入仍需另行 gate 与授权。
-
-## 离线验证
-
-```bash
-.venv/bin/python .claude/skills/maintain-work-extraction/scripts/test_backup_upgrade_work_offline.py
-.venv/bin/python .claude/skills/maintain-work-extraction/scripts/test_reset_upgrade_work_offline.py
-```
--- a/.claude/skills/prevent-ai-flavor/SKILL.md
+++ b/.claude/skills/prevent-ai-flavor/SKILL.md
@ -31,11 +31,3 @@ disable-model-invocation: true
 - 读：`example_voice_baseline` 当前 canonical 版本；规则/样例读 `humanization/` Git 资产。
 - 写：`example_run`（幂等 upsert，记录规则指纹、声音账 hash 和投影条数）。
 - 失败：作品不匹配、声音账非 canonical、规则装载门失败或 guidance 超合同直接失败关闭。
-
-## 自测
-
-```bash
-cd agent-example
-.venv/bin/python .claude/skills/prevent-ai-flavor/scripts/test_prevent_ai_flavor.py
-.venv/bin/python .claude/skills/assemble-context/scripts/test_assemble_writer_context.py
-```
--- a/.claude/skills/record-run-evidence/SKILL.md
+++ b/.claude/skills/record-run-evidence/SKILL.md
@ -32,11 +32,3 @@ disable-model-invocation: true
 - 读取运行、调用、raw 和回执当前状态，只为校验绑定与幂等。
 - 写入 `example_run`、`example_llm_call`、`example_raw_lease`、`example_raw_content`、`example_run_receipt` 和相应质量失败记录。
 - 不调用 Claude/New-API，不生成候选，不判断内容是否通过。
-
-## 离线验证
-
-```bash
-.venv/bin/python .claude/skills/record-run-evidence/scripts/test_file_cas.py
-.venv/bin/python .claude/skills/record-run-evidence/scripts/test_raw_vault.py
-.venv/bin/python .claude/skills/record-run-evidence/scripts/test_run_registry.py
-```
--- a/.claude/skills/revise-ai-flavor/SKILL.md
+++ b/.claude/skills/revise-ai-flavor/SKILL.md
@ -30,10 +30,3 @@ disable-model-invocation: true
 - 没有诊断产物、作者授权或功能仲裁不得修订。
 - 硬门失败不得以风格分、盲评结果抵消；旧诊断不能跨规则库/正文 hash 复用。
 - 语义级不变量（因果、POV、伏笔状态）机械未覆盖的，如实列 `unresolved_risks`，不假装验证完成。
-
-## 自测
-
-```bash
-cd agent-example
-.venv/bin/python .claude/skills/revise-ai-flavor/scripts/test_revise_ai_flavor.py
-```
--- a/AGENTS.md
+++ b/AGENTS.md
@ -23,7 +23,8 @@ SoT 按主题分域，不做跨主题的全局排序。可执行脚本与书面
 | [`.agent/docs/architecture/domains/`](.agent/docs/architecture/domains/_index.md) | 本仓领域边界、数据权威、落库合同和领域协作的 SoT。 |
 | [`meta/schemas/`](meta/schemas/) | 23 型结构本体的字段合同；库内 payload 结构以该合同为准。 |
 | [`meta/chains/`](meta/chains/) | scenario、purpose、功能 skill、角色槽位和保护节点的链路登记。 |
-| [`.claude/skills/`](.claude/skills/) | 每个 `SKILL.md` 定义能力的输入、输出、红线、数据库读写合同、输入产出落库和自测入口；同目录 `scripts/` 是确定性实现与机械门。 |
+| [`.claude/skills/`](.claude/skills/) | 每个 `SKILL.md` 定义项目运行时能力的输入、输出、红线、数据库读写合同和输入产出落库；同目录 `scripts/` 是运行时确定性实现与机械门，开发验证和行为评测入口由 `harness/manifests/` 登记。 |
+| [`tests/skills/`](tests/skills/) | 项目运行时 Skill 的实现测试、集成测试和 fake pipeline 测试；它们提供回归证据，不拥有运行时合同，也不等同于 Skill 行为评测。 |
 | [`.agent/`](.agent/_index.md) | 跨任务长期知识；领域设计位于 `docs/architecture/domains/`，新增、删除或重命名必须同步各级 `_index.md`。 |
 | [`docs/`](docs/) | 单次任务探索、计划、评测资料、样张和历史执行证据；任务完成后把稳定结论蒸馏到 `.agent/`，不得长期拥有领域定义。 |
 | [`README.md`](README.md) | 项目背景和历史路线概览。目录、阶段、技能数量和存储方式等描述可能陈旧，不得覆盖本文件、`meta/`、skill 或磁盘事实。 |
@ -46,7 +47,9 @@ agent-example/
 │   ├── ddl/                # 可审计 DDL / 迁移文件
 │   ├── 表映射.md
 │   └── 连接信息.md
-├── humanization/           # 去 AI 味与人感资产层（合同/规则/样例/执行骨架；验证期自治，升华进 Muse 时同步回父仓；20 项研究覆盖矩阵见 humanization/research/）
+├── humanization/           # 去 AI 味 Skill 能力域（规则/样例/执行骨架；运行时资产目标为数据库）
+├── harness/                # 项目验证与外部评测索引、清单和调度支架
+├── tests/                  # 运行时 Skill 实现测试（按 Skill 归档）
 ├── knowledge/              # 仓内参考资产；未经绑定、授权不得进入上下文
 ├── docs/                   # 设计、评测、样张与历史执行记录
 ├── .venv/                  # 本地 Python 运行环境
@ -57,15 +60,23 @@ agent-example/

 5 个角色：`writer`、`planner`、`extractor`、`detector`、`judge`。角色身份在 `.claude/agents/*.md`，具体功能合同不复制进角色文件。

-技能按能力分组如下，实际清单以 `.claude/skills/*/SKILL.md` 为准，不手填数量镜像：
+### Skill 领域索引（项目运行时 Skill）

- 数据与检索：`access-database`、`import-book`、`embed-knowledge`、`search-knowledge`、`call-content-model`。
- 执行、证据与上下文：`execute-claude-task`、`record-run-evidence`、`freeze-context`、`assemble-context`。
- 流程与主权治理：`decide-candidate`。
- 清洗、拆解与知识：`clean-book-text`、`deconstruct-book`、`extract-chapter-knowledge`、`review-knowledge-cards`、`extract-work-knowledge`、`maintain-work-extraction`。
- 去 AI 味与人感（父仓专题-09 五技能先行验证）：`establish-voice-baseline`（定基线）、`prevent-ai-flavor`（前置预防）、`diagnose-ai-flavor`（诊断）、`revise-ai-flavor`（修订）、`capture-ai-flavor-cases`（挖掘/案例采集）；共享资产层在 `humanization/`，接力铁律见 `meta/chains/`。
- 规划与写作：`design-story-foundation`、`plan-story`、`plan-chapter`、`write-next-chapter`、`rewrite-selection`、`expand-scene`、`polish-prose`。
- 检测与质量评测：`check-content-consistency`、`score-content-quality`、`optimize-content-quality`、`evaluate-frozen-replay`。
+实际清单以 `.claude/skills/*/SKILL.md` 为准；本表维护合同责任方、协作领域和领域 SoT，不复制各 Skill 的完整合同。每个运行时 Skill 必须登记一个合同责任方（业务领域或平台领域），但可以同时消费或影响多个协作领域；跨域调用、场景关系和保护节点在 `meta/chains/` 登记。合同责任方表示谁维护该 Skill 的稳定能力合同，不表示 Skill 只能属于一个业务领域。
+
+| 合同责任方 / 能力域 | 领域 SoT | 运行时 Skill |
+|---|---|---|
+| 平台运行与证据 | 07-Agent 与 Skill、08-数据权威与可视化 | `access-database`、`call-content-model`、`execute-claude-task`、`record-run-evidence` |
+| 上下文与知识检索 | 02-实体、04-上下文、08-数据权威与可视化 | `assemble-context`、`embed-knowledge`、`freeze-context`、`search-knowledge` |
+| 导入、清洗与抽取 | 02-实体、05-创作流程 | `clean-book-text`、`deconstruct-book`、`extract-chapter-knowledge`、`extract-work-knowledge`、`import-book`、`maintain-work-extraction`、`review-knowledge-cards` |
+| 规划与作品基础 | 05-创作流程 | `design-story-foundation`、`plan-chapter`、`plan-story` |
+| 写作与候选主权 | 01-作品、05-创作流程 | `decide-candidate`、`expand-scene`、`polish-prose`、`rewrite-selection`、`write-next-chapter` |
+| 质量与回放评测 | 06-质量与复利、05-创作流程 | `check-content-consistency`、`evaluate-frozen-replay`、`optimize-content-quality`、`score-content-quality` |
+| 去 AI 味与人感 | 06-质量与复利、父仓专题-09 | `capture-ai-flavor-cases`、`diagnose-ai-flavor`、`establish-voice-baseline`、`prevent-ai-flavor`、`revise-ai-flavor` |
+
+`humanization/` 是“去 AI 味与人感”运行时 Skill 家族的能力域：`src/deai/` 是共享实现，规则、样例、案例和声音资产是该域的数据依赖，运行时以数据库为权威；仓内 YAML/JSON 在数据库化完成前只作为迁移种子、离线夹具或结构合同。规则记录不各自注册为 Skill，Skill 负责动作和消费边界。
+
+Skill 领域列表的新增、删除、改名或主领域调整，必须同时检查 `.claude/skills/`、`meta/chains/README.md` 和相关领域 `_index.md`；不得只改本表造成索引漂移。

 上下文目标合同以 [上下文领域 SoT](.agent/docs/architecture/domains/04-上下文领域.md) 为准：数据库读取器是核心实现，库内检索加速（向量）只做候选召回；任何命中都要回读库行并校验 hash。Skill 对自己读写哪些表负责，并把经手的输入和产出落库；没落库的输入产出在系统视角里等于不存在。

@ -95,7 +106,7 @@ agent-example/

 ## 6. 模型边界

- 清洗、抽卡、范式拆取及其模型调用统一走 `call-content-model` Skill，不裸调 New-API。治理政策固定为 5 小时额度窗：MiniMax 模型累计花费上限 `$24`，全模型成功调用上限 `6000`；机械事实源是 `.claude/skills/call-content-model/scripts/llm.py` 及 `test_quota.py`，模型链切换必须由该 Skill 治理并留下日志。
+- 清洗、抽卡、范式拆取及其模型调用统一走 `call-content-model` Skill，不裸调 New-API。治理政策固定为 5 小时额度窗：MiniMax 模型累计花费上限 `$24`，全模型成功调用上限 `6000`；运行适配器、正式配置和账本是额度合同的事实源，`.claude/skills/call-content-model/scripts/llm.py` 是实现，`test_quota.py` 只提供回归证据；模型链切换必须由该 Skill 治理并留下日志。
 - 角色模型归属：`planner`/`writer`/`judge` 固定 `opus`；`extractor`/`detector` 可用其它模型（非必须降级）。拆书/导入侧抽取经 `call-content-model`/`deconstruct-book` Skill 走 MiniMax-M3，不走角色 model 派发；创作期章后抽取作为角色派发，可用 `opus`。
 - 确定性脚本、合同校验、快照冻结、泄漏审计和报告生成不调用模型；除非对应 `SKILL.md` 明确声明模型步骤，不得把机械任务升级为模型任务。
 - Claude 生成或评测只在对应任务 SoT、显式预算、固定执行配置和原文用途授权全部满足后运行；任一前置门失败都应关闭执行。
@ -130,12 +141,12 @@ agent-example/

 - 明确区分**已验证事实**、**推断**和**假设**。报告必须给出证据来源；没有机械输出或运行证据时，不声称完成、修复或通过。
 - 禁止从 `n=1` 样本推出普适结论。至少分析假阴、假阳、样本偏差和混淆因素；需要判断卡或模型效果时使用同任务、同模型、同预算、同公共上下文的对照，并把不稳定样本排除在方向结论之外。
- Python 一律使用仓内解释器 `.venv/bin/python`。先读目标 skill 的 `SKILL.md`，再运行其已存在的相关自测；例如：
+- Python 一律使用仓内解释器 `.venv/bin/python`。先读目标 Skill 的 `SKILL.md`，再按 `harness/manifests/` 登记的类别和依赖选择相关验证；下面命令仅是现有局部验证入口示例：

 ```bash
-.venv/bin/python .claude/skills/plan-chapter/scripts/test_contract.py
-.venv/bin/python .claude/skills/call-content-model/scripts/test_quota.py
-.venv/bin/python .claude/skills/assemble-context/scripts/test_writer_contract.py
+.venv/bin/python tests/skills/plan-chapter/test_contract.py
+.venv/bin/python tests/skills/call-content-model/test_quota.py
+.venv/bin/python tests/skills/assemble-context/test_writer_contract.py
 git diff --check
 ```

@ -163,7 +174,7 @@ git diff --check
 1. **给谁用**：主会话编排？某角色 agent？还是别的 skill？消费者必须明确、单一。
 2. **目的单一**：一个 skill 只实现一个能力。功能并列、又当编排又当执行的，拆。
 3. **可靠性与稳定性**：失败是否明确失败关闭（不静默降级、不返回假成功）？错误是否带稳定码、不泄漏原文/密钥？有无离线自测覆盖关键路径与失败路径？确定性步骤是否真不调模型？
-4. **符合 skill 规范**：`SKILL.md` 是否清楚定义输入、输出、红线、用法？`scripts/` 是否其确定性实现与自测入口？是否走 `.venv`、不裸调 PG/New-API/模型？
+4. **符合 skill 规范**：`SKILL.md` 是否清楚定义输入、输出、红线、用法？`scripts/` 是否只包含运行时确定性实现与机械门？相关实现测试和行为评测是否已在 `harness/manifests/` 登记？是否走 `.venv`、不裸调 PG/New-API/模型？
 5. **边界一致**：引用的路径、合同、字段是否与现行 SoT（本文件、`.agent/`、`meta/` 和库内正式内容）一致？数据库是正式内容权威，Skill 必须声明读写哪些表、失败如何关闭、输入产出如何落库；引用已失效合同的不得执行。

 命名采用小写 `动作-对象`：目录名必须等于 frontmatter `name`，名称表达可调用能力，不复用 `scenario`、内部模块名或含混阶段词。`scenario`、`source_type`、updater 和备份逻辑键是独立稳定标识，不随 Skill 改名。
--- a/harness/README.md
+++ b/harness/README.md
@ -0,0 +1,68 @@
+# agent-example Harness
+
+> 项目开发、验证和行为评测支架的入口与索引。
+> 这里的文档和工具服务开发者与评测编排，不会作为任何 Skill 的运行时提示词自动注入 Agent。
+
+## 1. 作用
+
+`agent-example` 同时包含三种不同性质的验证工作：
+
+- 运行时文档的静态卫生检查；
+- 确定性工具、数据库边界和 fake pipeline 的机械测试；
+- 将 Skill 交给 Agent 后的外部行为评测。
+
+Harness 负责把这三类工作分开编排并保留证据，不能用其中一类的通过结果冒充另一类结论。
+
+## 2. 目录索引
+
+```text
+harness/
+├── README.md                         # 本入口：项目 harness 介绍、边界和索引
+├── specs/
+│   └── skill-testing.md              # Skill 测试与评测规范（唯一规范正文）
+├── skill_harness.py                  # 运行时 Skill 文档静态审计入口
+├── run_selected.py                   # 按 manifest 选择性执行测试
+├── test_skill_harness.py             # 静态审计器离线测试
+├── test_run_selected.py              # 选择性执行器离线测试
+├── manifests/
+│   ├── skills.json                   # 运行时 Skill 责任方与协作领域清单
+│   └── test-inventory.json           # 测试分类、依赖、副作用与证据等级
+└── evals/<skill>/                    # Skill 行为评测夹具与适配器（尚未建立）
+
+项目级实现测试位于 `../tests/skills/<skill>/`，不进入运行时 Skill 目录。
+```
+
+现有专项支架：
+
+| 路径 | 责任 | 证据边界 |
+|---|---|---|
+| `humanization/eval/run_eval.py` | 去 AI 味规则合同和回归陷阱试跑 | 规则合同回放，不证明文学效果 |
+| `.claude/skills/evaluate-frozen-replay/` | 正文/细纲冻结回放、盲评和质量门 | 评测编排与质量证据，不是通用 Skill 单测入口 |
+| `harness/` | 跨 Skill 的开发验证与行为评测治理 | 只编排和裁决证据，不拥有业务合同 |
+
+## 3. 权威关系
+
+- `.claude/skills/*/SKILL.md`：Skill 的运行时行为合同；不放开发测试说明。
+- `.claude/skills/*/scripts/`：Skill 使用的运行时确定性实现与机械门。
+- `tests/skills/<skill>/`：对应 Skill 的实现测试、集成测试和 fake pipeline 测试；测试性质由 harness 清单标注。
+- `harness/specs/`：测试和评测规范；不被业务 Skill 当作运行时指令读取。
+- `harness/manifests/`：测试/评测登记与执行配置；不复制业务字段合同。
+- `harness/evals/`：外部行为评测案例、适配器和报告生成逻辑。
+- `.agent/docs/architecture/domains/`：领域与数据权威 SoT；harness 只引用，不复制领域定义。
+- `docs/`：单次设计、运行证据和历史资料；不替代 harness 规范。
+
+## 4. 当前状态
+
+- `skill_harness.py` 已实现运行时 Skill 文档与 `skills.json` 的只读静态审计。
+- `test-inventory.json` 已登记实现测试、集成测试、fake pipeline、领域评测和 harness 自测的依赖与证据等级。
+- `run_selected.py` 已实现显式选择、磁盘/manifest 对账、危险依赖阻断、超时和执行证据检查；它不提供默认全仓一键通过结论。
+- 项目运行时 Skill 的实现测试以 `tests/skills/<skill>/` 为目标位置，物理现状以 `test-inventory.json` 为准。
+- `skill_behavior_eval` 当前登记数量为 0；没有外部 Agent/模型行为证据时，不声称 Skill 内容有效。
+- PostgreSQL、网络和真实模型证据未由离线结果替代，是否执行仍受授权和预算约束。
+
+## 5. 运行边界
+
+- 默认只允许静态检查和明确标注的离线测试；
+- PostgreSQL、网络、真实模型和额度探针必须显式选择并满足授权、预算和环境前置条件；
+- 任何测试失败、依赖缺失或未发现测试都必须显式返回状态，不能静默跳过；
+- 报告必须区分已验证事实、推断和未验证假设。
--- a/harness/manifests/skills.json
+++ b/harness/manifests/skills.json
@ -0,0 +1,308 @@
+{
+  "schema_version": 1,
+  "skills": [
+    {
+      "name": "access-database",
+      "contract_owner": "平台运行与证据",
+      "collaborates_with": [
+        "上下文与知识检索",
+        "导入、清洗与抽取",
+        "规划与作品基础",
+        "写作与候选主权",
+        "质量与回放评测",
+        "去 AI 味与人感"
+      ],
+      "skill_path": ".claude/skills/access-database/SKILL.md"
+    },
+    {
+      "name": "assemble-context",
+      "contract_owner": "上下文与知识检索",
+      "collaborates_with": [
+        "规划与作品基础",
+        "写作与候选主权",
+        "质量与回放评测",
+        "去 AI 味与人感"
+      ],
+      "skill_path": ".claude/skills/assemble-context/SKILL.md"
+    },
+    {
+      "name": "call-content-model",
+      "contract_owner": "平台运行与证据",
+      "collaborates_with": [
+        "导入、清洗与抽取",
+        "质量与回放评测"
+      ],
+      "skill_path": ".claude/skills/call-content-model/SKILL.md"
+    },
+    {
+      "name": "capture-ai-flavor-cases",
+      "contract_owner": "去 AI 味与人感",
+      "collaborates_with": [
+        "质量与回放评测",
+        "写作与候选主权"
+      ],
+      "skill_path": ".claude/skills/capture-ai-flavor-cases/SKILL.md"
+    },
+    {
+      "name": "check-content-consistency",
+      "contract_owner": "质量与回放评测",
+      "collaborates_with": [
+        "上下文与知识检索",
+        "规划与作品基础",
+        "写作与候选主权"
+      ],
+      "skill_path": ".claude/skills/check-content-consistency/SKILL.md"
+    },
+    {
+      "name": "clean-book-text",
+      "contract_owner": "导入、清洗与抽取",
+      "collaborates_with": [
+        "平台运行与证据"
+      ],
+      "skill_path": ".claude/skills/clean-book-text/SKILL.md"
+    },
+    {
+      "name": "decide-candidate",
+      "contract_owner": "写作与候选主权",
+      "collaborates_with": [
+        "平台运行与证据",
+        "质量与回放评测",
+        "导入、清洗与抽取",
+        "规划与作品基础"
+      ],
+      "skill_path": ".claude/skills/decide-candidate/SKILL.md"
+    },
+    {
+      "name": "deconstruct-book",
+      "contract_owner": "导入、清洗与抽取",
+      "collaborates_with": [
+        "平台运行与证据",
+        "上下文与知识检索",
+        "质量与回放评测"
+      ],
+      "skill_path": ".claude/skills/deconstruct-book/SKILL.md"
+    },
+    {
+      "name": "design-story-foundation",
+      "contract_owner": "规划与作品基础",
+      "collaborates_with": [],
+      "skill_path": ".claude/skills/design-story-foundation/SKILL.md"
+    },
+    {
+      "name": "diagnose-ai-flavor",
+      "contract_owner": "去 AI 味与人感",
+      "collaborates_with": [
+        "质量与回放评测",
+        "写作与候选主权"
+      ],
+      "skill_path": ".claude/skills/diagnose-ai-flavor/SKILL.md"
+    },
+    {
+      "name": "embed-knowledge",
+      "contract_owner": "上下文与知识检索",
+      "collaborates_with": [
+        "导入、清洗与抽取"
+      ],
+      "skill_path": ".claude/skills/embed-knowledge/SKILL.md"
+    },
+    {
+      "name": "establish-voice-baseline",
+      "contract_owner": "去 AI 味与人感",
+      "collaborates_with": [
+        "上下文与知识检索",
+        "写作与候选主权"
+      ],
+      "skill_path": ".claude/skills/establish-voice-baseline/SKILL.md"
+    },
+    {
+      "name": "evaluate-frozen-replay",
+      "contract_owner": "质量与回放评测",
+      "collaborates_with": [
+        "平台运行与证据",
+        "上下文与知识检索",
+        "写作与候选主权"
+      ],
+      "skill_path": ".claude/skills/evaluate-frozen-replay/SKILL.md"
+    },
+    {
+      "name": "execute-claude-task",
+      "contract_owner": "平台运行与证据",
+      "collaborates_with": [
+        "质量与回放评测"
+      ],
+      "skill_path": ".claude/skills/execute-claude-task/SKILL.md"
+    },
+    {
+      "name": "expand-scene",
+      "contract_owner": "写作与候选主权",
+      "collaborates_with": [
+        "上下文与知识检索",
+        "质量与回放评测"
+      ],
+      "skill_path": ".claude/skills/expand-scene/SKILL.md"
+    },
+    {
+      "name": "extract-chapter-knowledge",
+      "contract_owner": "导入、清洗与抽取",
+      "collaborates_with": [
+        "平台运行与证据",
+        "上下文与知识检索",
+        "写作与候选主权"
+      ],
+      "skill_path": ".claude/skills/extract-chapter-knowledge/SKILL.md"
+    },
+    {
+      "name": "extract-work-knowledge",
+      "contract_owner": "导入、清洗与抽取",
+      "collaborates_with": [
+        "平台运行与证据",
+        "上下文与知识检索"
+      ],
+      "skill_path": ".claude/skills/extract-work-knowledge/SKILL.md"
+    },
+    {
+      "name": "freeze-context",
+      "contract_owner": "上下文与知识检索",
+      "collaborates_with": [
+        "平台运行与证据",
+        "质量与回放评测"
+      ],
+      "skill_path": ".claude/skills/freeze-context/SKILL.md"
+    },
+    {
+      "name": "import-book",
+      "contract_owner": "导入、清洗与抽取",
+      "collaborates_with": [
+        "平台运行与证据"
+      ],
+      "skill_path": ".claude/skills/import-book/SKILL.md"
+    },
+    {
+      "name": "maintain-work-extraction",
+      "contract_owner": "导入、清洗与抽取",
+      "collaborates_with": [
+        "平台运行与证据"
+      ],
+      "skill_path": ".claude/skills/maintain-work-extraction/SKILL.md"
+    },
+    {
+      "name": "optimize-content-quality",
+      "contract_owner": "质量与回放评测",
+      "collaborates_with": [
+        "上下文与知识检索",
+        "写作与候选主权"
+      ],
+      "skill_path": ".claude/skills/optimize-content-quality/SKILL.md"
+    },
+    {
+      "name": "plan-chapter",
+      "contract_owner": "规划与作品基础",
+      "collaborates_with": [
+        "上下文与知识检索",
+        "质量与回放评测"
+      ],
+      "skill_path": ".claude/skills/plan-chapter/SKILL.md"
+    },
+    {
+      "name": "plan-story",
+      "contract_owner": "规划与作品基础",
+      "collaborates_with": [
+        "平台运行与证据",
+        "上下文与知识检索",
+        "写作与候选主权"
+      ],
+      "skill_path": ".claude/skills/plan-story/SKILL.md"
+    },
+    {
+      "name": "polish-prose",
+      "contract_owner": "写作与候选主权",
+      "collaborates_with": [
+        "上下文与知识检索",
+        "质量与回放评测",
+        "去 AI 味与人感"
+      ],
+      "skill_path": ".claude/skills/polish-prose/SKILL.md"
+    },
+    {
+      "name": "prevent-ai-flavor",
+      "contract_owner": "去 AI 味与人感",
+      "collaborates_with": [
+        "上下文与知识检索",
+        "写作与候选主权"
+      ],
+      "skill_path": ".claude/skills/prevent-ai-flavor/SKILL.md"
+    },
+    {
+      "name": "record-run-evidence",
+      "contract_owner": "平台运行与证据",
+      "collaborates_with": [
+        "上下文与知识检索",
+        "导入、清洗与抽取",
+        "规划与作品基础",
+        "写作与候选主权",
+        "质量与回放评测",
+        "去 AI 味与人感"
+      ],
+      "skill_path": ".claude/skills/record-run-evidence/SKILL.md"
+    },
+    {
+      "name": "review-knowledge-cards",
+      "contract_owner": "导入、清洗与抽取",
+      "collaborates_with": [
+        "平台运行与证据",
+        "质量与回放评测"
+      ],
+      "skill_path": ".claude/skills/review-knowledge-cards/SKILL.md"
+    },
+    {
+      "name": "revise-ai-flavor",
+      "contract_owner": "去 AI 味与人感",
+      "collaborates_with": [
+        "质量与回放评测",
+        "写作与候选主权"
+      ],
+      "skill_path": ".claude/skills/revise-ai-flavor/SKILL.md"
+    },
+    {
+      "name": "rewrite-selection",
+      "contract_owner": "写作与候选主权",
+      "collaborates_with": [
+        "上下文与知识检索",
+        "质量与回放评测",
+        "去 AI 味与人感"
+      ],
+      "skill_path": ".claude/skills/rewrite-selection/SKILL.md"
+    },
+    {
+      "name": "score-content-quality",
+      "contract_owner": "质量与回放评测",
+      "collaborates_with": [
+        "平台运行与证据",
+        "上下文与知识检索",
+        "写作与候选主权"
+      ],
+      "skill_path": ".claude/skills/score-content-quality/SKILL.md"
+    },
+    {
+      "name": "search-knowledge",
+      "contract_owner": "上下文与知识检索",
+      "collaborates_with": [
+        "规划与作品基础",
+        "写作与候选主权",
+        "质量与回放评测"
+      ],
+      "skill_path": ".claude/skills/search-knowledge/SKILL.md"
+    },
+    {
+      "name": "write-next-chapter",
+      "contract_owner": "写作与候选主权",
+      "collaborates_with": [
+        "平台运行与证据",
+        "上下文与知识检索",
+        "质量与回放评测",
+        "去 AI 味与人感"
+      ],
+      "skill_path": ".claude/skills/write-next-chapter/SKILL.md"
+    }
+  ]
+}
--- a/harness/manifests/test-inventory.json
+++ b/harness/manifests/test-inventory.json
--- a/harness/run_selected.py
+++ b/harness/run_selected.py
--- a/harness/skill_harness.py
+++ b/harness/skill_harness.py
@ -0,0 +1,870 @@
+#!/usr/bin/env python3
+"""对项目运行时 Skill 文档执行只读静态审计。
+
+本模块只读取 `.claude/skills/*/SKILL.md`，不连接数据库、网络或模型，也不修改
+工作树。机器调用使用 :func:`audit_skills`，命令行入口同时支持人读摘要和 JSON。
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import sys
+from collections import Counter
+from pathlib import Path
+from typing import Any, Iterable, Optional, Sequence
+
+
+SCHEMA_VERSION = 1
+SKILLS_ROOT = Path(".claude") / "skills"
+DEFAULT_MANIFEST_PATH = Path("harness") / "manifests" / "skills.json"
+
+_FRONTMATTER_NAME = re.compile(r"^\s*name\s*:\s*(.*?)\s*$")
+_COVERAGE_TOOL_PATTERN = re.compile(
+    r"(?:\bcoverage\.py\b|\bpytest-cov\b|"
+    r"(?<![A-Za-z0-9_])--cov(?:[-A-Za-z0-9_]+)?(?:\s|=|:|$)|"
+    r"\bcoverage\s+report\b)",
+    re.IGNORECASE,
+)
+_EXPLICIT_TEST_COVERAGE_PATTERN = re.compile(
+    r"(?:"
+    r"(?:测试(?:用例|代码)?|单元(?:测试)?|集成(?:测试)?|回归测试|开发测试|开发验证|"
+    r"代码|分支|语句|行|函数|方法)\s*覆盖率"
+    r"|(?:unit(?:\s+test)?|integration(?:\s+test)?|regression\s+test|"
+    r"test|code|branch|statement|line|function|method)\s+coverage"
+    r")",
+    re.IGNORECASE,
+)
+_COVERAGE_PERCENT_PATTERN = re.compile(
+    r"覆盖率\s*(?:达到|为|至少|不低于|不少于|应达到|>=|>|:|：)?\s*"
+    r"\d+(?:\.\d+)?\s*[%％]",
+    re.IGNORECASE,
+)
+_BUSINESS_COVERAGE_CONTEXT_PATTERN = re.compile(
+    r"(?:细纲|事件|硬约束|软约束|伏笔|实体|设定|情节|场景|知识|事实|正文|"
+    r"作品|业务|候选|规划|大纲|章)[^。\n]{0,24}覆盖率"
+    r"|覆盖率[^。\n]{0,24}(?:细纲|事件|硬约束|软约束|伏笔|实体|设定|情节|"
+    r"场景|知识|事实|正文|作品|业务|候选|规划|大纲|章)",
+    re.IGNORECASE,
+)
+
+# 这些规则只针对运行时 Skill 文档中的明显开发验证残留；业务正文中的普通“检查”
+# 等词不在扫描范围内。规则保持确定性，方便报告中的 code 作为机械门禁输入。
+_DEVELOPMENT_HEADING_PATTERN = re.compile(r"^\s*(#{1,6})\s*(.*?)\s*$")
+_DEVELOPMENT_SECTION_TITLE_PATTERN = re.compile(
+    r"(?:自测|测试|开发验证|离线验证|开发测试|离线测试)",
+    re.IGNORECASE,
+)
+_DEVELOPMENT_COMMAND_PATTERN = re.compile(
+    r"(?<![A-Za-z0-9_])check_contract\.py(?![A-Za-z0-9_])",
+    re.IGNORECASE,
+)
+_POLLUTION_RULES = (
+    (
+        "pytest_reference",
+        re.compile(r"(?<![A-Za-z0-9_])pytest(?![A-Za-z0-9_])", re.IGNORECASE),
+        "发现 pytest 测试入口",
+    ),
+    (
+        "unittest_reference",
+        re.compile(r"(?<![A-Za-z0-9_])unittest(?![A-Za-z0-9_])", re.IGNORECASE),
+        "发现 unittest 测试入口",
+    ),
+    (
+        "test_file_reference",
+        re.compile(
+            r"(?:\btest_[A-Za-z0-9][A-Za-z0-9_.-]*\.py\b"
+            r"|\b[A-Za-z0-9][A-Za-z0-9_.-]*_test\.py\b"
+            r"|test_\*\.py|\*_test\.py)",
+            re.IGNORECASE,
+        ),
+        "发现测试文件路径或文件名模式",
+    ),
+    (
+        "test_pass_declaration",
+        re.compile(
+            r"(?:"
+            r"(?:所有|全部)?\s*测试\s*(?:已|已经|均|都|全部)?\s*(?:通过|成功)"
+            r"|(?:通过|成功)\s*测试"
+            r"|(?:all|the)\s+tests?\s+(?:all\s+)?(?:pass(?:ed|es|ing)?|succeed(?:ed|s|ing)?|successful)"
+            r"|tests?\s+(?:all\s+)?(?:pass(?:ed|es|ing)?|succeed(?:ed|s|ing)?|successful)"
+            r")",
+            re.IGNORECASE,
+        ),
+        "发现测试通过声明",
+    ),
+)
+
+
+def _resolve_root(root: str | Path) -> Path:
+    """把审计根目录解析为绝对路径，但不要求它已经存在。"""
+
+    path = Path(root).expanduser()
+    if not path.is_absolute():
+        path = Path.cwd() / path
+    return path.resolve()
+
+
+def _relative_path(path: Path, root: Path) -> str:
+    """返回报告中的稳定 POSIX 相对路径。"""
+
+    try:
+        return path.resolve().relative_to(root).as_posix()
+    except ValueError:
+        return path.resolve().as_posix()
+
+
+def _issue(
+    code: str,
+    message: str,
+    *,
+    path: Optional[str] = None,
+    line: Optional[int] = None,
+    **details: Any,
+) -> dict[str, Any]:
+    result: dict[str, Any] = {
+        "code": code,
+        "severity": "error",
+        "message": message,
+    }
+    if path is not None:
+        result["path"] = path
+    if line is not None:
+        result["line"] = line
+    result.update(details)
+    return result
+
+
+def _unquote_scalar(value: str) -> str:
+    """解析 Skill frontmatter 中足够支持 name 的简单标量。"""
+
+    value = value.strip()
+    if len(value) >= 2 and value[0] == value[-1] == '"':
+        try:
+            parsed = json.loads(value)
+        except json.JSONDecodeError:
+            return value[1:-1]
+        return parsed if isinstance(parsed, str) else value
+    if len(value) >= 2 and value[0] == value[-1] == "'":
+        return value[1:-1].replace("''", "'")
+    # 允许常见的 YAML 行尾注释，不引入 YAML 依赖。
+    return re.sub(r"\s+#.*$", "", value).strip()
+
+
+def _parse_frontmatter(text: str) -> tuple[Optional[dict[str, str]], Optional[int], Optional[str]]:
+    """读取 frontmatter 的简单键值视图。
+
+    返回值为 ``(fields, closing_line, error_code)``。该解析器只需要识别 name，
+    不试图替代完整 YAML 解析器；不合法或缺失的边界会明确进入失败报告。
+    """
+
+    lines = text.splitlines()
+    if not lines or lines[0].lstrip("\ufeff").strip() != "---":
+        return None, None, "frontmatter_missing"
+
+    closing_line: Optional[int] = None
+    for index in range(1, len(lines)):
+        if lines[index].strip() in {"---", "..."}:
+            closing_line = index + 1
+            break
+    if closing_line is None:
+        return None, None, "frontmatter_unclosed"
+
+    fields: dict[str, str] = {}
+    for line in lines[1 : closing_line - 1]:
+        match = _FRONTMATTER_NAME.match(line)
+        if match:
+            fields["name"] = _unquote_scalar(match.group(1))
+    return fields, closing_line, None
+
+
+def _snippet(line: str) -> str:
+    compact = line.strip()
+    return compact if len(compact) <= 200 else compact[:197] + "..."
+
+
+def _is_development_coverage_reference(line: str) -> bool:
+    """只把明确落在开发测试语境中的 coverage 文字判为污染。"""
+
+    if _COVERAGE_TOOL_PATTERN.search(line) or _EXPLICIT_TEST_COVERAGE_PATTERN.search(line):
+        return True
+    if not _COVERAGE_PERCENT_PATTERN.search(line):
+        return False
+    return not _BUSINESS_COVERAGE_CONTEXT_PATTERN.search(line)
+
+
+def _inspect_skill_file(path: Path, root: Path) -> tuple[dict[str, Any], list[dict[str, Any]]]:
+    relative = _relative_path(path, root)
+    entry: dict[str, Any] = {
+        "directory": path.parent.name,
+        "skill_path": relative,
+        "name": None,
+        "frontmatter_present": False,
+    }
+    issues: list[dict[str, Any]] = []
+
+    try:
+        text = path.read_text(encoding="utf-8")
+    except (OSError, UnicodeError) as exc:
+        issues.append(
+            _issue(
+                "skill_file_unreadable",
+                "无法读取 SKILL.md",
+                path=relative,
+                error=str(exc),
+            )
+        )
+        return entry, issues
+
+    if not text.strip():
+        issues.append(
+            _issue(
+                "empty_skill_file",
+                "SKILL.md 为空",
+                path=relative,
+                line=1,
+            )
+        )
+        return entry, issues
+
+    fields, closing_line, frontmatter_error = _parse_frontmatter(text)
+    if frontmatter_error is not None:
+        line = 1 if frontmatter_error == "frontmatter_missing" else max(1, len(text.splitlines()))
+        messages = {
+            "frontmatter_missing": "缺少以 --- 开始的 frontmatter",
+            "frontmatter_unclosed": "frontmatter 未闭合",
+        }
+        issues.append(
+            _issue(
+                frontmatter_error,
+                messages[frontmatter_error],
+                path=relative,
+                line=line,
+            )
+        )
+    else:
+        entry["frontmatter_present"] = True
+        name = fields.get("name") if fields is not None else None
+        entry["name"] = name or None
+        if not name:
+            issues.append(
+                _issue(
+                    "frontmatter_name_missing",
+                    "frontmatter 缺少非空 name",
+                    path=relative,
+                    line=1,
+                )
+            )
+        elif name != path.parent.name:
+            issues.append(
+                _issue(
+                    "name_directory_mismatch",
+                    "frontmatter name 与 Skill 目录名不一致",
+                    path=relative,
+                    line=2,
+                    expected=path.parent.name,
+                    actual=name,
+                )
+            )
+
+    development_section_level: Optional[int] = None
+    for line_number, line in enumerate(text.splitlines(), start=1):
+        heading_match = _DEVELOPMENT_HEADING_PATTERN.match(line)
+        if heading_match:
+            heading_level = len(heading_match.group(1))
+            heading_title = re.sub(r"\s+#+\s*$", "", heading_match.group(2)).strip()
+            if (
+                development_section_level is not None
+                and heading_level <= development_section_level
+            ):
+                development_section_level = None
+            if (
+                2 <= heading_level <= 6
+                and _DEVELOPMENT_SECTION_TITLE_PATTERN.search(heading_title)
+            ):
+                development_section_level = heading_level
+                issues.append(
+                    _issue(
+                        "development_test_heading",
+                        "发现开发测试章节标题",
+                        path=relative,
+                        line=line_number,
+                        snippet=_snippet(line),
+                    )
+                )
+
+        for code, pattern, message in _POLLUTION_RULES:
+            if pattern.search(line):
+                issues.append(
+                    _issue(
+                        code,
+                        message,
+                        path=relative,
+                        line=line_number,
+                        snippet=_snippet(line),
+                    )
+                )
+        if (
+            development_section_level is not None
+            and _DEVELOPMENT_COMMAND_PATTERN.search(line)
+        ):
+            issues.append(
+                _issue(
+                    "development_test_command_reference",
+                    "发现开发测试章节中的合同测试命令",
+                    path=relative,
+                    line=line_number,
+                    snippet=_snippet(line),
+                )
+            )
+        if _is_development_coverage_reference(line):
+            issues.append(
+                _issue(
+                    "coverage_reference",
+                    "发现开发测试语境中的覆盖率声明或工具名",
+                    path=relative,
+                    line=line_number,
+                    snippet=_snippet(line),
+                )
+            )
+
+    return entry, issues
+
+
+def _scan_skill_directories(root: Path) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
+    skills_root = root / SKILLS_ROOT
+    relative_skills_root = _relative_path(skills_root, root)
+    entries: list[dict[str, Any]] = []
+    issues: list[dict[str, Any]] = []
+
+    if not skills_root.exists():
+        issues.append(
+            _issue(
+                "skills_directory_missing",
+                "运行时 Skill 目录不存在",
+                path=relative_skills_root,
+            )
+        )
+        return entries, issues
+    if not skills_root.is_dir():
+        issues.append(
+            _issue(
+                "skills_directory_not_directory",
+                "运行时 Skill 路径不是目录",
+                path=relative_skills_root,
+            )
+        )
+        return entries, issues
+
+    try:
+        children = sorted(skills_root.iterdir(), key=lambda item: item.name)
+    except OSError as exc:
+        issues.append(
+            _issue(
+                "skills_directory_unreadable",
+                "无法列出运行时 Skill 目录",
+                path=relative_skills_root,
+                error=str(exc),
+            )
+        )
+        return entries, issues
+
+    directories = [child for child in children if child.is_dir()]
+    if not directories:
+        issues.append(
+            _issue(
+                "no_skills_found",
+                "运行时 Skill 目录中未发现 Skill 子目录",
+                path=relative_skills_root,
+            )
+        )
+        return entries, issues
+
+    for directory in directories:
+        skill_file = directory / "SKILL.md"
+        relative_skill_file = _relative_path(skill_file, root)
+        disk_entry: dict[str, Any] = {
+            "directory": directory.name,
+            "skill_path": relative_skill_file,
+            "name": None,
+            "frontmatter_present": False,
+        }
+        if not skill_file.exists():
+            entries.append(disk_entry)
+            issues.append(
+                _issue(
+                    "missing_skill_file",
+                    "Skill 目录缺少 SKILL.md",
+                    path=relative_skill_file,
+                )
+            )
+            continue
+        if not skill_file.is_file():
+            entries.append(disk_entry)
+            issues.append(
+                _issue(
+                    "skill_file_not_file",
+                    "SKILL.md 路径不是普通文件",
+                    path=relative_skill_file,
+                )
+            )
+            continue
+
+        entry, file_issues = _inspect_skill_file(skill_file, root)
+        entries.append(entry)
+        issues.extend(file_issues)
+
+    return entries, issues
+
+
+def _resolve_manifest_path(root: Path, manifest: str | Path | None) -> Path:
+    path = DEFAULT_MANIFEST_PATH if manifest is None else Path(manifest).expanduser()
+    if not path.is_absolute():
+        path = root / path
+    return path.resolve()
+
+
+def _normalise_manifest_skill_path(
+    root: Path, value: str
+) -> tuple[Optional[str], Optional[Path]]:
+    declared = Path(value)
+    if declared.is_absolute():
+        return None, None
+    candidate = (root / declared).resolve()
+    try:
+        relative = candidate.relative_to(root).as_posix()
+    except ValueError:
+        return None, None
+    return relative, candidate
+
+
+def _validate_manifest(
+    root: Path,
+    entries: Sequence[dict[str, Any]],
+    manifest_path: Path,
+) -> tuple[dict[str, Any], list[dict[str, Any]]]:
+    """验证 Skill 登记表，并要求它与磁盘目录一一对应。"""
+
+    relative_manifest = _relative_path(manifest_path, root)
+    metadata: dict[str, Any] = {
+        "path": relative_manifest,
+        "loaded": False,
+        "skills_declared": 0,
+    }
+    issues: list[dict[str, Any]] = []
+
+    if not manifest_path.exists():
+        issues.append(
+            _issue(
+                "manifest_missing",
+                "Skill manifest 不存在",
+                path=relative_manifest,
+            )
+        )
+        return metadata, issues
+    if not manifest_path.is_file():
+        issues.append(
+            _issue(
+                "manifest_not_file",
+                "Skill manifest 路径不是普通文件",
+                path=relative_manifest,
+            )
+        )
+        return metadata, issues
+
+    try:
+        manifest_text = manifest_path.read_text(encoding="utf-8")
+        manifest = json.loads(manifest_text)
+    except UnicodeError as exc:
+        issues.append(
+            _issue(
+                "manifest_unreadable",
+                "无法按 UTF-8 读取 Skill manifest",
+                path=relative_manifest,
+                error=str(exc),
+            )
+        )
+        return metadata, issues
+    except OSError as exc:
+        issues.append(
+            _issue(
+                "manifest_unreadable",
+                "无法读取 Skill manifest",
+                path=relative_manifest,
+                error=str(exc),
+            )
+        )
+        return metadata, issues
+    except json.JSONDecodeError as exc:
+        issues.append(
+            _issue(
+                "manifest_invalid_json",
+                "Skill manifest 不是合法 JSON",
+                path=relative_manifest,
+                line=exc.lineno,
+                column=exc.colno,
+                error=exc.msg,
+            )
+        )
+        return metadata, issues
+
+    metadata["loaded"] = True
+    if not isinstance(manifest, dict):
+        issues.append(
+            _issue(
+                "manifest_invalid_structure",
+                "Skill manifest 顶层必须是对象",
+                path=relative_manifest,
+            )
+        )
+        return metadata, issues
+
+    raw_skills = manifest.get("skills")
+    if not isinstance(raw_skills, list):
+        issues.append(
+            _issue(
+                "manifest_skills_invalid",
+                "Skill manifest.skills 必须是数组",
+                path=relative_manifest,
+            )
+        )
+        return metadata, issues
+
+    metadata["skills_declared"] = len(raw_skills)
+    disk_paths = {
+        str(entry["skill_path"])
+        for entry in entries
+        if isinstance(entry.get("skill_path"), str)
+    }
+    declared_paths: set[str] = set()
+    path_occurrences: dict[str, list[int]] = {}
+    name_occurrences: dict[str, list[int]] = {}
+    skills_root = (root / SKILLS_ROOT).resolve()
+
+    for index, item in enumerate(raw_skills):
+        if not isinstance(item, dict):
+            issues.append(
+                _issue(
+                    "manifest_entry_invalid",
+                    "Skill manifest 条目必须是对象",
+                    path=relative_manifest,
+                    entry_index=index,
+                )
+            )
+            continue
+
+        name = item.get("name")
+        if (
+            not isinstance(name, str)
+            or not name.strip()
+            or "/" in name
+            or "\\" in name
+            or name in {".", ".."}
+        ):
+            issues.append(
+                _issue(
+                    "manifest_name_invalid",
+                    "manifest 条目的 name 必须是非空目录名",
+                    path=relative_manifest,
+                    entry_index=index,
+                    actual=name,
+                )
+            )
+        else:
+            name_occurrences.setdefault(name, []).append(index)
+
+        contract_owner = item.get("contract_owner")
+        if not isinstance(contract_owner, str) or not contract_owner.strip():
+            issues.append(
+                _issue(
+                    "manifest_contract_owner_invalid",
+                    "manifest 条目的 contract_owner 必须非空",
+                    path=relative_manifest,
+                    entry_index=index,
+                )
+            )
+
+        collaborations = item.get("collaborates_with")
+        if (
+            not isinstance(collaborations, list)
+            or any(not isinstance(value, str) or not value.strip() for value in collaborations)
+        ):
+            issues.append(
+                _issue(
+                    "manifest_collaborates_with_invalid",
+                    "manifest 条目的 collaborates_with 必须是非空字符串数组",
+                    path=relative_manifest,
+                    entry_index=index,
+                )
+            )
+
+        declared_skill_path = item.get("skill_path")
+        if not isinstance(declared_skill_path, str) or not declared_skill_path.strip():
+            issues.append(
+                _issue(
+                    "manifest_skill_path_invalid",
+                    "manifest 条目的 skill_path 必须是非空相对路径",
+                    path=relative_manifest,
+                    entry_index=index,
+                    actual=declared_skill_path,
+                )
+            )
+            continue
+
+        normalised_path, candidate = _normalise_manifest_skill_path(
+            root, declared_skill_path
+        )
+        if normalised_path is None or candidate is None:
+            issues.append(
+                _issue(
+                    "manifest_skill_path_invalid",
+                    "manifest 条目的 skill_path 必须位于项目根目录内",
+                    path=relative_manifest,
+                    entry_index=index,
+                    actual=declared_skill_path,
+                )
+            )
+            continue
+
+        declared_paths.add(normalised_path)
+        path_occurrences.setdefault(normalised_path, []).append(index)
+        if candidate != candidate.parent / "SKILL.md" or not candidate.is_relative_to(skills_root):
+            issues.append(
+                _issue(
+                    "manifest_skill_path_invalid",
+                    "manifest 条目的 skill_path 必须指向 .claude/skills/*/SKILL.md",
+                    path=normalised_path,
+                    entry_index=index,
+                )
+            )
+
+        if not candidate.exists():
+            issues.append(
+                _issue(
+                    "manifest_skill_path_missing",
+                    "manifest 登记的 Skill 路径不存在",
+                    path=normalised_path,
+                    entry_index=index,
+                )
+            )
+        elif not candidate.is_file():
+            issues.append(
+                _issue(
+                    "manifest_skill_path_not_file",
+                    "manifest 登记的 Skill 路径不是普通文件",
+                    path=normalised_path,
+                    entry_index=index,
+                )
+            )
+
+        directory_name = candidate.parent.name
+        if isinstance(name, str) and name.strip() and name != directory_name:
+            issues.append(
+                _issue(
+                    "manifest_name_directory_mismatch",
+                    "manifest name 与 skill_path 的目录名不一致",
+                    path=normalised_path,
+                    entry_index=index,
+                    expected=directory_name,
+                    actual=name,
+                )
+            )
+        if isinstance(name, str) and name.strip():
+            expected_path = (root / SKILLS_ROOT / name / "SKILL.md").resolve()
+            expected_relative = _relative_path(expected_path, root)
+            if normalised_path != expected_relative:
+                issues.append(
+                    _issue(
+                        "manifest_skill_path_mismatch",
+                        "manifest skill_path 与 name 对应的 Skill 路径不一致",
+                        path=normalised_path,
+                        entry_index=index,
+                        expected=expected_relative,
+                        actual=normalised_path,
+                    )
+                )
+
+    for path, indexes in sorted(path_occurrences.items()):
+        if len(indexes) > 1:
+            issues.append(
+                _issue(
+                    "manifest_duplicate_skill_path",
+                    "manifest 重复登记同一个 Skill 路径",
+                    path=path,
+                    entry_indexes=indexes,
+                )
+            )
+    for name, indexes in sorted(name_occurrences.items()):
+        if len(indexes) > 1:
+            issues.append(
+                _issue(
+                    "manifest_duplicate_name",
+                    "manifest 重复登记同一个 Skill name",
+                    path=relative_manifest,
+                    name=name,
+                    entry_indexes=indexes,
+                )
+            )
+
+    for path in sorted(disk_paths - declared_paths):
+        issues.append(
+            _issue(
+                "manifest_skill_missing",
+                "磁盘 Skill 未在 manifest 中登记",
+                path=path,
+            )
+        )
+    for path in sorted(declared_paths - disk_paths):
+        issues.append(
+            _issue(
+                "manifest_skill_extra",
+                "manifest 登记了磁盘中不存在的额外 Skill",
+                path=path,
+            )
+        )
+
+    return metadata, issues
+
+
+def _duplicate_name_issues(entries: Iterable[dict[str, Any]]) -> list[dict[str, Any]]:
+    by_name: dict[str, list[str]] = {}
+    for entry in entries:
+        name = entry.get("name")
+        if isinstance(name, str) and name:
+            by_name.setdefault(name, []).append(str(entry["skill_path"]))
+
+    issues: list[dict[str, Any]] = []
+    for name in sorted(by_name):
+        paths = sorted(by_name[name])
+        if len(paths) > 1:
+            issues.append(
+                _issue(
+                    "duplicate_name",
+                    "多个 Skill 使用相同的 frontmatter name",
+                    path=paths[0],
+                    name=name,
+                    paths=paths,
+                )
+            )
+    return issues
+
+
+def _summarize(issues: Sequence[dict[str, Any]], skill_count: int) -> dict[str, Any]:
+    counts = Counter(str(issue["code"]) for issue in issues)
+    return {
+        "status": "passed" if not issues else "failed",
+        "skills_scanned": skill_count,
+        "issue_count": len(issues),
+        "issue_codes": {code: counts[code] for code in sorted(counts)},
+    }
+
+
+def audit_skills(
+    root: str | Path = ".", manifest: str | Path | None = None
+) -> dict[str, Any]:
+    """审计 Skill 文档与 manifest，并返回 JSON 可序列化报告。"""
+
+    root_path = _resolve_root(root)
+    manifest_path = _resolve_manifest_path(root_path, manifest)
+    report: dict[str, Any] = {
+        "schema_version": SCHEMA_VERSION,
+        "root": str(root_path),
+        "skills_root": SKILLS_ROOT.as_posix(),
+        "manifest": {
+            "path": _relative_path(manifest_path, root_path),
+            "loaded": False,
+        },
+        "skills_scanned": 0,
+        "skills": [],
+        "issues": [],
+    }
+
+    if not root_path.exists():
+        report["issues"] = [
+            _issue("root_missing", "审计根目录不存在", path=".")
+        ]
+    elif not root_path.is_dir():
+        report["issues"] = [
+            _issue("root_not_directory", "审计根路径不是目录", path=".")
+        ]
+    else:
+        entries, issues = _scan_skill_directories(root_path)
+        issues.extend(_duplicate_name_issues(entries))
+        manifest_info, manifest_issues = _validate_manifest(
+            root_path, entries, manifest_path
+        )
+        issues.extend(manifest_issues)
+        report["manifest"] = manifest_info
+        report["skills"] = entries
+        report["skills_scanned"] = len(entries)
+        report["issues"] = issues
+
+    report["summary"] = _summarize(report["issues"], report["skills_scanned"])
+    report["issue_count"] = len(report["issues"])
+    report["ok"] = not report["issues"]
+    report["status"] = "passed" if report["ok"] else "failed"
+    return report
+
+
+def human_summary(report: dict[str, Any]) -> str:
+    """把结构化报告渲染为简洁的人读摘要。"""
+
+    status = "通过" if report["ok"] else "失败"
+    lines = [
+        f"Skill 静态审计：{status}",
+        f"扫描 Skill：{report['skills_scanned']} 个；问题：{report['issue_count']} 个",
+    ]
+    issues = report["issues"]
+    if not issues:
+        lines.append("未发现 frontmatter、目录命名、空文件或开发测试污染问题。")
+        return "\n".join(lines)
+
+    lines.append("问题明细：")
+    for issue in issues:
+        location = str(issue.get("path", "<root>"))
+        if issue.get("line") is not None:
+            location += f":{issue['line']}"
+        lines.append(f"- {location} [{issue['code']}] {issue['message']}")
+    return "\n".join(lines)
+
+
+def _build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        description="只读扫描项目运行时 .claude/skills/*/SKILL.md 的静态卫生审计器。"
+    )
+    parser.add_argument(
+        "--root",
+        default=".",
+        help="项目根目录，默认为当前目录",
+    )
+    parser.add_argument(
+        "--manifest",
+        default=None,
+        help="Skill manifest 路径；相对路径按项目根目录解析，默认 harness/manifests/skills.json",
+    )
+    parser.add_argument(
+        "--json",
+        action="store_true",
+        help="输出结构化 JSON 报告",
+    )
+    parser.add_argument(
+        "--quiet",
+        action="store_true",
+        help="不输出人读摘要；与 --json 同用时仍输出 JSON",
+    )
+    return parser
+
+
+def main(argv: Optional[Sequence[str]] = None) -> int:
+    args = _build_parser().parse_args(argv)
+    report = audit_skills(args.root, args.manifest)
+    if args.json:
+        print(json.dumps(report, ensure_ascii=False, indent=2, sort_keys=True))
+    elif not args.quiet:
+        print(human_summary(report))
+    return 0 if report["ok"] else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())
--- a/harness/specs/skill-testing.md
+++ b/harness/specs/skill-testing.md
@ -0,0 +1,115 @@
+# Skill 测试与评测规范
+
+> 状态：生效规范
+> 适用范围：`agent-example/.claude/skills/*/SKILL.md`、其确定性工具和 Skill 行为评测。
+> Owner：`agent-example/harness/`
+> 本文件是开发与评测支架规范，不属于任何 Skill 的运行时提示词。
+
+## 1. 三层边界
+
+### 1.1 Skill 运行时指令
+
+`SKILL.md` 只描述 Agent 执行该 Skill 时必须知道的稳定合同：唯一目的、消费者、输入、输出、schema/version、允许读取、允许副作用、禁止动作、数据库读写、授权、预算、raw 和失败关闭规则，以及实际执行所需的业务工具与流程。
+
+`SKILL.md` 不承载开发者测试说明、测试文件路径、测试框架命令、夹具细节或测试结论。
+
+禁止在运行时 Skill 文本中出现：
+
+- `## 自测`、`## 测试`、`## 开发验证` 等开发测试章节；
+- `test_*.py`、`*_test.py`、pytest、unittest 等实现测试入口；
+- “本测试通过”“测试覆盖 N 项”等开发证据；
+- 把测试文件、测试常量或测试输出说成事实源或业务合同。
+
+评测 Skill 的生产/评测执行命令可以保留，但必须是用户请求该评测时实际执行的业务步骤，而不是验证实现代码的开发测试。
+
+### 1.2 确定性工具验证
+
+工具验证回答“Python、数据库或状态机实现是否守住机械合同”，不回答模型是否写得好，也不证明 Skill 的自然语言指令有效。
+
+- 单元测试、离线合同测试和数据库集成测试属于这一层；
+- 项目运行时 Skill 的实现测试统一放在 `tests/skills/<skill>/`，领域包保留自己的 `tests/`；两者都必须在 harness 的开发验证清单中登记；
+- 测试通过只能证明列出的机械行为，不得升级为模型质量或产品可用性结论；
+- 真实数据库、网络、模型和额度探针必须显式标记并单独授权。
+
+### 1.3 Skill 行为评测
+
+行为评测回答“把 Skill 内容交给 Agent 后，Agent 是否按合同行动”。评测由 harness 从外部驱动，不能由 `SKILL.md` 自己宣布通过。
+
+行为评测至少区分：
+
+- 正向触发：应该使用该 Skill 的任务；
+- 负向触发：不应该使用该 Skill 的任务；
+- 输入缺失与越界；
+- 输出合同与失败关闭；
+- 关键禁止动作；
+- 多样例稳定性和已知混淆项。
+
+行为评测结果必须保留输入、Skill 版本指纹、模型/运行配置、输出摘要、判定证据和失败原因。小样本合同回放不等于文学质量、通用效果或生产完成。
+
+## 2. Harness 职责
+
+Harness 只负责从外部发现、运行、收集和裁决测试/评测证据：
+
+1. 扫描运行时 Skill 文档中的开发测试污染；
+2. 登记和分类确定性工具测试、集成测试、fake pipeline 测试和真实探针；
+3. 驱动 Skill 行为评测案例；
+4. 生成带证据类型和失败原因的结构化报告；
+5. 防止空跑、静默跳过、宽泛异常吞错和证据级别越权。
+
+Harness 不负责：
+
+- 修改 `SKILL.md` 或业务代码；
+- 判断文学质量；
+- 用字符串出现证明自然语言合同有效；
+- 把 fake、离线回放或小样本结果升级成生产结论。
+
+## 3. 证据分级
+
+从低到高仅表示证据类型，不允许自动跨级：
+
+1. 静态结构检查：frontmatter、路径、禁用污染模式；
+2. 确定性离线测试：纯函数、schema、状态机和失败分支；
+3. 真实依赖集成测试：PostgreSQL、文件系统或外部服务；
+4. Skill 行为评测：外部 Agent/模型运行与结构化裁决；
+5. 人工内容评审：文学质量、声音和语义效果。
+
+任一层通过都不能代替更高层证据。没有行为评测，不得声称 Skill 内容有效；没有真实依赖证据，不得声称生产链路可用。
+
+## 4. 改动门禁
+
+### 只改运行时合同
+
+- 通过 harness 的运行时文档污染扫描；
+- 更新受影响的行为评测案例，或记录只改措辞、不改变行为的理由；
+- 不新增把测试细节塞回 `SKILL.md` 的说明。
+
+### 只改确定性工具
+
+- 更新针对变更不变量的离线测试；
+- 对真实数据库、网络和模型测试单独标记；
+- 保留失败关闭和副作用边界证据。
+
+### 改变 Skill 意图或边界
+
+- 更新行为评测清单和正/负向案例；
+- 检查消费者、Agent 槽位和 `meta/chains` 映射；
+- 重新执行相关层级的验证，不得只跑 Python 单测。
+
+## 5. 明确禁止
+
+- 用 `assertIn` 检查几个词出现，就宣称 Skill 内容正确；
+- 用测试文件或测试输出作为运行时事实源；
+- 用 `except Exception: pass` 把错误依赖、连接失败或实现错误当成预期拒绝；
+- 用任意总入口的“零测试”结果当成通过；
+- 把离线 fake、模型探针或小样本合同回放写成真实质量结论；
+- 为了让 harness 变绿而修改业务合同、降低断言或静默跳过测试。
+
+## 6. 完成定义
+
+本治理任务只有同时满足以下条件，才可称为完成：
+
+- 运行时 `SKILL.md` 不再携带开发测试说明；
+- harness 能机械发现并阻断明显的测试污染；
+- 实现测试、集成测试和 Skill 行为评测的证据类型可区分；
+- 改动范围内的回归测试真实执行，失败不会被吞掉；
+- 报告明确区分已验证事实、推断和未验证的模型质量假设。
--- a/harness/test_run_selected.py
+++ b/harness/test_run_selected.py
@ -0,0 +1,438 @@
+#!/usr/bin/env python3
+"""run_selected 的标准库离线回归测试。
+
+测试项目、manifest 和 .venv/bin/python 均在临时目录中生成，不连接数据库、网络或模型。
+"""
+
+from __future__ import annotations
+
+import io
+import json
+import stat
+import sys
+import tempfile
+import unittest
+from contextlib import redirect_stdout
+from pathlib import Path
+from typing import Any, Iterable
+
+try:
+    from .run_selected import main
+except ImportError:  # 允许直接执行 `.venv/bin/python harness/test_run_selected.py`
+    from run_selected import main
+
+
+class RunSelectedTests(unittest.TestCase):
+    def make_project(
+        self,
+        entries: Iterable[dict[str, Any]],
+        *,
+        generated_scope: str | None = None,
+    ) -> tuple[tempfile.TemporaryDirectory[str], Path]:
+        temporary = tempfile.TemporaryDirectory()
+        root = Path(temporary.name)
+        (root / ".venv" / "bin").mkdir(parents=True)
+        fake_python = root / ".venv" / "bin" / "python"
+        fake_python.write_text(
+            f"#!{sys.executable}\n"
+            "import pathlib\n"
+            "import sys\n"
+            "import time\n"
+            "script = pathlib.Path(sys.argv[1])\n"
+            "(pathlib.Path.cwd() / 'invocations.log').open('a', encoding='utf-8').write(script.name + '\\n')\n"
+            "if script.name == 'empty.py':\n"
+            "    raise SystemExit(0)\n"
+            "if script.name == 'skip.py':\n"
+            "    print('1 skipped')\n"
+            "    raise SystemExit(0)\n"
+            "if script.name == 'fail.py':\n"
+            "    print('failure stdout')\n"
+            "    print('failure stderr', file=sys.stderr)\n"
+            "    raise SystemExit(7)\n"
+            "if script.name == 'sleep.py':\n"
+            "    time.sleep(2)\n"
+            "if script.name == 'noisy.py':\n"
+            "    print('x' * 5000)\n"
+            "else:\n"
+            "    print('ran ' + script.name)\n",
+            encoding="utf-8",
+        )
+        fake_python.chmod(fake_python.stat().st_mode | stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH)
+
+        manifest_entries = []
+        for entry in entries:
+            entry_copy = dict(entry)
+            path = root / str(entry_copy["path"])
+            if entry_copy.pop("create", True):
+                path.parent.mkdir(parents=True, exist_ok=True)
+                path.write_text("# temporary test script\n", encoding="utf-8")
+            manifest_entries.append(entry_copy)
+
+        manifest_path = root / "manifest.json"
+        manifest_payload: dict[str, Any] = {
+            "schema_version": 1,
+            "entries": manifest_entries,
+        }
+        if generated_scope is not None:
+            manifest_payload["generated_scope"] = generated_scope
+        manifest_path.write_text(
+            json.dumps(manifest_payload, indent=2) + "\n",
+            encoding="utf-8",
+        )
+        self.addCleanup(temporary.cleanup)
+        return temporary, manifest_path
+
+    def invoke(self, root: Path, manifest: Path, *arguments: str) -> tuple[int, dict[str, Any]]:
+        output = io.StringIO()
+        with redirect_stdout(output):
+            return_code = main(
+                [
+                    "--root",
+                    str(root),
+                    "--manifest",
+                    str(manifest),
+                    *arguments,
+                    "--json",
+                ]
+            )
+        return return_code, json.loads(output.getvalue())
+
+    @staticmethod
+    def entry(
+        path: str,
+        *,
+        owner: str = "alpha",
+        kind: str = "tool_unit",
+        scope: str = "runtime_skill",
+        requires: list[str] | None = None,
+        create: bool = True,
+    ) -> dict[str, Any]:
+        return {
+            "path": path,
+            "owner_skill_or_domain": owner,
+            "kind": kind,
+            "scope": scope,
+            "requires": ["offline"] if requires is None else requires,
+            "create": create,
+        }
+
+    def test_selector_alias_and_intersection_are_reported_as_json(self) -> None:
+        _, manifest = self.make_project(
+            [
+                self.entry("pass.py", owner="alpha", kind="tool_unit"),
+                self.entry("other.py", owner="beta", kind="tool_contract"),
+            ]
+        )
+        root = manifest.parent
+
+        return_code, report = self.invoke(
+            root,
+            manifest,
+            "--skill",
+            "alpha",
+            "--kind",
+            "tool_unit",
+            "--path",
+            "pass.py",
+        )
+
+        self.assertEqual(return_code, 0)
+        self.assertEqual(report["status"], "passed")
+        self.assertEqual(report["selected_count"], 1)
+        self.assertEqual(report["entries"][0]["path"], "pass.py")
+        self.assertEqual(report["entries"][0]["returncode"], 0)
+        self.assertIn("ran pass.py", report["entries"][0]["stdout"])
+
+    def test_offline_dependency_is_blocked_without_running_child(self) -> None:
+        _, manifest = self.make_project(
+            [self.entry("db.py", requires=["offline", "postgresql"])]
+        )
+        root = manifest.parent
+
+        return_code, report = self.invoke(root, manifest, "--path", "db.py")
+
+        result = report["entries"][0]
+        self.assertEqual(return_code, 1)
+        self.assertEqual(report["status"], "failed")
+        self.assertEqual(result["status"], "blocked_dependency")
+        self.assertEqual(result["requires"], ["offline", "postgresql"])
+        self.assertEqual(result["blocked_requires"], ["postgresql"])
+        self.assertFalse((root / "invocations.log").exists())
+
+    def test_allow_requires_runs_and_still_reports_dependency(self) -> None:
+        _, manifest = self.make_project(
+            [self.entry("db.py", requires=["offline", "postgresql"])]
+        )
+        root = manifest.parent
+
+        return_code, report = self.invoke(
+            root,
+            manifest,
+            "--path",
+            "db.py",
+            "--allow-requires",
+            "postgresql",
+        )
+
+        result = report["entries"][0]
+        self.assertEqual(return_code, 0)
+        self.assertEqual(result["status"], "passed")
+        self.assertEqual(result["requires"], ["offline", "postgresql"])
+        self.assertEqual((root / "invocations.log").read_text(encoding="utf-8"), "db.py\n")
+
+    def test_nonzero_and_timeout_have_distinct_structured_statuses(self) -> None:
+        _, manifest = self.make_project(
+            [self.entry("fail.py"), self.entry("sleep.py")]
+        )
+        root = manifest.parent
+
+        return_code, report = self.invoke(
+            root,
+            manifest,
+            "--path",
+            "fail.py",
+            "--path",
+            "sleep.py",
+            "--timeout-seconds",
+            "0.5",
+        )
+
+        self.assertEqual(return_code, 1)
+        results = {entry["path"]: entry for entry in report["entries"]}
+        self.assertEqual(results["fail.py"]["status"], "failed")
+        self.assertEqual(results["fail.py"]["returncode"], 7)
+        self.assertIn("failure stdout", results["fail.py"]["stdout"])
+        self.assertIn("failure stderr", results["fail.py"]["stderr"])
+        self.assertEqual(results["sleep.py"]["status"], "timeout")
+        self.assertIsNone(results["sleep.py"]["returncode"])
+        self.assertEqual(results["sleep.py"]["error"]["code"], "timeout")
+
+    def test_no_matches_and_missing_path_are_not_silent(self) -> None:
+        _, manifest = self.make_project(
+            [
+                self.entry("missing.py", create=False),
+                self.entry("present.py"),
+            ]
+        )
+        root = manifest.parent
+
+        return_code, no_match = self.invoke(root, manifest, "--kind", "does_not_exist")
+        self.assertEqual(return_code, 1)
+        self.assertEqual(no_match["status"], "no_matches")
+        self.assertEqual(no_match["entries"], [])
+        self.assertEqual(no_match["issues"][0]["code"], "no_matches")
+
+        return_code, missing = self.invoke(root, manifest, "--path", "missing.py")
+        self.assertEqual(return_code, 1)
+        self.assertEqual(missing["entries"][0]["status"], "not_found")
+        self.assertEqual(missing["entries"][0]["error"]["code"], "path_not_found")
+
+    def test_invalid_manifest_and_selector_requirement_are_structured(self) -> None:
+        temporary = tempfile.TemporaryDirectory()
+        self.addCleanup(temporary.cleanup)
+        root = Path(temporary.name)
+        manifest = root / "manifest.json"
+        manifest.write_text("{broken", encoding="utf-8")
+
+        return_code, invalid_manifest = self.invoke(root, manifest, "--path", "x.py")
+        self.assertEqual(return_code, 1)
+        self.assertEqual(invalid_manifest["status"], "manifest_invalid")
+        self.assertEqual(invalid_manifest["issues"][0]["code"], "manifest_invalid_json")
+
+        valid_root, valid_manifest = self.make_project([self.entry("pass.py")])
+        return_code, invalid_selector = self.invoke(valid_root, valid_manifest)
+        self.assertEqual(return_code, 1)
+        self.assertEqual(invalid_selector["status"], "invalid_selector")
+        self.assertEqual(invalid_selector["issues"][0]["code"], "selector_required")
+
+    def test_all_offline_is_explicit_opt_in_and_output_is_summary_only(self) -> None:
+        _, manifest = self.make_project(
+            [
+                self.entry("pass.py"),
+                self.entry("db.py", requires=["postgresql"]),
+                self.entry("noisy.py"),
+            ]
+        )
+        root = manifest.parent
+
+        return_code, report = self.invoke(root, manifest, "--all-offline")
+
+        self.assertEqual(return_code, 1)
+        self.assertEqual(report["selected_count"], 3)
+        statuses = {entry["path"]: entry["status"] for entry in report["entries"]}
+        self.assertEqual(statuses["pass.py"], "passed")
+        self.assertEqual(statuses["db.py"], "blocked_dependency")
+        self.assertEqual(statuses["noisy.py"], "passed")
+        noisy = next(entry for entry in report["entries"] if entry["path"] == "noisy.py")
+        self.assertLess(len(noisy["stdout"]), 2001)
+        self.assertIn("output summary truncated", noisy["stdout"])
+
+    def test_generated_scope_rejects_unregistered_disk_asset(self) -> None:
+        _, manifest = self.make_project(
+            [self.entry("tests/skills/registered/test_registered.py")],
+            generated_scope="temporary test asset inventory",
+        )
+        root = manifest.parent
+        unregistered = root / "tests" / "skills" / "new" / "test_unregistered.py"
+        unregistered.parent.mkdir(parents=True)
+        unregistered.write_text("# unregistered test asset\n", encoding="utf-8")
+
+        return_code, report = self.invoke(root, manifest, "--all-offline")
+
+        self.assertEqual(return_code, 1)
+        self.assertEqual(report["status"], "manifest_invalid")
+        self.assertEqual(report["entries"], [])
+        self.assertEqual(
+            [
+                issue["path"]
+                for issue in report["issues"]
+                if issue["code"] == "manifest_test_asset_missing"
+            ],
+            ["tests/skills/new/test_unregistered.py"],
+        )
+        self.assertFalse((root / "invocations.log").exists())
+
+    def test_generated_scope_rejects_manifest_extra_asset(self) -> None:
+        _, manifest = self.make_project(
+            [
+                self.entry(
+                    "tests/skills/removed/test_removed.py",
+                    create=False,
+                )
+            ],
+            generated_scope="temporary test asset inventory",
+        )
+        root = manifest.parent
+
+        return_code, report = self.invoke(root, manifest, "--all-offline")
+
+        self.assertEqual(return_code, 1)
+        self.assertEqual(report["status"], "manifest_invalid")
+        self.assertEqual(report["entries"], [])
+        self.assertEqual(
+            [
+                issue["path"]
+                for issue in report["issues"]
+                if issue["code"] == "manifest_test_asset_extra"
+            ],
+            ["tests/skills/removed/test_removed.py"],
+        )
+        self.assertFalse((root / "invocations.log").exists())
+
+    def test_generated_scope_ignores_non_test_helpers_under_test_roots(self) -> None:
+        _, manifest = self.make_project(
+            [self.entry("tests/skills/registered/test_registered.py")],
+            generated_scope="temporary test asset inventory",
+        )
+        root = manifest.parent
+        registered_test = root / "tests" / "skills" / "registered" / "test_registered.py"
+        registered_test.write_text("def test_registered():\n    pass\n", encoding="utf-8")
+        (root / "tests" / "skills" / "registered" / "helper.py").write_text(
+            "VALUE = 1\n", encoding="utf-8"
+        )
+        (root / "humanization" / "tests" / "helper.py").parent.mkdir(
+            parents=True, exist_ok=True
+        )
+        (root / "humanization" / "tests" / "helper.py").write_text(
+            "VALUE = 2\n", encoding="utf-8"
+        )
+
+        return_code, report = self.invoke(root, manifest, "--all-offline")
+
+        self.assertEqual(return_code, 0)
+        self.assertEqual(report["status"], "passed")
+        self.assertEqual(report["manifest"]["test_assets_scanned"], 1)
+        self.assertEqual(report["entries"][0]["path"], "tests/skills/registered/test_registered.py")
+
+    def test_generated_scope_blocks_empty_and_print_only_scripts_before_child(self) -> None:
+        entries = [
+            self.entry("tests/skills/empty/test_empty.py"),
+            self.entry("tests/skills/print_only/test_print_only.py"),
+        ]
+        _, manifest = self.make_project(
+            entries,
+            generated_scope="temporary test asset inventory",
+        )
+        root = manifest.parent
+        (root / "tests" / "skills" / "empty" / "test_empty.py").write_text(
+            "", encoding="utf-8"
+        )
+        (root / "tests" / "skills" / "print_only" / "test_print_only.py").write_text(
+            "print('not a test')\n", encoding="utf-8"
+        )
+
+        return_code, report = self.invoke(root, manifest, "--all-offline")
+
+        self.assertEqual(return_code, 1)
+        self.assertEqual(report["status"], "failed")
+        results = {entry["path"]: entry for entry in report["entries"]}
+        for path in (
+            "tests/skills/empty/test_empty.py",
+            "tests/skills/print_only/test_print_only.py",
+        ):
+            self.assertEqual(results[path]["status"], "failed")
+            self.assertIsNone(results[path]["returncode"])
+            self.assertEqual(results[path]["error"]["code"], "test_shape_missing")
+        self.assertFalse((root / "invocations.log").exists())
+
+    def test_generated_scope_accepts_supported_python_shapes(self) -> None:
+        entries = [
+            self.entry("tests/skills/function/test_function.py"),
+            self.entry("tests/skills/class/test_class.py"),
+            self.entry("harness/test_main_entry.py"),
+        ]
+        _, manifest = self.make_project(
+            entries,
+            generated_scope="temporary test asset inventory",
+        )
+        root = manifest.parent
+        (root / "tests" / "skills" / "function" / "test_function.py").write_text(
+            "def test_function():\n    pass\n", encoding="utf-8"
+        )
+        (root / "tests" / "skills" / "class" / "test_class.py").write_text(
+            "import unittest\n\nclass Fixture(unittest.TestCase):\n    pass\n",
+            encoding="utf-8",
+        )
+        (root / "harness" / "test_main_entry.py").write_text(
+            "if __name__ == \"__main__\":\n    print(\"script\")\n",
+            encoding="utf-8",
+        )
+
+        return_code, report = self.invoke(root, manifest, "--all-offline")
+
+        self.assertEqual(return_code, 0)
+        self.assertEqual(report["status"], "passed")
+        self.assertEqual(
+            [entry["status"] for entry in report["entries"]],
+            ["passed", "passed", "passed"],
+        )
+        self.assertEqual(
+            (root / "invocations.log").read_text(encoding="utf-8").splitlines(),
+            ["test_function.py", "test_class.py", "test_main_entry.py"],
+        )
+
+    def test_zero_exit_without_execution_evidence_fails_closed(self) -> None:
+        _, manifest = self.make_project(
+            [self.entry("empty.py"), self.entry("skip.py")]
+        )
+        root = manifest.parent
+
+        return_code, report = self.invoke(
+            root,
+            manifest,
+            "--path",
+            "empty.py",
+            "--path",
+            "skip.py",
+        )
+
+        self.assertEqual(return_code, 1)
+        results = {entry["path"]: entry for entry in report["entries"]}
+        for path in ("empty.py", "skip.py"):
+            self.assertEqual(results[path]["status"], "failed")
+            self.assertEqual(results[path]["returncode"], 0)
+            self.assertEqual(results[path]["error"]["code"], "no_execution_evidence")
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)
--- a/harness/test_skill_harness.py
+++ b/harness/test_skill_harness.py
@ -0,0 +1,420 @@
+#!/usr/bin/env python3
+"""skill_harness 的纯标准库离线回归测试。
+
+所有夹具都在临时目录中构造，不读取当前仓库的 Skill，也不依赖数据库、网络或模型。
+"""
+
+from __future__ import annotations
+
+import io
+import json
+import tempfile
+import unittest
+from contextlib import redirect_stdout
+from pathlib import Path
+from typing import Optional
+
+try:
+    from .skill_harness import audit_skills, main
+except ImportError:  # 允许直接执行 `.venv/bin/python harness/test_skill_harness.py`
+    from skill_harness import audit_skills, main
+
+
+class SkillHarnessTests(unittest.TestCase):
+    def make_skill(
+        self,
+        root: Path,
+        directory: str,
+        *,
+        name: Optional[str] = None,
+        body: str = "# 合同\n\n只描述运行时行为。\n",
+    ) -> Path:
+        skill_dir = root / ".claude" / "skills" / directory
+        skill_dir.mkdir(parents=True, exist_ok=True)
+        skill_path = skill_dir / "SKILL.md"
+        frontmatter_name = directory if name is None else name
+        skill_path.write_text(
+            f"---\nname: {frontmatter_name}\ndescription: 离线夹具\n---\n{body}",
+            encoding="utf-8",
+        )
+        self.write_manifest(root)
+        return skill_path
+
+    def write_manifest(
+        self,
+        root: Path,
+        entries: Optional[list[dict[str, object]]] = None,
+        *,
+        path: Optional[Path] = None,
+    ) -> Path:
+        if entries is None:
+            entries = []
+            skills_root = root / ".claude" / "skills"
+            if skills_root.exists():
+                for directory in sorted(skills_root.iterdir(), key=lambda item: item.name):
+                    if directory.is_dir():
+                        entries.append(
+                            {
+                                "name": directory.name,
+                                "contract_owner": "测试夹具",
+                                "collaborates_with": [],
+                                "skill_path": (
+                                    Path(".claude")
+                                    / "skills"
+                                    / directory.name
+                                    / "SKILL.md"
+                                ).as_posix(),
+                            }
+                        )
+        manifest_path = path or root / "harness" / "manifests" / "skills.json"
+        manifest_path.parent.mkdir(parents=True, exist_ok=True)
+        manifest_path.write_text(
+            json.dumps(
+                {"schema_version": 1, "skills": entries},
+                ensure_ascii=False,
+                indent=2,
+            )
+            + "\n",
+            encoding="utf-8",
+        )
+        return manifest_path
+
+    def test_passing_project_is_independent_of_current_repository(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(root, "alpha")
+
+            report = audit_skills(root)
+
+            self.assertTrue(report["ok"])
+            self.assertEqual(report["status"], "passed")
+            self.assertEqual(report["skills_scanned"], 1)
+            self.assertEqual(report["issues"], [])
+
+    def test_pollution_is_reported_for_each_obvious_category(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(
+                root,
+                "polluted",
+                body=(
+                    "## 自测\n"
+                    "pytest -q\n"
+                    "import unittest\n"
+                    "run test_sample.py and sample_test.py\n"
+                    "覆盖率达到 100%。\n"
+                    "测试通过。\n"
+                ),
+            )
+
+            report = audit_skills(root)
+            codes = {issue["code"] for issue in report["issues"]}
+
+            self.assertFalse(report["ok"])
+            self.assertTrue(
+                {
+                    "development_test_heading",
+                    "pytest_reference",
+                    "unittest_reference",
+                    "test_file_reference",
+                    "coverage_reference",
+                    "test_pass_declaration",
+                }.issubset(codes)
+            )
+
+    def test_development_headings_support_multiple_levels(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(
+                root,
+                "nested-development-sections",
+                body=(
+                    "### 自测\n"
+                    "这里不能登记测试命令。\n"
+                    "## 离线验证\n"
+                    "这里也不能登记测试命令。\n"
+                ),
+            )
+
+            report = audit_skills(root)
+            headings = [
+                issue
+                for issue in report["issues"]
+                if issue["code"] == "development_test_heading"
+            ]
+
+            self.assertFalse(report["ok"])
+            self.assertEqual(len(headings), 2)
+            self.assertEqual([issue["line"] for issue in headings], [5, 7])
+
+    def test_check_contract_command_requires_development_section_context(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(
+                root,
+                "contextual-contract-command",
+                body=(
+                    "业务合同字段名可以写 check_contract.py。\n"
+                    "## 离线验证\n"
+                    "python check_contract.py\n"
+                    "## 业务说明\n"
+                    "普通业务 check_contract.py 不是测试入口。\n"
+                ),
+            )
+
+            report = audit_skills(root)
+            command_issues = [
+                issue
+                for issue in report["issues"]
+                if issue["code"] == "development_test_command_reference"
+            ]
+
+            self.assertFalse(report["ok"])
+            self.assertEqual(len(command_issues), 1)
+            self.assertEqual(command_issues[0]["line"], 7)
+
+    def test_business_coverage_terms_are_not_development_pollution(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(
+                root,
+                "business-coverage",
+                body=(
+                    "细纲覆盖率达到 100%。\n"
+                    "事件覆盖率为 80%。\n"
+                    "硬约束覆盖率 100%。\n"
+                ),
+            )
+
+            report = audit_skills(root)
+
+            self.assertTrue(report["ok"])
+            self.assertNotIn(
+                "coverage_reference",
+                {issue["code"] for issue in report["issues"]},
+            )
+
+    def test_development_coverage_terms_are_reported(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(
+                root,
+                "development-coverage",
+                body=(
+                    "coverage.py\n"
+                    "pytest-cov\n"
+                    "pytest --cov=harness\n"
+                    "coverage report\n"
+                    "测试覆盖率达到 90%。\n"
+                    "覆盖率达到 80%。\n"
+                ),
+            )
+
+            report = audit_skills(root)
+
+            self.assertFalse(report["ok"])
+            self.assertIn(
+                "coverage_reference",
+                {issue["code"] for issue in report["issues"]},
+            )
+
+    def test_frontmatter_name_must_match_skill_directory(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(root, "alpha", name="beta")
+
+            report = audit_skills(root)
+
+            self.assertFalse(report["ok"])
+            self.assertIn(
+                "name_directory_mismatch",
+                {issue["code"] for issue in report["issues"]},
+            )
+
+    def test_duplicate_frontmatter_names_are_reported(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(root, "alpha", name="shared")
+            self.make_skill(root, "beta", name="shared")
+
+            report = audit_skills(root)
+            duplicates = [
+                issue for issue in report["issues"] if issue["code"] == "duplicate_name"
+            ]
+
+            self.assertFalse(report["ok"])
+            self.assertEqual(len(duplicates), 1)
+            self.assertEqual(duplicates[0]["name"], "shared")
+            self.assertEqual(len(duplicates[0]["paths"]), 2)
+
+    def test_clean_manifest_can_be_selected_explicitly(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(root, "alpha")
+            manifest_path = self.write_manifest(
+                root,
+                path=root / "clean-manifest.json",
+            )
+            output = io.StringIO()
+
+            with redirect_stdout(output):
+                return_code = main(
+                    [
+                        "--root",
+                        str(root),
+                        "--manifest",
+                        str(manifest_path),
+                        "--json",
+                        "--quiet",
+                    ]
+                )
+
+            payload = json.loads(output.getvalue())
+            self.assertEqual(return_code, 0)
+            self.assertTrue(payload["ok"])
+            self.assertTrue(payload["manifest"]["loaded"])
+            self.assertEqual(payload["manifest"]["path"], "clean-manifest.json")
+
+    def test_error_manifest_fails_with_structured_issues(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(root, "alpha")
+            self.make_skill(root, "beta")
+            broken_manifest = self.write_manifest(
+                root,
+                entries=[
+                    {
+                        "name": "alpha",
+                        "contract_owner": "",
+                        "collaborates_with": ["上下文与知识检索", 3],
+                        "skill_path": ".claude/skills/beta/SKILL.md",
+                    },
+                    {
+                        "name": "alpha",
+                        "contract_owner": "测试夹具",
+                        "collaborates_with": [],
+                        "skill_path": ".claude/skills/beta/SKILL.md",
+                    },
+                    {
+                        "name": "extra",
+                        "contract_owner": "测试夹具",
+                        "collaborates_with": [],
+                        "skill_path": ".claude/skills/extra/SKILL.md",
+                    },
+                ],
+                path=root / "broken-manifest.json",
+            )
+            output = io.StringIO()
+
+            with redirect_stdout(output):
+                return_code = main(
+                    [
+                        "--root",
+                        str(root),
+                        "--manifest",
+                        str(broken_manifest),
+                        "--json",
+                    ]
+                )
+
+            payload = json.loads(output.getvalue())
+            codes = {issue["code"] for issue in payload["issues"]}
+            self.assertEqual(return_code, 1)
+            self.assertFalse(payload["ok"])
+            self.assertTrue(all(isinstance(issue, dict) for issue in payload["issues"]))
+            self.assertTrue(
+                {
+                    "manifest_contract_owner_invalid",
+                    "manifest_collaborates_with_invalid",
+                    "manifest_duplicate_skill_path",
+                    "manifest_duplicate_name",
+                    "manifest_skill_missing",
+                    "manifest_skill_extra",
+                    "manifest_skill_path_mismatch",
+                    "manifest_name_directory_mismatch",
+                    "manifest_skill_path_missing",
+                }.issubset(codes)
+            )
+
+    def test_missing_manifest_fails_closed(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(root, "alpha")
+            (root / "harness" / "manifests" / "skills.json").unlink()
+
+            report = audit_skills(root)
+
+            self.assertFalse(report["ok"])
+            self.assertIn(
+                "manifest_missing",
+                {issue["code"] for issue in report["issues"]},
+            )
+
+    def test_missing_skills_directory_fails_closed(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+
+            report = audit_skills(root)
+
+            self.assertFalse(report["ok"])
+            self.assertIn(
+                "skills_directory_missing",
+                {issue["code"] for issue in report["issues"]},
+            )
+
+    def test_skill_directory_without_skill_file_fails(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            (root / ".claude" / "skills" / "missing").mkdir(parents=True)
+
+            report = audit_skills(root)
+
+            self.assertFalse(report["ok"])
+            self.assertIn(
+                "missing_skill_file",
+                {issue["code"] for issue in report["issues"]},
+            )
+
+    def test_empty_skill_file_is_reported(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            skill_path = self.make_skill(root, "empty")
+            skill_path.write_text("", encoding="utf-8")
+
+            report = audit_skills(root)
+
+            self.assertFalse(report["ok"])
+            self.assertIn(
+                "empty_skill_file",
+                {issue["code"] for issue in report["issues"]},
+            )
+
+    def test_json_cli_and_exit_code(self) -> None:
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            self.make_skill(root, "alpha")
+            output = io.StringIO()
+
+            with redirect_stdout(output):
+                return_code = main(["--root", str(root), "--json", "--quiet"])
+
+            payload = json.loads(output.getvalue())
+            self.assertEqual(return_code, 0)
+            self.assertTrue(payload["ok"])
+            self.assertEqual(payload["skills_scanned"], 1)
+
+            (root / ".claude" / "skills" / "alpha" / "SKILL.md").write_text(
+                "---\nname: alpha\n---\n## 测试\n",
+                encoding="utf-8",
+            )
+            output = io.StringIO()
+            with redirect_stdout(output):
+                return_code = main(["--root", str(root), "--json"])
+
+            failed_payload = json.loads(output.getvalue())
+            self.assertEqual(return_code, 1)
+            self.assertFalse(failed_payload["ok"])
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)
--- a/humanization/research/20-project-skill-coverage.yaml
+++ b/humanization/research/20-project-skill-coverage.yaml
@ -73,7 +73,7 @@ capabilities:
    evidence_sources: [no-ai-slop, neuro-book, humanizer]
    owner_skill: prevent-ai-flavor
    implementation: .claude/skills/prevent-ai-flavor/scripts/prevent_ai_flavor.py
-    test: .claude/skills/assemble-context/scripts/test_assemble_writer_context.py
+    test: tests/skills/assemble-context/test_assemble_writer_context.py
    status: implemented
  - id: sf_snf_boundary_regression
    evidence_sources: [speak-human-tw, shuorenhua, neuro-book]
@ -98,14 +98,14 @@ capabilities:
    evidence_sources: [inkos, oh-story-claudecode, neuro-book]
    owner_skill: revise-ai-flavor
    implementation: humanization/src/deai/pipeline.py
-    test: .claude/skills/revise-ai-flavor/scripts/test_revise_ai_flavor.py
+    test: tests/skills/revise-ai-flavor/test_revise_ai_flavor.py
    status: implemented
    note: 当前以最多3轮合同、复扫和pairwise no_gain实现；跨轮最佳快照编排仍由上层负责
  - id: source_revalidation
    evidence_sources: [shuorenhua, neuro-book]
    owner_skill: capture-ai-flavor-cases
    implementation: .claude/skills/capture-ai-flavor-cases/scripts/capture_cases.py
-    test: .claude/skills/capture-ai-flavor-cases/scripts/test_capture_cases.py
+    test: tests/skills/capture-ai-flavor-cases/test_capture_cases.py
    status: implemented
  - id: case_to_rule_lifecycle
    evidence_sources: [unslop, shuorenhua, neuro-book]
--- a/.claude/skills/access-database/scripts/test_authorization_snapshot_ddl.py
+++ b/.claude/skills/access-database/scripts/test_authorization_snapshot_ddl.py
@ -6,7 +6,8 @@ import re
 import unittest


-DDL_PATH = pathlib.Path(__file__).resolve().parents[4] / "db" / "ddl" / "96-example参考作品授权快照.sql"
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+DDL_PATH = PROJECT_ROOT / "db" / "ddl" / "96-example参考作品授权快照.sql"


 class AuthorizationSnapshotDdlTest(unittest.TestCase):
--- a/.claude/skills/access-database/scripts/test_db_params.py
+++ b/.claude/skills/access-database/scripts/test_db_params.py
@ -1,14 +1,16 @@
 #!/usr/bin/env python3
 """db execparams 参数装载逻辑离线自测（不连库）。

-跑法（仓库根目录）：.venv/bin/python .claude/skills/access-database/scripts/test_db_params.py
+跑法（仓库根目录）：.venv/bin/python tests/skills/access-database/test_db_params.py
 """
 import io
 import json
-import os
 import sys
+from pathlib import Path

-sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+PROJECT_ROOT = Path(__file__).resolve().parents[3]
+SCRIPTS_DIR = PROJECT_ROOT / ".claude" / "skills" / "access-database" / "scripts"
+sys.path.insert(0, str(SCRIPTS_DIR))

 import click  # noqa: E402

--- a/.claude/skills/access-database/scripts/test_skill_catalog.py
+++ b/.claude/skills/access-database/scripts/test_skill_catalog.py
@ -2,9 +2,14 @@
 """Skill 目录命名与 frontmatter 一致性的离线测试。"""

 import pathlib
+import sys
 import tempfile
 import unittest

+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPTS_DIR = PROJECT_ROOT / ".claude" / "skills" / "access-database" / "scripts"
+sys.path.insert(0, str(SCRIPTS_DIR))
+
 from sync_agent_registry import validate_skill_catalog


--- a/.claude/skills/assemble-context/scripts/test_assemble_writer_context.py
+++ b/.claude/skills/assemble-context/scripts/test_assemble_writer_context.py
@ -9,7 +9,9 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "assemble-context" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 from assemble_writer_context import (  # noqa: E402
    AssemblyError,
--- a/.claude/skills/assemble-context/scripts/test_fine_outline_reader.py
+++ b/.claude/skills/assemble-context/scripts/test_fine_outline_reader.py
@ -5,7 +5,9 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "assemble-context" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 from retrieve_writer_sources import RetrievalError, load_confirmed_fine_outline  # noqa: E402

--- a/.claude/skills/assemble-context/scripts/test_fine_outline_unification.py
+++ b/.claude/skills/assemble-context/scripts/test_fine_outline_unification.py
@ -14,7 +14,9 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "assemble-context" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 from assemble_writer_context import AssemblyError, _outline_contract  # noqa: E402

--- a/.claude/skills/assemble-context/scripts/test_pattern_binding_reader.py
+++ b/.claude/skills/assemble-context/scripts/test_pattern_binding_reader.py
@ -9,7 +9,9 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "assemble-context" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 from retrieve_writer_sources import load_confirmed_pattern_bindings  # noqa: E402

--- a/.claude/skills/assemble-context/scripts/test_retrieve_writer_sources.py
+++ b/.claude/skills/assemble-context/scripts/test_retrieve_writer_sources.py
@ -8,10 +8,18 @@ import pathlib
 import sys
 import unittest

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-sys.path.insert(0, str(SCRIPT_DIR))
-sys.path.insert(0, str(SCRIPT_DIR.parents[1] / "freeze-context" / "scripts"))
-sys.path.insert(0, str(SCRIPT_DIR.parents[1] / "search-knowledge" / "scripts"))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+ASSEMBLE_CONTEXT_SCRIPTS = PROJECT_ROOT / ".claude" / "skills" / "assemble-context" / "scripts"
+FREEZE_CONTEXT_SCRIPTS = PROJECT_ROOT / ".claude" / "skills" / "freeze-context" / "scripts"
+SEARCH_KNOWLEDGE_SCRIPTS = PROJECT_ROOT / ".claude" / "skills" / "search-knowledge" / "scripts"
+EMBED_KNOWLEDGE_SCRIPTS = PROJECT_ROOT / ".claude" / "skills" / "embed-knowledge" / "scripts"
+for script_dir in (
+    ASSEMBLE_CONTEXT_SCRIPTS,
+    FREEZE_CONTEXT_SCRIPTS,
+    SEARCH_KNOWLEDGE_SCRIPTS,
+    EMBED_KNOWLEDGE_SCRIPTS,
+):
+    sys.path.insert(0, str(script_dir))

 from load_reference_work import begin_read_snapshot  # noqa: E402
 from search import search_cards  # noqa: E402
--- a/.claude/skills/assemble-context/scripts/test_style_loader.py
+++ b/.claude/skills/assemble-context/scripts/test_style_loader.py
@ -9,7 +9,9 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "assemble-context" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 from retrieve_writer_sources import derive_style_constraints, load_confirmed_style  # noqa: E402

--- a/.claude/skills/assemble-context/scripts/test_writer_contract.py
+++ b/.claude/skills/assemble-context/scripts/test_writer_contract.py
@ -9,7 +9,9 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "assemble-context" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 from writer_contract import (  # noqa: E402
    ContractError,
--- a/.claude/skills/call-content-model/scripts/test_call_persistence.py
+++ b/.claude/skills/call-content-model/scripts/test_call_persistence.py
@ -9,7 +9,9 @@ import sys
 import types


-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "call-content-model" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 import llm  # noqa: E402


--- a/.claude/skills/call-content-model/scripts/test_quota.py
+++ b/.claude/skills/call-content-model/scripts/test_quota.py
@ -10,7 +10,9 @@ import sys
 import types
 from datetime import datetime, timedelta

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "call-content-model" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 import llm  # noqa: E402

 # 费率缓存预置为兜底表：cost_usd/chat_governed 记账时 get_pricing() 直接命中缓存，绝不触网
--- a/.claude/skills/capture-ai-flavor-cases/scripts/test_capture_cases.py
+++ b/.claude/skills/capture-ai-flavor-cases/scripts/test_capture_cases.py
@ -3,10 +3,16 @@

 from __future__ import annotations

+import sys
 import tempfile
 import unittest
 from pathlib import Path
 from unittest.mock import patch
+
+PROJECT_ROOT = Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "capture-ai-flavor-cases" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
+
 import yaml

 from capture_cases import (
@ -233,20 +239,20 @@ class CaptureCasesTest(unittest.TestCase):
        self.assertEqual("narration", sample["carrier"])

    def test_shipped_fixtures_pass_the_same_validator(self):
-        root = Path(__file__).resolve().parents[1] / "references" / "fixtures"
+        root = SCRIPT_DIR.parent / "references" / "fixtures"
        for name in ("backfill-hash-only.yaml", "canonical-samples.yaml"):
            data = yaml.safe_load((root / name).read_text(encoding="utf-8"))
            for card in data["cards"]:
                validate_card(card)

    def test_shipped_rule_seed_is_candidate_only(self):
-        root = Path(__file__).resolve().parents[1] / "references" / "fixtures"
+        root = SCRIPT_DIR.parent / "references" / "fixtures"
        data = yaml.safe_load((root / "rule-candidates.yaml").read_text(encoding="utf-8"))
        self.assertTrue(data["rules"])
        self.assertTrue(all(rule["status"] == "candidate" for rule in data["rules"]))

    def test_shipped_revalidation_report_is_structurally_usable(self):
-        root = Path(__file__).resolve().parents[1] / "references" / "fixtures"
+        root = SCRIPT_DIR.parent / "references" / "fixtures"
        data = yaml.safe_load((root / "revalidation-2026-08-14.json").read_text(encoding="utf-8"))
        self.assertEqual("ai-flavor-revalidation-v1", data["schema_version"])
        self.assertTrue(data["usable"])
--- a/.claude/skills/check-content-consistency/scripts/test_build_semantic_input.py
+++ b/.claude/skills/check-content-consistency/scripts/test_build_semantic_input.py
@ -4,7 +4,7 @@
 验证：WriterContext + 候选能被投影成通过闭集校验的 semantic-detector-input-v3；
 sourceRef 多余字段被清洗；身份字段严格绑定；哈希自洽。

-跑法：.venv/bin/python .claude/skills/check-content-consistency/scripts/test_build_semantic_input.py
+跑法：.venv/bin/python tests/skills/check-content-consistency/test_build_semantic_input.py
 """

 from __future__ import annotations
@ -14,12 +14,13 @@ import pathlib
 import sys
 import unittest

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
-for path in (SCRIPT_DIR,
-             SKILLS_DIR / "check-content-consistency" / "scripts",
-             SKILLS_DIR / "write-next-chapter" / "scripts",
-             SKILLS_DIR / "assemble-context" / "scripts"):
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+TEST_DIR = PROJECT_ROOT / "tests" / "skills" / "check-content-consistency"
+SCRIPT_DIR = SKILLS_DIR / "check-content-consistency" / "scripts"
+CONTINUATION_DIR = SKILLS_DIR / "write-next-chapter" / "scripts"
+READ_CONTEXT_DIR = SKILLS_DIR / "assemble-context" / "scripts"
+for path in (TEST_DIR, SCRIPT_DIR, CONTINUATION_DIR, READ_CONTEXT_DIR):
    if str(path) not in sys.path:
        sys.path.insert(0, str(path))

--- a/.claude/skills/check-content-consistency/scripts/test_check_writer_candidate.py
+++ b/.claude/skills/check-content-consistency/scripts/test_check_writer_candidate.py
@ -9,12 +9,15 @@ import pathlib
 import sys
 import unittest

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "check-content-consistency" / "scripts"
 CONTINUATION_DIR = SKILLS_DIR / "write-next-chapter" / "scripts"
 READ_CONTEXT_DIR = SKILLS_DIR / "assemble-context" / "scripts"
-for path in (SCRIPT_DIR, CONTINUATION_DIR, READ_CONTEXT_DIR):
-    sys.path.insert(0, str(path))
+WRITER_TEST_DIR = PROJECT_ROOT / "tests" / "skills" / "write-next-chapter"
+for path in (SCRIPT_DIR, CONTINUATION_DIR, READ_CONTEXT_DIR, WRITER_TEST_DIR):
+    if str(path) not in sys.path:
+        sys.path.insert(0, str(path))

 from test_run_writer import _bound_context  # noqa: E402
 from check_writer_candidate import check_writer_candidate  # noqa: E402
--- a/.claude/skills/check-content-consistency/scripts/test_run_writer_semantic_detector.py
+++ b/.claude/skills/check-content-consistency/scripts/test_run_writer_semantic_detector.py
@ -10,7 +10,10 @@ import sys
 import unittest
 from typing import Any, Mapping, Sequence

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "check-content-consistency" / "scripts"
+if str(SCRIPT_DIR) not in sys.path:
+    sys.path.insert(0, str(SCRIPT_DIR))

 from run_writer_semantic_detector import (  # noqa: E402
    SEMANTIC_DETECTOR_REPORT_JSON_SCHEMA,
--- a/.claude/skills/clean-book-text/scripts/test_clean_detect_offline.py
+++ b/.claude/skills/clean-book-text/scripts/test_clean_detect_offline.py
@ -16,7 +16,9 @@ from unittest.mock import patch

 from click.testing import CliRunner

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "clean-book-text" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 import clean_detect  # noqa: E402


--- a/.claude/skills/decide-candidate/scripts/test_confirm_knowledge_offline.py
+++ b/.claude/skills/decide-candidate/scripts/test_confirm_knowledge_offline.py
@ -5,7 +5,8 @@ import sys
 import unittest


-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "decide-candidate" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 import confirm_knowledge as confirm  # noqa: E402
--- a/.claude/skills/decide-candidate/scripts/test_fact_delta.py
+++ b/.claude/skills/decide-candidate/scripts/test_fact_delta.py
@ -4,7 +4,7 @@
 模型只能提类型化增量，证据引文必须真实出现在候选正文中（不得编造证据）；
 字段闭集、类型闭集、payload 合同、重复 ID 一律失败关闭。

-跑法：.venv/bin/python .claude/skills/decide-candidate/scripts/test_fact_delta.py
+跑法：.venv/bin/python tests/skills/decide-candidate/test_fact_delta.py
 """

 from __future__ import annotations
@ -14,8 +14,9 @@ import pathlib
 import sys
 import unittest

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "decide-candidate" / "scripts"
 for path in (SCRIPT_DIR, SKILLS_DIR / "assemble-context" / "scripts"):
    if str(path) not in sys.path:
        sys.path.insert(0, str(path))
--- a/.claude/skills/decide-candidate/scripts/test_fact_delta_db.py
+++ b/.claude/skills/decide-candidate/scripts/test_fact_delta_db.py
@ -11,7 +11,7 @@
 测试数据 unittest-delta- 前缀隔离；清理时短暂禁用账本防删触发器（try/finally 恢复）。

 跑法（需 Tailscale 内网可达 muse-example）：
-.venv/bin/python .claude/skills/decide-candidate/scripts/test_fact_delta_db.py
+.venv/bin/python tests/skills/decide-candidate/test_fact_delta_db.py
 """

 from __future__ import annotations
@ -21,8 +21,11 @@ import pathlib
 import sys
 import uuid

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
+from psycopg.errors import RaiseException
+
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "decide-candidate" / "scripts"
 for path in (SCRIPT_DIR, SKILLS_DIR / "access-database" / "scripts"):
    if str(path) not in sys.path:
        sys.path.insert(0, str(path))
@ -266,7 +269,7 @@ def test_ledger_append_only(work_id: int) -> None:
        raise AssertionError("账本必须 append-only")
    except AssertionError:
        raise
-    except Exception:
+    except RaiseException:
        pass
    try:
        with connect() as conn:
@ -275,7 +278,7 @@ def test_ledger_append_only(work_id: int) -> None:
        raise AssertionError("账本必须 append-only")
    except AssertionError:
        raise
-    except Exception:
+    except RaiseException:
        pass


--- a/.claude/skills/decide-candidate/scripts/test_projection_db.py
+++ b/.claude/skills/decide-candidate/scripts/test_projection_db.py
@ -10,7 +10,7 @@
 测试数据 unittest-proj- 前缀隔离，结束物理清理（本表可变，直接 DELETE）。

 跑法（需 Tailscale 内网可达 muse-example）：
-.venv/bin/python .claude/skills/decide-candidate/scripts/test_projection_db.py
+.venv/bin/python tests/skills/decide-candidate/test_projection_db.py
 """

 from __future__ import annotations
@ -20,8 +20,11 @@ import pathlib
 import sys
 import uuid

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
+from psycopg.errors import RaiseException
+
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "decide-candidate" / "scripts"
 for path in (SCRIPT_DIR, SKILLS_DIR / "access-database" / "scripts"):
    if str(path) not in sys.path:
        sys.path.insert(0, str(path))
@ -175,7 +178,7 @@ def test_failed_never_masquerades_completed(work_id: int) -> None:
        raise AssertionError("触发器必须拒绝 failed→completed")
    except AssertionError:
        raise
-    except Exception:
+    except RaiseException:
        pass
    # 显式 retry 才能回 pending，且 attempt+1
    retried = retry_projection(pending[0])
@ -202,7 +205,7 @@ def test_stale_projections_cannot_report_outcomes(work_id: int) -> None:
        raise AssertionError("stale→completed 必须被拒绝")
    except AssertionError:
        raise
-    except Exception:
+    except RaiseException:
        pass
    try:
        with connect() as conn:
@ -212,7 +215,7 @@ def test_stale_projections_cannot_report_outcomes(work_id: int) -> None:
        raise AssertionError("stale→failed 必须被拒绝")
    except AssertionError:
        raise
-    except Exception:
+    except RaiseException:
        pass
    # stale 走 retry 恢复：attempt+1 回 pending，随后可以正常完成
    retried = retry_projection(stale_row[0])
--- a/.claude/skills/decide-candidate/scripts/test_write_canonical_db.py
+++ b/.claude/skills/decide-candidate/scripts/test_write_canonical_db.py
@ -12,7 +12,7 @@
 清理时短暂禁用其防删触发器（try/finally 保证恢复）。

 跑法（需 Tailscale 内网可达 muse-example）：
-.venv/bin/python .claude/skills/decide-candidate/scripts/test_write_canonical_db.py
+.venv/bin/python tests/skills/decide-candidate/test_write_canonical_db.py
 """

 from __future__ import annotations
@ -24,8 +24,9 @@ import uuid
 from typing import Any
 from unittest import mock

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "decide-candidate" / "scripts"
 for path in (SCRIPT_DIR, SKILLS_DIR / "access-database" / "scripts"):
    if str(path) not in sys.path:
        sys.path.insert(0, str(path))
--- a/.claude/skills/decide-candidate/scripts/test_writer_acceptance.py
+++ b/.claude/skills/decide-candidate/scripts/test_writer_acceptance.py
@ -14,8 +14,9 @@ import pathlib
 import sys
 import unittest

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "decide-candidate" / "scripts"
 CONTINUATION_DIR = SKILLS_DIR / "write-next-chapter" / "scripts"
 READ_CONTEXT_DIR = SKILLS_DIR / "assemble-context" / "scripts"
 for path in (SCRIPT_DIR, CONTINUATION_DIR, READ_CONTEXT_DIR):
--- a/.claude/skills/deconstruct-book/scripts/test_parse_llm_offline.py
+++ b/.claude/skills/deconstruct-book/scripts/test_parse_llm_offline.py
@ -11,7 +11,9 @@ import unittest
 from unittest.mock import Mock, patch


-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "deconstruct-book" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 import parse_llm as pll  # noqa: E402


--- a/.claude/skills/deconstruct-book/scripts/test_parse_outline_offline.py
+++ b/.claude/skills/deconstruct-book/scripts/test_parse_outline_offline.py
@ -6,8 +6,9 @@ import sys

 import click

-HERE = pathlib.Path(__file__).resolve().parent
-sys.path.insert(0, str(HERE))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "deconstruct-book" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 import parse_outline as po  # noqa: E402

--- a/.claude/skills/diagnose-ai-flavor/scripts/test_diagnose_ai_flavor.py
+++ b/.claude/skills/diagnose-ai-flavor/scripts/test_diagnose_ai_flavor.py
@ -7,7 +7,9 @@ import tempfile
 import unittest
 from unittest.mock import patch

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "diagnose-ai-flavor" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 import diagnose_ai_flavor as diag  # noqa: E402

--- a/.claude/skills/embed-knowledge/scripts/test_embed_drafts_offline.py
+++ b/.claude/skills/embed-knowledge/scripts/test_embed_drafts_offline.py
@ -11,7 +11,8 @@ from unittest.mock import MagicMock, Mock, patch
 from click.testing import CliRunner


-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "embed-knowledge" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 import embed_drafts as embed  # noqa: E402
--- a/.claude/skills/establish-voice-baseline/scripts/test_establish_voice_baseline.py
+++ b/.claude/skills/establish-voice-baseline/scripts/test_establish_voice_baseline.py
@ -6,7 +6,9 @@ import tempfile
 import unittest
 from unittest.mock import patch

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "establish-voice-baseline" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 import establish_voice_baseline as base  # noqa: E402

--- a/.claude/skills/evaluate-frozen-replay/scripts/test_fine_outline_detector.py
+++ b/.claude/skills/evaluate-frozen-replay/scripts/test_fine_outline_detector.py
@ -7,7 +7,11 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "evaluate-frozen-replay" / "scripts"
+if str(SCRIPT_DIR) not in sys.path:
+    sys.path.insert(0, str(SCRIPT_DIR))
+
 from fine_outline_detector import validate_detector_report  # noqa: E402


@ -76,7 +80,9 @@ class FineOutlineDetectorTest(unittest.TestCase):
        self.assertTrue(validate_detector_report(report, "blind-1")["ok"])

        skill_path = (
-            pathlib.Path(__file__).resolve().parents[2]
+            PROJECT_ROOT
+            / ".claude"
+            / "skills"
            / "check-content-consistency"
            / "SKILL.md"
        )
--- a/.claude/skills/evaluate-frozen-replay/scripts/test_gate_input_builder.py
+++ b/.claude/skills/evaluate-frozen-replay/scripts/test_gate_input_builder.py
@ -8,9 +8,12 @@ import pathlib
 import sys
 import unittest

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-QUALITY_GATE_DIR = SCRIPT_DIR.parents[1] / "score-content-quality" / "scripts"
-for _import_dir in (SCRIPT_DIR, QUALITY_GATE_DIR):
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "evaluate-frozen-replay" / "scripts"
+TEST_DIR = PROJECT_ROOT / "tests" / "skills" / "evaluate-frozen-replay"
+QUALITY_GATE_DIR = SKILLS_DIR / "score-content-quality" / "scripts"
+for _import_dir in (SCRIPT_DIR, TEST_DIR, QUALITY_GATE_DIR):
    if str(_import_dir) not in sys.path:
        sys.path.insert(0, str(_import_dir))

--- a/.claude/skills/evaluate-frozen-replay/scripts/test_load_writer_reference_work.py
+++ b/.claude/skills/evaluate-frozen-replay/scripts/test_load_writer_reference_work.py
@ -14,12 +14,24 @@ import unittest
 from datetime import datetime, timedelta, timezone
 from unittest.mock import patch

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-READ_CONTEXT_SCRIPTS = SCRIPT_DIR.parents[1] / "assemble-context" / "scripts"
-DETECT_SCRIPTS = SCRIPT_DIR.parents[1] / "check-content-consistency" / "scripts"
-CONTINUATION_SCRIPTS = SCRIPT_DIR.parents[1] / "write-next-chapter" / "scripts"
-for _path in (SCRIPT_DIR, READ_CONTEXT_SCRIPTS, DETECT_SCRIPTS, CONTINUATION_SCRIPTS):
-    sys.path.insert(0, str(_path))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "evaluate-frozen-replay" / "scripts"
+TEST_DIR = PROJECT_ROOT / "tests" / "skills" / "evaluate-frozen-replay"
+READ_CONTEXT_SCRIPTS = SKILLS_DIR / "assemble-context" / "scripts"
+DETECT_SCRIPTS = SKILLS_DIR / "check-content-consistency" / "scripts"
+CONTINUATION_SCRIPTS = SKILLS_DIR / "write-next-chapter" / "scripts"
+CONTINUATION_TESTS = PROJECT_ROOT / "tests" / "skills" / "write-next-chapter"
+for _path in (
+    SCRIPT_DIR,
+    TEST_DIR,
+    READ_CONTEXT_SCRIPTS,
+    DETECT_SCRIPTS,
+    CONTINUATION_SCRIPTS,
+    CONTINUATION_TESTS,
+):
+    if str(_path) not in sys.path:
+        sys.path.insert(0, str(_path))

 import load_writer_reference_work as loader  # noqa: E402
 import run_writer_replay as replay_module  # noqa: E402
--- a/.claude/skills/evaluate-frozen-replay/scripts/test_pattern_reference_injection.py
+++ b/.claude/skills/evaluate-frozen-replay/scripts/test_pattern_reference_injection.py
@ -13,8 +13,11 @@ import pathlib
 import sys
 import unittest

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "evaluate-frozen-replay" / "scripts"
+TEST_DIR = PROJECT_ROOT / "tests" / "skills" / "evaluate-frozen-replay"
+ASSEMBLE_CONTEXT_TESTS = PROJECT_ROOT / "tests" / "skills" / "assemble-context"
 READ_CONTEXT_SCRIPTS = SKILLS_DIR / "assemble-context" / "scripts"
 # 端到端链路测试要导入回放包（run_writer_replay.sample），其依赖执行、证据与
 # 评分三个 Skill 的 scripts 目录，路径口径与 test_run_writer_replay 保持一致。
@ -23,12 +26,15 @@ EVIDENCE_SCRIPTS = SKILLS_DIR / "record-run-evidence" / "scripts"
 QUALITY_GATE_SCRIPTS = SKILLS_DIR / "score-content-quality" / "scripts"
 for _path in (
    SCRIPT_DIR,
+    TEST_DIR,
+    ASSEMBLE_CONTEXT_TESTS,
    READ_CONTEXT_SCRIPTS,
    EXECUTION_SCRIPTS,
    EVIDENCE_SCRIPTS,
    QUALITY_GATE_SCRIPTS,
 ):
-    sys.path.insert(0, str(_path))
+    if str(_path) not in sys.path:
+        sys.path.insert(0, str(_path))

 import load_writer_reference_work as loader  # noqa: E402
 from load_writer_reference_work import (  # noqa: E402
--- a/.claude/skills/evaluate-frozen-replay/scripts/test_refresh_runtime_probe.py
+++ b/.claude/skills/evaluate-frozen-replay/scripts/test_refresh_runtime_probe.py
@ -21,15 +21,14 @@ import unittest
 from typing import Any, Mapping
 from unittest import mock

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-EXECUTION_DIR = SCRIPT_DIR.parents[1] / "execute-claude-task" / "scripts"
-QUALITY_GATE_DIR = SCRIPT_DIR.parents[1] / "score-content-quality" / "scripts"
-if str(SCRIPT_DIR) not in sys.path:
-    sys.path.insert(0, str(SCRIPT_DIR))
-if str(EXECUTION_DIR) not in sys.path:
-    sys.path.insert(0, str(EXECUTION_DIR))
-if str(QUALITY_GATE_DIR) not in sys.path:
-    sys.path.insert(0, str(QUALITY_GATE_DIR))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "evaluate-frozen-replay" / "scripts"
+EXECUTION_DIR = SKILLS_DIR / "execute-claude-task" / "scripts"
+QUALITY_GATE_DIR = SKILLS_DIR / "score-content-quality" / "scripts"
+for _import_dir in (SCRIPT_DIR, EXECUTION_DIR, QUALITY_GATE_DIR):
+    if str(_import_dir) not in sys.path:
+        sys.path.insert(0, str(_import_dir))

 import refresh_runtime_probe as refresh_module  # noqa: E402
 from refresh_runtime_probe import (  # noqa: E402
--- a/.claude/skills/evaluate-frozen-replay/scripts/test_run_replay.py
+++ b/.claude/skills/evaluate-frozen-replay/scripts/test_run_replay.py
@ -11,10 +11,14 @@ import tempfile
 import unittest
 from unittest.mock import patch

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-QUALITY_GATE_DIR = SCRIPT_DIR.parents[1] / "score-content-quality" / "scripts"
-sys.path.insert(0, str(SCRIPT_DIR))
-sys.path.insert(0, str(QUALITY_GATE_DIR))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "evaluate-frozen-replay" / "scripts"
+TEST_DIR = PROJECT_ROOT / "tests" / "skills" / "evaluate-frozen-replay"
+QUALITY_GATE_DIR = SKILLS_DIR / "score-content-quality" / "scripts"
+for _import_dir in (SCRIPT_DIR, TEST_DIR, QUALITY_GATE_DIR):
+    if str(_import_dir) not in sys.path:
+        sys.path.insert(0, str(_import_dir))
 from run_replay import _parse_args, _planner_prompt, run_replay as _run_replay  # noqa: E402
 from test_writer_gate import build_input, source_bundle  # noqa: E402
 from writer_gate import issue_gate_report_and_receipt  # noqa: E402
--- a/.claude/skills/evaluate-frozen-replay/scripts/test_run_writer_replay.py
+++ b/.claude/skills/evaluate-frozen-replay/scripts/test_run_writer_replay.py
@ -19,8 +19,10 @@ from dataclasses import replace
 from decimal import Decimal
 from datetime import datetime, timedelta, timezone

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
+TEST_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "evaluate-frozen-replay" / "scripts"
 READ_CONTEXT_DIR = SKILLS_DIR / "assemble-context" / "scripts"
 EXECUTION_DIR = SKILLS_DIR / "execute-claude-task" / "scripts"
 EVIDENCE_DIR = SKILLS_DIR / "record-run-evidence" / "scripts"
@ -32,7 +34,8 @@ for import_path in (
    EVIDENCE_DIR,
    QUALITY_GATE_DIR,
 ):
-    sys.path.insert(0, str(import_path))
+    if str(import_path) not in sys.path:
+        sys.path.insert(0, str(import_path))

 import run_writer_replay as replay_module  # noqa: E402
 import run_writer_replay.execute as execute_module  # noqa: E402
--- a/.claude/skills/evaluate-frozen-replay/scripts/test_writer_eval_preregister.py
+++ b/.claude/skills/evaluate-frozen-replay/scripts/test_writer_eval_preregister.py
@ -8,7 +8,10 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "evaluate-frozen-replay" / "scripts"
+if str(SCRIPT_DIR) not in sys.path:
+    sys.path.insert(0, str(SCRIPT_DIR))

 from writer_eval_preregister import (  # noqa: E402
    PreregistrationError,
--- a/.claude/skills/evaluate-frozen-replay/scripts/test_writer_gate.py
+++ b/.claude/skills/evaluate-frozen-replay/scripts/test_writer_gate.py
@ -10,10 +10,13 @@ import sys
 import tempfile
 import unittest

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-QUALITY_GATE_DIR = SCRIPT_DIR.parents[1] / "score-content-quality" / "scripts"
-EVIDENCE_DIR = SCRIPT_DIR.parents[1] / "record-run-evidence" / "scripts"
-for _import_dir in (SCRIPT_DIR, QUALITY_GATE_DIR, EVIDENCE_DIR):
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "evaluate-frozen-replay" / "scripts"
+TEST_DIR = PROJECT_ROOT / "tests" / "skills" / "evaluate-frozen-replay"
+QUALITY_GATE_DIR = SKILLS_DIR / "score-content-quality" / "scripts"
+EVIDENCE_DIR = SKILLS_DIR / "record-run-evidence" / "scripts"
+for _import_dir in (SCRIPT_DIR, TEST_DIR, QUALITY_GATE_DIR, EVIDENCE_DIR):
    if str(_import_dir) not in sys.path:
        sys.path.insert(0, str(_import_dir))

--- a/.claude/skills/execute-claude-task/scripts/test_claude_runtime.py
+++ b/.claude/skills/execute-claude-task/scripts/test_claude_runtime.py
@ -17,7 +17,8 @@ import unittest
 from unittest import mock
 from decimal import Decimal

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "execute-claude-task" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 from claude_runtime import (  # noqa: E402
--- a/.claude/skills/extract-chapter-knowledge/scripts/test_extract_knowledge_offline.py
+++ b/.claude/skills/extract-chapter-knowledge/scripts/test_extract_knowledge_offline.py
@ -3,7 +3,9 @@
 import pathlib
 import sys

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "extract-chapter-knowledge" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 from extract_knowledge import ExtractionContractError, normalize_extraction, salvage_extraction  # noqa: E402


--- a/.claude/skills/extract-work-knowledge/scripts/test_parse_upgrade_offline.py
+++ b/.claude/skills/extract-work-knowledge/scripts/test_parse_upgrade_offline.py
@ -13,13 +13,12 @@
 本测试**不触碰真实服务**——质量修复命令只通过 fake DB 验证 preview、CAS 和原子写入，
 不在离线自测中执行真实 work/window。

-跑法：仓根 `.venv/bin/python .claude/skills/extract-work-knowledge/scripts/test_parse_upgrade_offline.py`
+跑法：仓库根目录 `.venv/bin/python tests/skills/extract-work-knowledge/test_parse_upgrade_offline.py`
 """
 import json
 import hashlib
 import inspect
 from io import StringIO
-import os
 import pathlib
 import sys
 from contextlib import nullcontext
@ -29,8 +28,10 @@ from unittest.mock import patch

 from click.testing import CliRunner

-# 与 upgrade 同目录：直接 import 触发其 sys.path 装配（含 parse-book/embed/llm scripts），随后可导 embed_drafts
-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+# 从生产 scripts 导入 upgrade；upgrade 自己装配 parse-book/embed/llm 的生产模块路径。
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "extract-work-knowledge" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 import upgrade as pu  # noqa: E402
 import embed_drafts  # noqa: E402  （型修正验证在 embed skill 本体）

@ -4724,17 +4725,6 @@ def test_recovery_and_redo_rejection_contracts():
        check("redo-run首行机械拒绝", False, detail="_run 接受了 redo")


-def test_real_pg_rollback_smoke_entry():
-    """真实 PG 入口默认不运行；显式开启后验证 public 窗状态写入可回滚。"""
-
-    if os.getenv("MUSE_REAL_PG_ROLLBACK_SMOKE") != "1":
-        check("real-pg-rollback-smoke默认跳过", True)
-        return
-    work_id = int(os.getenv("MUSE_REAL_PG_ROLLBACK_SMOKE_WORK_ID", "8"))
-    pu.real_pg_rollback_smoke(work_id)
-    check("real-pg-rollback-smoke真实public窗回滚", True)
-
-
 class _LegacyRecoveryConn:
    """恢复命令离线夹具：用安全墓碑及异常软删验证 SQL 过滤。"""

@ -5847,538 +5837,7 @@ def test_quality_repair_rejects_alias_and_presence_duplicates():
        )


-# ── presence 冗余收口（work8 十组双行：保留 MIN(id) 软删 MAX(id)）──

-# 离线 fixture 用字符串时间戳代替数据库驱动的 datetime：窄合同只比较同组两行是否相等，
-# _sha256_json 计算摘要时统一 default=str 转写，二者行为一致。
-_PRESENCE_DEDUPE_BASE_TIME = "2026-07-01 12:00:00"
-
-
-def _presence_dedupe_rows():
-    """构造满足窄合同的 work8 presence 行 fixture。
-
-    十组双行：每组 observation 不同、create_time 相同、恰好一个待删 ID 命中 DELETE_IDS、
-    保留 ID 是组内较小者；另有四个单行，证明快照会跳过不构成冗余的 key。
-    """
-
-    rows = []
-    for index, delete_id in enumerate(sorted(pu.PRESENCE_DEDUPE_DELETE_IDS)):
-        keep_id = 11501 + index
-        entity_type = "location" if index % 2 == 0 else "item"
-        name = f"重复实体{index}"
-        for row_id, observation in ((keep_id, f"观察A{index}"),
-                                    (delete_id, f"观察B{index}")):
-            rows.append((row_id, 8, 57 + index, 403 + index, entity_type, name,
-                         observation, "upgrade", _PRESENCE_DEDUPE_BASE_TIME,
-                         False, pu.TENANT))
-    for index in range(4):
-        rows.append((11400 + index, 8, 10 + index, 20 + index, "character",
-                     f"单体实体{index}", f"单体观察{index}",
-                     "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT))
-    return sorted(rows, key=lambda row: row[0])
-
-
-def _presence_dedupe_state():
-    return {
-        "rows": _presence_dedupe_rows(),
-        "writes": [],
-        "queries": [],
-        "commits": 0,
-    }
-
-
-def _mutate_presence_row(state, row_id, column, value):
-    """替换指定行的单列；元组不可变，整体重建后写回。"""
-
-    state["rows"] = [
-        row[:column] + (value,) + row[column + 1:] if row[0] == row_id else row
-        for row in state["rows"]
-    ]
-
-
-class _PresenceDedupeConn:
-    """presence 冗余收口 fake DB：只实现本命令的读快照与软删 SQL。
-
-    其他任何域（draft/window/alias/card_state/audit/embedding）的 SQL 会落到末尾
-    AssertionError，因此「零副作用」无需逐条枚举禁写语句即可离线断言。fake 只做
-    「未软删」粗过滤；列数/work/tenant/deleted 类型等窄合同行像校验正是离线断言对象，
-    fake 不能代为过滤。
-    """
-
-    def __init__(self, state):
-        self.state = state
-
-    def __enter__(self):
-        return self
-
-    def __exit__(self, exc_type, exc, tb):
-        return False
-
-    def execute(self, query, params=()):
-        normalized = " ".join(query.split())
-        self.state.setdefault("queries", []).append(normalized)
-        if normalized.startswith("SET TRANSACTION") or normalized.startswith("LOCK TABLE"):
-            return _CardResult()
-        if normalized.startswith(
-                "SELECT id, work_id, window_no, chapter_no, entity_type, name, observation,"):
-            rows = [
-                row for row in sorted(self.state["rows"], key=lambda item: item[0])
-                if len(row) > 9 and not row[9]
-            ]
-            return _CardResult(rows=rows)
-        if normalized.startswith("UPDATE example_upgrade_presence SET deleted=TRUE"):
-            tenant_id, work_id, delete_ids = params
-            assert tenant_id == pu.TENANT
-            delete_set = set(delete_ids)
-            hit = []
-            new_rows = []
-            for row in self.state["rows"]:
-                if row[1] == work_id and row[0] in delete_set and not row[9]:
-                    row = row[:9] + (True,) + row[10:]
-                    hit.append((row[0],))
-                new_rows.append(row)
-            self.state["rows"] = new_rows
-            self.state["writes"].append("UPDATE example_upgrade_presence")
-            return _CardResult(rows=hit)
-        raise AssertionError(f"presence 冗余收口 fake DB 未覆盖 SQL：{normalized}")
-
-    def commit(self):
-        self.state["commits"] += 1
-
-
-def _presence_dedupe_cli(state, args):
-    """离线执行 repair-presence-duplicates：注入 fake DB 与空放同书锁。"""
-
-    conn = _PresenceDedupeConn(state)
-    with patch.object(pu.psycopg, "connect", return_value=conn), \
-            patch.object(pu, "upgrade_work_lock", return_value=nullcontext()):
-        result = CliRunner().invoke(pu.maintenance_cli, ["repair-presence-duplicates"] + args)
-    return result, conn
-
-
-def _presence_dedupe_preview_sha(state):
-    """离线取 preview 的 confirmation_sha，作为 execute 的合法输入。"""
-
-    result, _ = _presence_dedupe_cli(state, ["--work-id", "8", "--preview"])
-    assert result.exit_code == 0, result.output
-    return json.loads(result.output)["confirmation_sha"]
-
-
-def _presence_snapshot_raises(label, state, work_id=pu.PRESENCE_DEDUPE_WORK_ID):
-    """断言快照以 CompensationFenceConflict 拒绝，且没有发出任何写入。"""
-
-    conn = _PresenceDedupeConn(state)
-    try:
-        pu._presence_dedupe_capture_snapshot(conn, work_id)
-    except pu.CompensationFenceConflict as exc:
-        check(f"presence-dedupe-{label}失败关闭", True, detail=str(exc))
-    else:
-        check(f"presence-dedupe-{label}失败关闭", False,
-              detail="未抛 CompensationFenceConflict")
-    check(
-        f"presence-dedupe-{label}零副作用",
-        state["writes"] == [] and state["commits"] == 0,
-    )
-
-
-def test_presence_dedupe_actions_planner_guard():
-    """planner 级收口守卫：同观察幂等留首条、不同观察失败关闭、非 presence 原样透传。"""
-
-    presence_a = ("presence", None, 403, "location", "重复星体", "观察一")
-    presence_a_dup = ("presence", None, 403, "location", "重复星体", "观察一")
-    presence_b = ("presence", None, 404, "item", "镜面护盾", "观察二")
-    alias_action = ("alias", 7, "规范名", "别名", "ai")
-    new_action = ("new", -1, {"名称": "新实体"}, set(), [])
-    actions = [presence_a, alias_action, presence_a_dup, new_action, presence_b, ()]
-    original = list(actions)
-    deduped = pu._dedupe_presence_actions(actions)
-    check(
-        "presence-dedupe-planner同观察去重保留首条",
-        deduped == [presence_a, alias_action, new_action, presence_b, ()]
-        and deduped[0] is presence_a,
-        detail=str(deduped),
-    )
-    check("presence-dedupe-重复应用幂等",
-          pu._dedupe_presence_actions(deduped) == deduped)
-    check("presence-dedupe-不修改输入列表", actions == original)
-
-    conflict = [presence_a,
-                ("presence", None, 403, "location", "重复星体", "另一个观察")]
-    try:
-        pu._dedupe_presence_actions(conflict)
-    except RuntimeError as exc:
-        check("presence-dedupe-不同观察失败关闭",
-              "observation 冲突" in str(exc), detail=str(exc))
-    else:
-        check("presence-dedupe-不同观察失败关闭", False, detail="未抛 RuntimeError")
-
-    for label, bad, keyword in (
-        ("结构缺一元", ("presence", None, 403, "location", "重复星体"),
-         "结构非法"),
-        ("结构多一元",
-         ("presence", None, 403, "location", "重复星体", "观察一", "extra"),
-         "结构非法"),
-        ("observation非字符串", ("presence", None, 403, "location", "重复星体", None),
-         "必须是字符串"),
-    ):
-        try:
-            pu._dedupe_presence_actions([bad])
-        except RuntimeError as exc:
-            check(f"presence-dedupe-{label}抛错", keyword in str(exc), detail=str(exc))
-        else:
-            check(f"presence-dedupe-{label}抛错", False, detail="未抛 RuntimeError")
-
-
-def test_presence_dedupe_snapshot_success():
-    """快照成功路径：恰好 10 组/20 行，待删集合精确等于 DELETE_IDS，保留为组内较小者。"""
-
-    state = _presence_dedupe_state()
-    snapshot = pu._presence_dedupe_capture_snapshot(
-        _PresenceDedupeConn(state), pu.PRESENCE_DEDUPE_WORK_ID,
-    )
-    check(
-        "presence-dedupe-快照组数行数精确",
-        len(snapshot["groups"]) == pu.PRESENCE_DEDUPE_GROUP_COUNT
-        and len(snapshot["rows"]) == pu.PRESENCE_DEDUPE_ROW_COUNT
-        and snapshot["contract"] == pu.PRESENCE_DEDUPE_CONTRACT
-        and snapshot["work_id"] == pu.PRESENCE_DEDUPE_WORK_ID,
-    )
-    check(
-        "presence-dedupe-待删集合精确等于DELETE_IDS",
-        snapshot["delete_ids"] == sorted(pu.PRESENCE_DEDUPE_DELETE_IDS),
-        detail=str(snapshot["delete_ids"]),
-    )
-    check(
-        "presence-dedupe-保留为组内较小ID",
-        snapshot["keep_ids"] == list(range(11501, 11511))
-        and all(group["keep_id"] < group["delete_id"] for group in snapshot["groups"])
-        and all(len(group["rows"]) == 2 for group in snapshot["groups"]),
-        detail=str(snapshot["keep_ids"]),
-    )
-    check(
-        "presence-dedupe-行按id排序且排除单体行",
-        [row[0] for row in snapshot["rows"]]
-        == sorted(row[0] for row in snapshot["rows"])
-        and not any(row[0] in range(11400, 11404) for row in snapshot["rows"]),
-    )
-
-
-def test_presence_dedupe_snapshot_fail_closed():
-    """各类窄合同漂移都必须在写入前以 CompensationFenceConflict 逐条拒绝。"""
-
-    # 额外冗余组：组数 11 ≠ 10（新增组复用合同内待删 ID，确保失败点落在组数校验）。
-    state = _presence_dedupe_state()
-    state["rows"].extend((
-        (11581, 8, 90, 500, "event", "额外重复组", "观察C",
-         "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT),
-        (11582, 8, 90, 500, "event", "额外重复组", "观察D",
-         "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT),
-    ))
-    _presence_snapshot_raises("额外冗余组", state)
-
-    # 三行组不是「恰好两行」。
-    state = _presence_dedupe_state()
-    state["rows"].append(
-        (11601, 8, 57, 403, "location", "重复实体0", "观察E",
-         "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT)
-    )
-    _presence_snapshot_raises("三行冗余组", state)
-
-    # 缺行：一组只剩单行被跳过 → 组数 9 ≠ 10。
-    state = _presence_dedupe_state()
-    state["rows"] = [row for row in state["rows"]
-                     if row[0] != sorted(pu.PRESENCE_DEDUPE_DELETE_IDS)[0]]
-    _presence_snapshot_raises("缺行", state)
-
-    # 同组两行 create_time 漂移。
-    state = _presence_dedupe_state()
-    _mutate_presence_row(state, 11501, 8, "2026-07-02 09:30:00")
-    _presence_snapshot_raises("create_time漂移", state)
-
-    # 待删 ID 命中 0：组内两行都不在 DELETE_IDS。
-    state = _presence_dedupe_state()
-    _mutate_presence_row(state, 11501, 0, 12001)
-    _mutate_presence_row(state, 11582, 0, 12002)
-    _presence_snapshot_raises("待删ID无命中", state)
-
-    # 待删 ID 命中 2：组内两行都在 DELETE_IDS。
-    state = _presence_dedupe_state()
-    _mutate_presence_row(state, 11501, 0, 11693)
-    _presence_snapshot_raises("待删ID双命中", state)
-
-    # 保留不是组内较小者：待删 ID 反而是组内较小者。
-    state = _presence_dedupe_state()
-    _mutate_presence_row(state, 11501, 0, 11900)
-    _presence_snapshot_raises("保留非组内较小", state)
-
-    # 同组两行 observation 相同（byte-exact 重复违反窄合同）。
-    state = _presence_dedupe_state()
-    _mutate_presence_row(state, 11582, 6, "观察A0")
-    _presence_snapshot_raises("observation相同", state)
-
-    # work_id ≠ 8 拒绝。
-    _presence_snapshot_raises("work不符", _presence_dedupe_state(), work_id=9)
-
-    # 窄合同行像漂移：列数不足。
-    state = _presence_dedupe_state()
-    state["rows"] = [row[:10] if row[0] == 11501 else row for row in state["rows"]]
-    _presence_snapshot_raises("列数不符", state)
-
-    # 窄合同行像漂移：work 列与查询目标不符。
-    state = _presence_dedupe_state()
-    _mutate_presence_row(state, 11501, 1, 9)
-    _presence_snapshot_raises("work列不符", state)
-
-    # 窄合同行像漂移：deleted 不是严格 False（整数 0 也必须拒绝）。
-    state = _presence_dedupe_state()
-    _mutate_presence_row(state, 11501, 9, 0)
-    _presence_snapshot_raises("deleted类型不符", state)
-
-    # 窄合同行像漂移：tenant 列不符。
-    state = _presence_dedupe_state()
-    _mutate_presence_row(state, 11501, 10, "other-tenant")
-    _presence_snapshot_raises("tenant不符", state)
-
-
-def test_presence_dedupe_confirmation_sha_binds_rows_and_ids():
-    """confirmation_sha 必须确定且绑定二十行完整内容与保留/删除 ID，任一来源漂移即变化。"""
-
-    base_snapshot = pu._presence_dedupe_capture_snapshot(
-        _PresenceDedupeConn(_presence_dedupe_state()), pu.PRESENCE_DEDUPE_WORK_ID,
-    )
-    base_sha = pu._presence_dedupe_confirmation_sha(base_snapshot)
-    again_snapshot = pu._presence_dedupe_capture_snapshot(
-        _PresenceDedupeConn(_presence_dedupe_state()), pu.PRESENCE_DEDUPE_WORK_ID,
-    )
-    check(
-        "presence-dedupe-confirmation_sha确定性",
-        base_sha == pu._presence_dedupe_confirmation_sha(again_snapshot)
-        and len(base_sha) == 64,
-        detail=base_sha,
-    )
-    for label, row_id, column, value in (
-        ("observation", 11582, 6, "被篡改的观察"),
-        ("保留行ID", 11501, 0, 11001),
-        ("creator", 11501, 7, "篡改者"),
-    ):
-        state = _presence_dedupe_state()
-        _mutate_presence_row(state, row_id, column, value)
-        drifted = pu._presence_dedupe_capture_snapshot(
-            _PresenceDedupeConn(state), pu.PRESENCE_DEDUPE_WORK_ID,
-        )
-        check(
-            f"presence-dedupe-sha感知{label}漂移",
-            pu._presence_dedupe_confirmation_sha(drifted) != base_sha,
-        )
-
-
-def test_presence_dedupe_cli_preview_read_only():
-    """preview：RR READ ONLY 输出十组摘要与 SHA，不写任何表，不触嵌入/模型。"""
-
-    state = _presence_dedupe_state()
-    result, _ = _presence_dedupe_cli(state, ["--work-id", "8", "--preview"])
-    check("presence-dedupe-preview成功", result.exit_code == 0, detail=result.output)
-    data = json.loads(result.output)
-    check(
-        "presence-dedupe-preview摘要完整",
-        data["mode"] == "preview"
-        and data["work_id"] == pu.PRESENCE_DEDUPE_WORK_ID
-        and data["group_count"] == pu.PRESENCE_DEDUPE_GROUP_COUNT
-        and data["row_count"] == pu.PRESENCE_DEDUPE_ROW_COUNT
-        and data["delete_ids"] == sorted(pu.PRESENCE_DEDUPE_DELETE_IDS)
-        and data["keep_ids"] == list(range(11501, 11511))
-        and len(data["confirmation_sha"]) == 64
-        and data["deleted_ids"] is None
-        and data["audit_rows_written"] == 0
-        and len(data["groups"]) == pu.PRESENCE_DEDUPE_GROUP_COUNT,
-        detail=result.output[:500],
-    )
-    check(
-        "presence-dedupe-preview不输出observation原文",
-        "观察A0" not in result.output and "观察B0" not in result.output,
-    )
-    check(
-        "presence-dedupe-preview零副作用",
-        state["writes"] == [] and state["commits"] == 0
-        and all(not row[9] for row in state["rows"]),
-    )
-    check(
-        "presence-dedupe-preview首条SQL为RR只读",
-        state["queries"][0] == "SET TRANSACTION ISOLATION LEVEL REPEATABLE READ, READ ONLY",
-        detail=str(state["queries"][:2]),
-    )
-    check(
-        "presence-dedupe-preview读快照不加行锁",
-        not any(query.endswith("FOR UPDATE") for query in state["queries"]),
-        detail=str(state["queries"]),
-    )
-
-
-def test_presence_dedupe_cli_execute_soft_deletes_exactly_ten():
-    """execute：锁内重算精确匹配后同事务只软删十个精确 ID；其他域零写入。"""
-
-    state = _presence_dedupe_state()
-    sha = _presence_dedupe_preview_sha(_presence_dedupe_state())
-    result, _ = _presence_dedupe_cli(
-        state,
-        ["--work-id", "8", "--execute", "--confirmation-sha", sha,
-         "--confirm-no-live-process"],
-    )
-    check("presence-dedupe-execute成功", result.exit_code == 0, detail=result.output)
-    data = json.loads(result.output)
-    check(
-        "presence-dedupe-execute只软删十个精确ID",
-        data["mode"] == "execute"
-        and data["deleted_ids"] == sorted(pu.PRESENCE_DEDUPE_DELETE_IDS)
-        and sorted(row[0] for row in state["rows"] if row[9] is True)
-        == sorted(pu.PRESENCE_DEDUPE_DELETE_IDS),
-        detail=result.output[:500],
-    )
-    check(
-        "presence-dedupe-execute保留行与单体行全部留存",
-        sorted(row[0] for row in state["rows"] if not row[9])
-        == sorted(list(range(11501, 11511)) + list(range(11400, 11404))),
-        detail=str(sorted(row[0] for row in state["rows"] if not row[9])),
-    )
-    check(
-        "presence-dedupe-execute同事务单次提交",
-        state["writes"] == ["UPDATE example_upgrade_presence"] and state["commits"] == 1,
-        detail=str(state["writes"]),
-    )
-    mutating = [
-        query for query in state["queries"]
-        if query.startswith(("UPDATE ", "INSERT ", "DELETE "))
-    ]
-    check(
-        "presence-dedupe-execute写SQL仅presence软删",
-        len(mutating) == 1
-        and mutating[0].startswith("UPDATE example_upgrade_presence SET deleted=TRUE")
-        and "RETURNING id" in mutating[0],
-        detail=str(mutating),
-    )
-    check(
-        "presence-dedupe-execute锁内带行锁重算快照",
-        any(
-            query.startswith("SELECT id, work_id, window_no, chapter_no")
-            and query.endswith("FOR UPDATE")
-            for query in state["queries"]
-        ),
-        detail=str(state["queries"]),
-    )
-
-
-def test_presence_dedupe_execute_second_run_fails_closed():
-    """execute 成功后二次运行：软删后每组只剩单行不成十组，必须失败关闭且无新写入。"""
-
-    state = _presence_dedupe_state()
-    sha = _presence_dedupe_preview_sha(_presence_dedupe_state())
-    args = ["--work-id", "8", "--execute", "--confirmation-sha", sha,
-            "--confirm-no-live-process"]
-    first, _ = _presence_dedupe_cli(state, args)
-    check("presence-dedupe-首次execute成功", first.exit_code == 0,
-          detail=first.output)
-    second, _ = _presence_dedupe_cli(state, args)
-    check(
-        "presence-dedupe-二次execute失败关闭",
-        second.exit_code != 0 and "冗余组数量不精确" in second.output,
-        detail=second.output,
-    )
-    check(
-        "presence-dedupe-二次execute无新写入",
-        state["writes"] == ["UPDATE example_upgrade_presence"] and state["commits"] == 1,
-        detail=str(state["writes"]),
-    )
-
-
-def test_presence_dedupe_execute_rejects_stale_or_drifted_snapshot():
-    """旧 SHA 与 preview/execute 间快照漂移都必须在写入前失败关闭。"""
-
-    # 旧/伪 confirmation-sha：锁内重算后精确匹配拒绝。
-    state = _presence_dedupe_state()
-    stale, _ = _presence_dedupe_cli(
-        state,
-        ["--work-id", "8", "--execute", "--confirmation-sha", "0" * 64,
-         "--confirm-no-live-process"],
-    )
-    check("presence-dedupe-旧SHA失败关闭", stale.exit_code != 0,
-          detail=stale.output)
-    check(
-        "presence-dedupe-旧SHA零写入",
-        state["writes"] == [] and state["commits"] == 0
-        and all(not row[9] for row in state["rows"]),
-    )
-
-    # preview 与 execute 之间漂移（observation 被改）：锁内重算 SHA 不匹配。
-    sha = _presence_dedupe_preview_sha(_presence_dedupe_state())
-    drifted = _presence_dedupe_state()
-    _mutate_presence_row(drifted, 11582, 6, "被篡改的观察")
-    result, _ = _presence_dedupe_cli(
-        drifted,
-        ["--work-id", "8", "--execute", "--confirmation-sha", sha,
-         "--confirm-no-live-process"],
-    )
-    check(
-        "presence-dedupe-锁内漂移失败关闭",
-        result.exit_code != 0 and "不匹配" in result.output,
-        detail=result.output,
-    )
-    check(
-        "presence-dedupe-锁内漂移零写入",
-        drifted["writes"] == [] and drifted["commits"] == 0
-        and all(not row[9] for row in drifted["rows"]),
-    )
-
-
-def test_presence_dedupe_cli_argument_contracts():
-    """preview/execute 互斥；execute 必须同时提供 SHA 与无活进程确认。"""
-
-    cases = (
-        ("preview与execute同给", ["--work-id", "8", "--preview", "--execute"]),
-        ("两模式都不给", ["--work-id", "8"]),
-        ("execute缺SHA", ["--work-id", "8", "--execute", "--confirm-no-live-process"]),
-        ("execute缺进程确认",
-         ["--work-id", "8", "--execute", "--confirmation-sha", "ab" * 32]),
-        ("work不符", ["--work-id", "9", "--preview"]),
-    )
-    for label, args in cases:
-        state = _presence_dedupe_state()
-        result, _ = _presence_dedupe_cli(state, args)
-        check(f"presence-dedupe-{label}拒绝", result.exit_code != 0,
-              detail=result.output)
-        check(
-            f"presence-dedupe-{label}零副作用",
-            state["writes"] == [] and state["commits"] == 0,
-        )
-
-
-def test_presence_dedupe_cli_preview_fail_closed_on_drift():
-    """CLI 层各类快照漂移都在 preview 被拒，且零副作用。"""
-
-    first_delete_id = sorted(pu.PRESENCE_DEDUPE_DELETE_IDS)[0]
-    cases = (
-        ("额外冗余组", lambda state: state["rows"].extend((
-            (11581, 8, 90, 500, "event", "额外重复组", "观察C",
-             "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT),
-            (11582, 8, 90, 500, "event", "额外重复组", "观察D",
-             "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT),
-        ))),
-        ("observation相同",
-         lambda state: _mutate_presence_row(state, first_delete_id, 6, "观察A0")),
-        ("缺行", lambda state: state["rows"].remove(
-            next(row for row in state["rows"] if row[0] == first_delete_id))),
-        ("tenant不符",
-         lambda state: _mutate_presence_row(state, 11501, 10, "other-tenant")),
-    )
-    for label, mutate in cases:
-        state = _presence_dedupe_state()
-        mutate(state)
-        result, _ = _presence_dedupe_cli(state, ["--work-id", "8", "--preview"])
-        check(f"presence-dedupe-preview{label}拒绝", result.exit_code != 0,
-              detail=result.output)
-        check(
-            f"presence-dedupe-preview{label}零副作用",
-            state["writes"] == [] and state["commits"] == 0,
-        )


 if __name__ == "__main__":
@ -6443,7 +5902,7 @@ if __name__ == "__main__":
               test_prompts_disciplines, test_window_fence_hashes_and_markers,
               test_capture_window_input_rejects_incomplete_or_empty_content,
               test_two_phase_external_calls_and_atomic_embedding_contract,
-               test_recovery_and_redo_rejection_contracts, test_real_pg_rollback_smoke_entry,
+               test_recovery_and_redo_rejection_contracts,
               test_compensation_attempt_baseline_exact_allows_old_non_allowlist_tombstone,
               test_compensation_attempt_baseline_input_drift_blocks,
               test_compensation_attempt_baseline_state_drift_blocks,
@ -6458,21 +5917,8 @@ if __name__ == "__main__":
               test_quality_repair_rejects_final_snapshot_drift_and_bad_embedding,
               test_quality_repair_rejects_invalid_vector_dimension_and_values,
               test_quality_repair_rejects_alias_and_presence_duplicates,
-               test_presence_dedupe_actions_planner_guard,
-               test_presence_dedupe_snapshot_success,
-               test_presence_dedupe_snapshot_fail_closed,
-               test_presence_dedupe_confirmation_sha_binds_rows_and_ids,
-               test_presence_dedupe_cli_preview_read_only,
-               test_presence_dedupe_cli_execute_soft_deletes_exactly_ten,
-               test_presence_dedupe_execute_second_run_fails_closed,
-               test_presence_dedupe_execute_rejects_stale_or_drifted_snapshot,
-               test_presence_dedupe_cli_argument_contracts,
-               test_presence_dedupe_cli_preview_fail_closed_on_drift,
               test_capture_window_state_binds_current_window_and_preserves_other_status,
               test_load_window_material_orders_blocks_like_capture_input,
               test_run_defaults_enforce_safety_contract):
        fn()
-    if os.getenv("MUSE_REAL_PG_ROLLBACK_SMOKE") == "1":
-        print(f"\n全部自测通过：{_passed} 项（含真实 PG public 窗 rollback smoke；未发模型/嵌入调用）")
-    else:
-        print(f"\n全部离线自测通过：{_passed} 项（真实 PG smoke 未启用；未发网络/嵌入/LLM 调用）")
+    print(f"\n全部离线自测通过：{_passed} 项（未发网络/嵌入/LLM 调用）")
--- a/tests/skills/extract-work-knowledge/test_parse_upgrade_pg_smoke.py
+++ b/tests/skills/extract-work-knowledge/test_parse_upgrade_pg_smoke.py
@ -0,0 +1,28 @@
+#!/usr/bin/env python3
+"""Opt-in real PostgreSQL rollback smoke entry point."""
+
+import os
+import pathlib
+import sys
+
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "extract-work-knowledge" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
+import upgrade as pu  # noqa: E402
+
+
+def main():
+    if os.getenv("MUSE_REAL_PG_ROLLBACK_SMOKE") != "1":
+        print(
+            "BLOCKED: set MUSE_REAL_PG_ROLLBACK_SMOKE=1 to run the real PostgreSQL rollback smoke"
+        )
+        return 2
+
+    work_id = int(os.getenv("MUSE_REAL_PG_ROLLBACK_SMOKE_WORK_ID", "8"))
+    pu.real_pg_rollback_smoke(work_id)
+    print(f"PASS: real PostgreSQL rollback smoke completed for work_id={work_id}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
--- a/tests/skills/extract-work-knowledge/test_presence_dedupe.py
+++ b/tests/skills/extract-work-knowledge/test_presence_dedupe.py
@ -0,0 +1,599 @@
+#!/usr/bin/env python3
+"""presence 冗余收口（work8 十组双行）的纯逻辑离线自测。
+
+红线：**不连真实库、不发任何网络/嵌入/LLM 调用**——只使用内存 fake DB 验证
+presence 去重的 planner、快照、确认摘要和 preview/execute 事务合同。
+
+跑法：仓库根目录 `.venv/bin/python tests/skills/extract-work-knowledge/test_presence_dedupe.py`；
+也支持从其他工作目录通过该文件的绝对路径运行。
+"""
+import json
+import pathlib
+import sys
+from contextlib import nullcontext
+from unittest.mock import patch
+
+from click.testing import CliRunner
+
+# 从原生产 scripts 导入 upgrade；测试替身只隔离数据库连接，不替代生产逻辑。
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "extract-work-knowledge" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
+import upgrade as pu  # noqa: E402
+
+_passed = 0
+
+
+def check(name, cond, detail=""):
+    """单项断言：通过打 [PASS]，失败抛 AssertionError（带上下文，令 CI/人工一眼定位）。"""
+    global _passed
+    assert cond, f"[FAIL] {name} :: {detail}"
+    _passed += 1
+    print(f"[PASS] {name}")
+
+
+class _CardResult:
+    """为 presence 去重离线测试提供最小查询结果对象。"""
+
+    def __init__(self, row=None, rows=None):
+        self.row = row
+        self.rows = rows or []
+
+    def fetchone(self):
+        """返回预置的单行结果。"""
+
+        return self.row
+
+    def fetchall(self):
+        """返回预置的多行结果。"""
+
+        return self.rows
+
+
+# ── presence 冗余收口（work8 十组双行：保留 MIN(id) 软删 MAX(id)）──
+
+# 离线 fixture 用字符串时间戳代替数据库驱动的 datetime：窄合同只比较同组两行是否相等，
+# _sha256_json 计算摘要时统一 default=str 转写，二者行为一致。
+_PRESENCE_DEDUPE_BASE_TIME = "2026-07-01 12:00:00"
+
+
+def _presence_dedupe_rows():
+    """构造满足窄合同的 work8 presence 行 fixture。
+
+    十组双行：每组 observation 不同、create_time 相同、恰好一个待删 ID 命中 DELETE_IDS、
+    保留 ID 是组内较小者；另有四个单行，证明快照会跳过不构成冗余的 key。
+    """
+
+    rows = []
+    for index, delete_id in enumerate(sorted(pu.PRESENCE_DEDUPE_DELETE_IDS)):
+        keep_id = 11501 + index
+        entity_type = "location" if index % 2 == 0 else "item"
+        name = f"重复实体{index}"
+        for row_id, observation in ((keep_id, f"观察A{index}"),
+                                    (delete_id, f"观察B{index}")):
+            rows.append((row_id, 8, 57 + index, 403 + index, entity_type, name,
+                         observation, "upgrade", _PRESENCE_DEDUPE_BASE_TIME,
+                         False, pu.TENANT))
+    for index in range(4):
+        rows.append((11400 + index, 8, 10 + index, 20 + index, "character",
+                     f"单体实体{index}", f"单体观察{index}",
+                     "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT))
+    return sorted(rows, key=lambda row: row[0])
+
+
+def _presence_dedupe_state():
+    return {
+        "rows": _presence_dedupe_rows(),
+        "writes": [],
+        "queries": [],
+        "commits": 0,
+    }
+
+
+def _mutate_presence_row(state, row_id, column, value):
+    """替换指定行的单列；元组不可变，整体重建后写回。"""
+
+    state["rows"] = [
+        row[:column] + (value,) + row[column + 1:] if row[0] == row_id else row
+        for row in state["rows"]
+    ]
+
+
+class _PresenceDedupeConn:
+    """presence 冗余收口 fake DB：只实现本命令的读快照与软删 SQL。
+
+    其他任何域（draft/window/alias/card_state/audit/embedding）的 SQL 会落到末尾
+    AssertionError，因此「零副作用」无需逐条枚举禁写语句即可离线断言。fake 只做
+    「未软删」粗过滤；列数/work/tenant/deleted 类型等窄合同行像校验正是离线断言对象，
+    fake 不能代为过滤。
+    """
+
+    def __init__(self, state):
+        self.state = state
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc, tb):
+        return False
+
+    def execute(self, query, params=()):
+        normalized = " ".join(query.split())
+        self.state.setdefault("queries", []).append(normalized)
+        if normalized.startswith("SET TRANSACTION") or normalized.startswith("LOCK TABLE"):
+            return _CardResult()
+        if normalized.startswith(
+                "SELECT id, work_id, window_no, chapter_no, entity_type, name, observation,"):
+            rows = [
+                row for row in sorted(self.state["rows"], key=lambda item: item[0])
+                if len(row) > 9 and not row[9]
+            ]
+            return _CardResult(rows=rows)
+        if normalized.startswith("UPDATE example_upgrade_presence SET deleted=TRUE"):
+            tenant_id, work_id, delete_ids = params
+            assert tenant_id == pu.TENANT
+            delete_set = set(delete_ids)
+            hit = []
+            new_rows = []
+            for row in self.state["rows"]:
+                if row[1] == work_id and row[0] in delete_set and not row[9]:
+                    row = row[:9] + (True,) + row[10:]
+                    hit.append((row[0],))
+                new_rows.append(row)
+            self.state["rows"] = new_rows
+            self.state["writes"].append("UPDATE example_upgrade_presence")
+            return _CardResult(rows=hit)
+        raise AssertionError(f"presence 冗余收口 fake DB 未覆盖 SQL：{normalized}")
+
+    def commit(self):
+        self.state["commits"] += 1
+
+
+def _presence_dedupe_cli(state, args):
+    """离线执行 repair-presence-duplicates：注入 fake DB 与空放同书锁。"""
+
+    conn = _PresenceDedupeConn(state)
+    with patch.object(pu.psycopg, "connect", return_value=conn), \
+            patch.object(pu, "upgrade_work_lock", return_value=nullcontext()):
+        result = CliRunner().invoke(pu.maintenance_cli, ["repair-presence-duplicates"] + args)
+    return result, conn
+
+
+def _presence_dedupe_preview_sha(state):
+    """离线取 preview 的 confirmation_sha，作为 execute 的合法输入。"""
+
+    result, _ = _presence_dedupe_cli(state, ["--work-id", "8", "--preview"])
+    assert result.exit_code == 0, result.output
+    return json.loads(result.output)["confirmation_sha"]
+
+
+def _presence_snapshot_raises(label, state, work_id=pu.PRESENCE_DEDUPE_WORK_ID):
+    """断言快照以 CompensationFenceConflict 拒绝，且没有发出任何写入。"""
+
+    conn = _PresenceDedupeConn(state)
+    try:
+        pu._presence_dedupe_capture_snapshot(conn, work_id)
+    except pu.CompensationFenceConflict as exc:
+        check(f"presence-dedupe-{label}失败关闭", True, detail=str(exc))
+    else:
+        check(f"presence-dedupe-{label}失败关闭", False,
+              detail="未抛 CompensationFenceConflict")
+    check(
+        f"presence-dedupe-{label}零副作用",
+        state["writes"] == [] and state["commits"] == 0,
+    )
+
+
+def test_presence_dedupe_actions_planner_guard():
+    """planner 级收口守卫：同观察幂等留首条、不同观察失败关闭、非 presence 原样透传。"""
+
+    presence_a = ("presence", None, 403, "location", "重复星体", "观察一")
+    presence_a_dup = ("presence", None, 403, "location", "重复星体", "观察一")
+    presence_b = ("presence", None, 404, "item", "镜面护盾", "观察二")
+    alias_action = ("alias", 7, "规范名", "别名", "ai")
+    new_action = ("new", -1, {"名称": "新实体"}, set(), [])
+    actions = [presence_a, alias_action, presence_a_dup, new_action, presence_b, ()]
+    original = list(actions)
+    deduped = pu._dedupe_presence_actions(actions)
+    check(
+        "presence-dedupe-planner同观察去重保留首条",
+        deduped == [presence_a, alias_action, new_action, presence_b, ()]
+        and deduped[0] is presence_a,
+        detail=str(deduped),
+    )
+    check("presence-dedupe-重复应用幂等",
+          pu._dedupe_presence_actions(deduped) == deduped)
+    check("presence-dedupe-不修改输入列表", actions == original)
+
+    conflict = [presence_a,
+                ("presence", None, 403, "location", "重复星体", "另一个观察")]
+    try:
+        pu._dedupe_presence_actions(conflict)
+    except RuntimeError as exc:
+        check("presence-dedupe-不同观察失败关闭",
+              "observation 冲突" in str(exc), detail=str(exc))
+    else:
+        check("presence-dedupe-不同观察失败关闭", False, detail="未抛 RuntimeError")
+
+    for label, bad, keyword in (
+        ("结构缺一元", ("presence", None, 403, "location", "重复星体"),
+         "结构非法"),
+        ("结构多一元",
+         ("presence", None, 403, "location", "重复星体", "观察一", "extra"),
+         "结构非法"),
+        ("observation非字符串", ("presence", None, 403, "location", "重复星体", None),
+         "必须是字符串"),
+    ):
+        try:
+            pu._dedupe_presence_actions([bad])
+        except RuntimeError as exc:
+            check(f"presence-dedupe-{label}抛错", keyword in str(exc), detail=str(exc))
+        else:
+            check(f"presence-dedupe-{label}抛错", False, detail="未抛 RuntimeError")
+
+
+def test_presence_dedupe_snapshot_success():
+    """快照成功路径：恰好 10 组/20 行，待删集合精确等于 DELETE_IDS，保留为组内较小者。"""
+
+    state = _presence_dedupe_state()
+    snapshot = pu._presence_dedupe_capture_snapshot(
+        _PresenceDedupeConn(state), pu.PRESENCE_DEDUPE_WORK_ID,
+    )
+    check(
+        "presence-dedupe-快照组数行数精确",
+        len(snapshot["groups"]) == pu.PRESENCE_DEDUPE_GROUP_COUNT
+        and len(snapshot["rows"]) == pu.PRESENCE_DEDUPE_ROW_COUNT
+        and snapshot["contract"] == pu.PRESENCE_DEDUPE_CONTRACT
+        and snapshot["work_id"] == pu.PRESENCE_DEDUPE_WORK_ID,
+    )
+    check(
+        "presence-dedupe-待删集合精确等于DELETE_IDS",
+        snapshot["delete_ids"] == sorted(pu.PRESENCE_DEDUPE_DELETE_IDS),
+        detail=str(snapshot["delete_ids"]),
+    )
+    check(
+        "presence-dedupe-保留为组内较小ID",
+        snapshot["keep_ids"] == list(range(11501, 11511))
+        and all(group["keep_id"] < group["delete_id"] for group in snapshot["groups"])
+        and all(len(group["rows"]) == 2 for group in snapshot["groups"]),
+        detail=str(snapshot["keep_ids"]),
+    )
+    check(
+        "presence-dedupe-行按id排序且排除单体行",
+        [row[0] for row in snapshot["rows"]]
+        == sorted(row[0] for row in snapshot["rows"])
+        and not any(row[0] in range(11400, 11404) for row in snapshot["rows"]),
+    )
+
+
+def test_presence_dedupe_snapshot_fail_closed():
+    """各类窄合同漂移都必须在写入前以 CompensationFenceConflict 逐条拒绝。"""
+
+    # 额外冗余组：组数 11 ≠ 10（新增组复用合同内待删 ID，确保失败点落在组数校验）。
+    state = _presence_dedupe_state()
+    state["rows"].extend((
+        (11581, 8, 90, 500, "event", "额外重复组", "观察C",
+         "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT),
+        (11582, 8, 90, 500, "event", "额外重复组", "观察D",
+         "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT),
+    ))
+    _presence_snapshot_raises("额外冗余组", state)
+
+    # 三行组不是「恰好两行」。
+    state = _presence_dedupe_state()
+    state["rows"].append(
+        (11601, 8, 57, 403, "location", "重复实体0", "观察E",
+         "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT)
+    )
+    _presence_snapshot_raises("三行冗余组", state)
+
+    # 缺行：一组只剩单行被跳过 → 组数 9 ≠ 10。
+    state = _presence_dedupe_state()
+    state["rows"] = [row for row in state["rows"]
+                     if row[0] != sorted(pu.PRESENCE_DEDUPE_DELETE_IDS)[0]]
+    _presence_snapshot_raises("缺行", state)
+
+    # 同组两行 create_time 漂移。
+    state = _presence_dedupe_state()
+    _mutate_presence_row(state, 11501, 8, "2026-07-02 09:30:00")
+    _presence_snapshot_raises("create_time漂移", state)
+
+    # 待删 ID 命中 0：组内两行都不在 DELETE_IDS。
+    state = _presence_dedupe_state()
+    _mutate_presence_row(state, 11501, 0, 12001)
+    _mutate_presence_row(state, 11582, 0, 12002)
+    _presence_snapshot_raises("待删ID无命中", state)
+
+    # 待删 ID 命中 2：组内两行都在 DELETE_IDS。
+    state = _presence_dedupe_state()
+    _mutate_presence_row(state, 11501, 0, 11693)
+    _presence_snapshot_raises("待删ID双命中", state)
+
+    # 保留不是组内较小者：待删 ID 反而是组内较小者。
+    state = _presence_dedupe_state()
+    _mutate_presence_row(state, 11501, 0, 11900)
+    _presence_snapshot_raises("保留非组内较小", state)
+
+    # 同组两行 observation 相同（byte-exact 重复违反窄合同）。
+    state = _presence_dedupe_state()
+    _mutate_presence_row(state, 11582, 6, "观察A0")
+    _presence_snapshot_raises("observation相同", state)
+
+    # work_id ≠ 8 拒绝。
+    _presence_snapshot_raises("work不符", _presence_dedupe_state(), work_id=9)
+
+    # 窄合同行像漂移：列数不足。
+    state = _presence_dedupe_state()
+    state["rows"] = [row[:10] if row[0] == 11501 else row for row in state["rows"]]
+    _presence_snapshot_raises("列数不符", state)
+
+    # 窄合同行像漂移：work 列与查询目标不符。
+    state = _presence_dedupe_state()
+    _mutate_presence_row(state, 11501, 1, 9)
+    _presence_snapshot_raises("work列不符", state)
+
+    # 窄合同行像漂移：deleted 不是严格 False（整数 0 也必须拒绝）。
+    state = _presence_dedupe_state()
+    _mutate_presence_row(state, 11501, 9, 0)
+    _presence_snapshot_raises("deleted类型不符", state)
+
+    # 窄合同行像漂移：tenant 列不符。
+    state = _presence_dedupe_state()
+    _mutate_presence_row(state, 11501, 10, "other-tenant")
+    _presence_snapshot_raises("tenant不符", state)
+
+
+def test_presence_dedupe_confirmation_sha_binds_rows_and_ids():
+    """confirmation_sha 必须确定且绑定二十行完整内容与保留/删除 ID，任一来源漂移即变化。"""
+
+    base_snapshot = pu._presence_dedupe_capture_snapshot(
+        _PresenceDedupeConn(_presence_dedupe_state()), pu.PRESENCE_DEDUPE_WORK_ID,
+    )
+    base_sha = pu._presence_dedupe_confirmation_sha(base_snapshot)
+    again_snapshot = pu._presence_dedupe_capture_snapshot(
+        _PresenceDedupeConn(_presence_dedupe_state()), pu.PRESENCE_DEDUPE_WORK_ID,
+    )
+    check(
+        "presence-dedupe-confirmation_sha确定性",
+        base_sha == pu._presence_dedupe_confirmation_sha(again_snapshot)
+        and len(base_sha) == 64,
+        detail=base_sha,
+    )
+    for label, row_id, column, value in (
+        ("observation", 11582, 6, "被篡改的观察"),
+        ("保留行ID", 11501, 0, 11001),
+        ("creator", 11501, 7, "篡改者"),
+    ):
+        state = _presence_dedupe_state()
+        _mutate_presence_row(state, row_id, column, value)
+        drifted = pu._presence_dedupe_capture_snapshot(
+            _PresenceDedupeConn(state), pu.PRESENCE_DEDUPE_WORK_ID,
+        )
+        check(
+            f"presence-dedupe-sha感知{label}漂移",
+            pu._presence_dedupe_confirmation_sha(drifted) != base_sha,
+        )
+
+
+def test_presence_dedupe_cli_preview_read_only():
+    """preview：RR READ ONLY 输出十组摘要与 SHA，不写任何表，不触嵌入/模型。"""
+
+    state = _presence_dedupe_state()
+    result, _ = _presence_dedupe_cli(state, ["--work-id", "8", "--preview"])
+    check("presence-dedupe-preview成功", result.exit_code == 0, detail=result.output)
+    data = json.loads(result.output)
+    check(
+        "presence-dedupe-preview摘要完整",
+        data["mode"] == "preview"
+        and data["work_id"] == pu.PRESENCE_DEDUPE_WORK_ID
+        and data["group_count"] == pu.PRESENCE_DEDUPE_GROUP_COUNT
+        and data["row_count"] == pu.PRESENCE_DEDUPE_ROW_COUNT
+        and data["delete_ids"] == sorted(pu.PRESENCE_DEDUPE_DELETE_IDS)
+        and data["keep_ids"] == list(range(11501, 11511))
+        and len(data["confirmation_sha"]) == 64
+        and data["deleted_ids"] is None
+        and data["audit_rows_written"] == 0
+        and len(data["groups"]) == pu.PRESENCE_DEDUPE_GROUP_COUNT,
+        detail=result.output[:500],
+    )
+    check(
+        "presence-dedupe-preview不输出observation原文",
+        "观察A0" not in result.output and "观察B0" not in result.output,
+    )
+    check(
+        "presence-dedupe-preview零副作用",
+        state["writes"] == [] and state["commits"] == 0
+        and all(not row[9] for row in state["rows"]),
+    )
+    check(
+        "presence-dedupe-preview首条SQL为RR只读",
+        state["queries"][0] == "SET TRANSACTION ISOLATION LEVEL REPEATABLE READ, READ ONLY",
+        detail=str(state["queries"][:2]),
+    )
+    check(
+        "presence-dedupe-preview读快照不加行锁",
+        not any(query.endswith("FOR UPDATE") for query in state["queries"]),
+        detail=str(state["queries"]),
+    )
+
+
+def test_presence_dedupe_cli_execute_soft_deletes_exactly_ten():
+    """execute：锁内重算精确匹配后同事务只软删十个精确 ID；其他域零写入。"""
+
+    state = _presence_dedupe_state()
+    sha = _presence_dedupe_preview_sha(_presence_dedupe_state())
+    result, _ = _presence_dedupe_cli(
+        state,
+        ["--work-id", "8", "--execute", "--confirmation-sha", sha,
+         "--confirm-no-live-process"],
+    )
+    check("presence-dedupe-execute成功", result.exit_code == 0, detail=result.output)
+    data = json.loads(result.output)
+    check(
+        "presence-dedupe-execute只软删十个精确ID",
+        data["mode"] == "execute"
+        and data["deleted_ids"] == sorted(pu.PRESENCE_DEDUPE_DELETE_IDS)
+        and sorted(row[0] for row in state["rows"] if row[9] is True)
+        == sorted(pu.PRESENCE_DEDUPE_DELETE_IDS),
+        detail=result.output[:500],
+    )
+    check(
+        "presence-dedupe-execute保留行与单体行全部留存",
+        sorted(row[0] for row in state["rows"] if not row[9])
+        == sorted(list(range(11501, 11511)) + list(range(11400, 11404))),
+        detail=str(sorted(row[0] for row in state["rows"] if not row[9])),
+    )
+    check(
+        "presence-dedupe-execute同事务单次提交",
+        state["writes"] == ["UPDATE example_upgrade_presence"] and state["commits"] == 1,
+        detail=str(state["writes"]),
+    )
+    mutating = [
+        query for query in state["queries"]
+        if query.startswith(("UPDATE ", "INSERT ", "DELETE "))
+    ]
+    check(
+        "presence-dedupe-execute写SQL仅presence软删",
+        len(mutating) == 1
+        and mutating[0].startswith("UPDATE example_upgrade_presence SET deleted=TRUE")
+        and "RETURNING id" in mutating[0],
+        detail=str(mutating),
+    )
+    check(
+        "presence-dedupe-execute锁内带行锁重算快照",
+        any(
+            query.startswith("SELECT id, work_id, window_no, chapter_no")
+            and query.endswith("FOR UPDATE")
+            for query in state["queries"]
+        ),
+        detail=str(state["queries"]),
+    )
+
+
+def test_presence_dedupe_execute_second_run_fails_closed():
+    """execute 成功后二次运行：软删后每组只剩单行不成十组，必须失败关闭且无新写入。"""
+
+    state = _presence_dedupe_state()
+    sha = _presence_dedupe_preview_sha(_presence_dedupe_state())
+    args = ["--work-id", "8", "--execute", "--confirmation-sha", sha,
+            "--confirm-no-live-process"]
+    first, _ = _presence_dedupe_cli(state, args)
+    check("presence-dedupe-首次execute成功", first.exit_code == 0,
+          detail=first.output)
+    second, _ = _presence_dedupe_cli(state, args)
+    check(
+        "presence-dedupe-二次execute失败关闭",
+        second.exit_code != 0 and "冗余组数量不精确" in second.output,
+        detail=second.output,
+    )
+    check(
+        "presence-dedupe-二次execute无新写入",
+        state["writes"] == ["UPDATE example_upgrade_presence"] and state["commits"] == 1,
+        detail=str(state["writes"]),
+    )
+
+
+def test_presence_dedupe_execute_rejects_stale_or_drifted_snapshot():
+    """旧 SHA 与 preview/execute 间快照漂移都必须在写入前失败关闭。"""
+
+    # 旧/伪 confirmation-sha：锁内重算后精确匹配拒绝。
+    state = _presence_dedupe_state()
+    stale, _ = _presence_dedupe_cli(
+        state,
+        ["--work-id", "8", "--execute", "--confirmation-sha", "0" * 64,
+         "--confirm-no-live-process"],
+    )
+    check("presence-dedupe-旧SHA失败关闭", stale.exit_code != 0,
+          detail=stale.output)
+    check(
+        "presence-dedupe-旧SHA零写入",
+        state["writes"] == [] and state["commits"] == 0
+        and all(not row[9] for row in state["rows"]),
+    )
+
+    # preview 与 execute 之间漂移（observation 被改）：锁内重算 SHA 不匹配。
+    sha = _presence_dedupe_preview_sha(_presence_dedupe_state())
+    drifted = _presence_dedupe_state()
+    _mutate_presence_row(drifted, 11582, 6, "被篡改的观察")
+    result, _ = _presence_dedupe_cli(
+        drifted,
+        ["--work-id", "8", "--execute", "--confirmation-sha", sha,
+         "--confirm-no-live-process"],
+    )
+    check(
+        "presence-dedupe-锁内漂移失败关闭",
+        result.exit_code != 0 and "不匹配" in result.output,
+        detail=result.output,
+    )
+    check(
+        "presence-dedupe-锁内漂移零写入",
+        drifted["writes"] == [] and drifted["commits"] == 0
+        and all(not row[9] for row in drifted["rows"]),
+    )
+
+
+def test_presence_dedupe_cli_argument_contracts():
+    """preview/execute 互斥；execute 必须同时提供 SHA 与无活进程确认。"""
+
+    cases = (
+        ("preview与execute同给", ["--work-id", "8", "--preview", "--execute"]),
+        ("两模式都不给", ["--work-id", "8"]),
+        ("execute缺SHA", ["--work-id", "8", "--execute", "--confirm-no-live-process"]),
+        ("execute缺进程确认",
+         ["--work-id", "8", "--execute", "--confirmation-sha", "ab" * 32]),
+        ("work不符", ["--work-id", "9", "--preview"]),
+    )
+    for label, args in cases:
+        state = _presence_dedupe_state()
+        result, _ = _presence_dedupe_cli(state, args)
+        check(f"presence-dedupe-{label}拒绝", result.exit_code != 0,
+              detail=result.output)
+        check(
+            f"presence-dedupe-{label}零副作用",
+            state["writes"] == [] and state["commits"] == 0,
+        )
+
+
+def test_presence_dedupe_cli_preview_fail_closed_on_drift():
+    """CLI 层各类快照漂移都在 preview 被拒，且零副作用。"""
+
+    first_delete_id = sorted(pu.PRESENCE_DEDUPE_DELETE_IDS)[0]
+    cases = (
+        ("额外冗余组", lambda state: state["rows"].extend((
+            (11581, 8, 90, 500, "event", "额外重复组", "观察C",
+             "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT),
+            (11582, 8, 90, 500, "event", "额外重复组", "观察D",
+             "upgrade", _PRESENCE_DEDUPE_BASE_TIME, False, pu.TENANT),
+        ))),
+        ("observation相同",
+         lambda state: _mutate_presence_row(state, first_delete_id, 6, "观察A0")),
+        ("缺行", lambda state: state["rows"].remove(
+            next(row for row in state["rows"] if row[0] == first_delete_id))),
+        ("tenant不符",
+         lambda state: _mutate_presence_row(state, 11501, 10, "other-tenant")),
+    )
+    for label, mutate in cases:
+        state = _presence_dedupe_state()
+        mutate(state)
+        result, _ = _presence_dedupe_cli(state, ["--work-id", "8", "--preview"])
+        check(f"presence-dedupe-preview{label}拒绝", result.exit_code != 0,
+              detail=result.output)
+        check(
+            f"presence-dedupe-preview{label}零副作用",
+            state["writes"] == [] and state["commits"] == 0,
+        )
+
+
+if __name__ == "__main__":
+    for fn in (test_presence_dedupe_actions_planner_guard,
+               test_presence_dedupe_snapshot_success,
+               test_presence_dedupe_snapshot_fail_closed,
+               test_presence_dedupe_confirmation_sha_binds_rows_and_ids,
+               test_presence_dedupe_cli_preview_read_only,
+               test_presence_dedupe_cli_execute_soft_deletes_exactly_ten,
+               test_presence_dedupe_execute_second_run_fails_closed,
+               test_presence_dedupe_execute_rejects_stale_or_drifted_snapshot,
+               test_presence_dedupe_cli_argument_contracts,
+               test_presence_dedupe_cli_preview_fail_closed_on_drift):
+        fn()
+    print(f"\n全部离线自测通过：{_passed} 项（未发网络/嵌入/LLM 调用）")
--- a/.claude/skills/extract-work-knowledge/scripts/test_upgrade_work_lock_offline.py
+++ b/.claude/skills/extract-work-knowledge/scripts/test_upgrade_work_lock_offline.py
@ -6,7 +6,8 @@ import sys
 import unittest


-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "extract-work-knowledge" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 from upgrade_work_lock import (  # noqa: E402
--- a/.claude/skills/freeze-context/scripts/test_audit_leakage.py
+++ b/.claude/skills/freeze-context/scripts/test_audit_leakage.py
@ -8,7 +8,9 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "freeze-context" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 from audit_leakage import audit_snapshot  # noqa: E402


--- a/.claude/skills/freeze-context/scripts/test_build_snapshot.py
+++ b/.claude/skills/freeze-context/scripts/test_build_snapshot.py
@ -5,7 +5,9 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "freeze-context" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 from build_snapshot import (  # noqa: E402
    SnapshotError,
    build_snapshot,
--- a/.claude/skills/freeze-context/scripts/test_check_snapshot.py
+++ b/.claude/skills/freeze-context/scripts/test_check_snapshot.py
@ -5,7 +5,9 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "freeze-context" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 from check_snapshot import (  # noqa: E402
    STATUS_BLOCKED_AUTHORIZATION,
    STATUS_INVALID_ARM_DIFF,
--- a/.claude/skills/freeze-context/scripts/test_load_reference_work.py
+++ b/.claude/skills/freeze-context/scripts/test_load_reference_work.py
@ -10,7 +10,9 @@ import sys
 import unittest
 from unittest.mock import patch

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "freeze-context" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 import load_reference_work as loader  # noqa: E402
 from load_reference_work import (  # noqa: E402
    AdapterError,
--- a/.claude/skills/maintain-work-extraction/scripts/test_backup_upgrade_work_offline.py
+++ b/.claude/skills/maintain-work-extraction/scripts/test_backup_upgrade_work_offline.py
@ -12,8 +12,9 @@ from decimal import Decimal
 from unittest import mock


-HERE = pathlib.Path(__file__).resolve().parent
-sys.path.insert(0, str(HERE))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "maintain-work-extraction" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 import backup_upgrade_work as backup  # noqa: E402

--- a/.claude/skills/maintain-work-extraction/scripts/test_reset_upgrade_work_offline.py
+++ b/.claude/skills/maintain-work-extraction/scripts/test_reset_upgrade_work_offline.py
@ -11,9 +11,11 @@ from unittest.mock import Mock, patch
 from click.testing import CliRunner


-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "maintain-work-extraction" / "scripts"
+EXTRACTION_SCRIPTS = PROJECT_ROOT / ".claude" / "skills" / "extract-work-knowledge" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))
-sys.path.insert(0, str(SCRIPT_DIR.parents[1] / "extract-work-knowledge" / "scripts"))
+sys.path.insert(0, str(EXTRACTION_SCRIPTS))

 import upgrade as parse  # noqa: E402
 import backup_upgrade_work as backup  # noqa: E402
--- a/.claude/skills/plan-chapter/scripts/test_contract.py
+++ b/.claude/skills/plan-chapter/scripts/test_contract.py
@ -9,11 +9,15 @@ from pathlib import Path
 import unittest


-ROOT = Path(__file__).resolve().parents[4]
-SKILL = (ROOT / ".claude/skills/plan-chapter/SKILL.md").read_text(encoding="utf-8")
-PLANNER = (ROOT / ".claude/agents/planner.md").read_text(encoding="utf-8")
-CHAINS = (ROOT / "meta/chains/README.md").read_text(encoding="utf-8")
-SCHEMA = (ROOT / "meta/schemas/fine_outline.yaml").read_text(encoding="utf-8")
+PROJECT_ROOT = Path(__file__).resolve().parents[3]
+SKILL = (
+    PROJECT_ROOT / ".claude" / "skills" / "plan-chapter" / "SKILL.md"
+).read_text(encoding="utf-8")
+PLANNER = (PROJECT_ROOT / ".claude" / "agents" / "planner.md").read_text(encoding="utf-8")
+CHAINS = (PROJECT_ROOT / "meta" / "chains" / "README.md").read_text(encoding="utf-8")
+SCHEMA = (
+    PROJECT_ROOT / "meta" / "schemas" / "fine_outline.yaml"
+).read_text(encoding="utf-8")

 # 与 meta/schemas/fine_outline.yaml 的必填/推荐集保持一致。
 REQUIRED_FIELDS = (
--- a/.claude/skills/plan-story/scripts/test_field_coverage.py
+++ b/.claude/skills/plan-story/scripts/test_field_coverage.py
@ -10,8 +10,10 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parents[2] / "access-database" / "scripts"))
+
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "plan-story" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 from persist_planning import check_field_coverage, write_section, _load_schema  # noqa: E402

--- a/.claude/skills/plan-story/scripts/test_record_planning_execution.py
+++ b/.claude/skills/plan-story/scripts/test_record_planning_execution.py
@ -1,11 +1,13 @@
 #!/usr/bin/env python3
 """规划执行留痕脚本的离线合同测试。"""
+
 import pathlib
 import sys
 import unittest


-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "plan-story" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 import record_planning_execution as recorder  # noqa: E402
--- a/.claude/skills/plan-story/scripts/test_repair_deterministic_receipt.py
+++ b/.claude/skills/plan-story/scripts/test_repair_deterministic_receipt.py
@ -1,11 +1,13 @@
 #!/usr/bin/env python3
 """确定性规划回执修正的离线合同测试。"""
+
 import pathlib
 import sys
 import unittest


-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "plan-story" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 import repair_deterministic_receipt as repair  # noqa: E402
--- a/.claude/skills/plan-story/scripts/test_select_patterns_offline.py
+++ b/.claude/skills/plan-story/scripts/test_select_patterns_offline.py
@ -1,12 +1,14 @@
 #!/usr/bin/env python3
 """范式选择投影的离线测试，不调用向量服务。"""
+
 import pathlib
 import sys
 import unittest
 from unittest.mock import patch


-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "plan-story" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 import select_patterns as selector  # noqa: E402
--- a/.claude/skills/prevent-ai-flavor/scripts/test_prevent_ai_flavor.py
+++ b/.claude/skills/prevent-ai-flavor/scripts/test_prevent_ai_flavor.py
@ -5,7 +5,9 @@ import sys
 import unittest
 from unittest.mock import patch

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "prevent-ai-flavor" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))

 import prevent_ai_flavor as prev  # noqa: E402

--- a/.claude/skills/record-run-evidence/scripts/test_file_cas.py
+++ b/.claude/skills/record-run-evidence/scripts/test_file_cas.py
@ -12,7 +12,8 @@ import threading
 import unittest
 from unittest import mock

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "record-run-evidence" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 from file_cas import CasConflictError, CasRecoveryError, FileCasStore  # noqa: E402
--- a/.claude/skills/record-run-evidence/scripts/test_persist_raw.py
+++ b/.claude/skills/record-run-evidence/scripts/test_persist_raw.py
@ -3,7 +3,9 @@
 import pathlib
 import sys

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "record-run-evidence" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 from persist_raw import _check_no_secrets  # noqa: E402


--- a/.claude/skills/record-run-evidence/scripts/test_raw_vault.py
+++ b/.claude/skills/record-run-evidence/scripts/test_raw_vault.py
@ -13,7 +13,8 @@ import unittest
 from datetime import datetime, timedelta, timezone
 from unittest import mock

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "record-run-evidence" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 import raw_vault  # noqa: E402
--- a/.claude/skills/record-run-evidence/scripts/test_record_failed_run.py
+++ b/.claude/skills/record-run-evidence/scripts/test_record_failed_run.py
@ -5,7 +5,8 @@ import sys
 import unittest


-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "record-run-evidence" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 import record_failed_run as failure  # noqa: E402
--- a/.claude/skills/record-run-evidence/scripts/test_repair_receipt_evidence.py
+++ b/.claude/skills/record-run-evidence/scripts/test_repair_receipt_evidence.py
@ -5,7 +5,8 @@ import sys
 import unittest


-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "record-run-evidence" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 import repair_receipt_evidence as repair  # noqa: E402
--- a/.claude/skills/record-run-evidence/scripts/test_run_registry.py
+++ b/.claude/skills/record-run-evidence/scripts/test_run_registry.py
@ -3,7 +3,9 @@
 import pathlib
 import sys

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "record-run-evidence" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
 import run_registry  # noqa: E402


--- a/.claude/skills/revise-ai-flavor/scripts/test_revise_ai_flavor.py
+++ b/.claude/skills/revise-ai-flavor/scripts/test_revise_ai_flavor.py
@ -7,8 +7,11 @@ import tempfile
 import unittest
 from unittest.mock import patch

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parents[3] / "skills" / "diagnose-ai-flavor" / "scripts"))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "revise-ai-flavor" / "scripts"
+DIAGNOSE_SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "diagnose-ai-flavor" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
+sys.path.insert(0, str(DIAGNOSE_SCRIPT_DIR))

 import revise_ai_flavor as rev  # noqa: E402
 import diagnose_ai_flavor as diag  # noqa: E402
--- a/.claude/skills/score-content-quality/scripts/test_lesson_registry_db.py
+++ b/.claude/skills/score-content-quality/scripts/test_lesson_registry_db.py
@ -10,7 +10,7 @@
 测试行用 unittest-lesson- 前缀的 run_id 隔离，结束物理清理（本表可变）。

 跑法（需 Tailscale 内网可达 muse-example）：
-.venv/bin/python .claude/skills/score-content-quality/scripts/test_lesson_registry_db.py
+.venv/bin/python tests/skills/score-content-quality/test_lesson_registry_db.py
 """

 from __future__ import annotations
@ -19,9 +19,13 @@ import pathlib
 import sys
 import uuid

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
-for path in (SCRIPT_DIR, SKILLS_DIR / "access-database" / "scripts"):
+from psycopg.errors import RaiseException
+
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "score-content-quality" / "scripts"
+DB_SCRIPT_DIR = SKILLS_DIR / "access-database" / "scripts"
+for path in (SCRIPT_DIR, DB_SCRIPT_DIR):
    if str(path) not in sys.path:
        sys.path.insert(0, str(path))

@ -96,7 +100,7 @@ def test_no_auto_promotion() -> None:
        raise AssertionError("触发器必须拒绝自动升格")
    except AssertionError:
        raise
-    except Exception:
+    except RaiseException:
        pass
    rows = list_lessons(status="proposed")
    assert any(row["id"] == out["lesson_id"] for row in rows)
--- a/.claude/skills/score-content-quality/scripts/test_rubric.py
+++ b/.claude/skills/score-content-quality/scripts/test_rubric.py
@ -6,7 +6,10 @@ import inspect
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "score-content-quality" / "scripts"
+if str(SCRIPT_DIR) not in sys.path:
+    sys.path.insert(0, str(SCRIPT_DIR))
 from fine_outline_rubric import (  # noqa: E402
    DIMENSIONS,
    RUBRIC_PROFILE,
--- a/.claude/skills/score-content-quality/scripts/test_run_writer_blind_judge.py
+++ b/.claude/skills/score-content-quality/scripts/test_run_writer_blind_judge.py
@ -11,7 +11,8 @@ import sys
 import unittest
 from typing import Any, Mapping

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "score-content-quality" / "scripts"
 if str(SCRIPT_DIR) not in sys.path:
    sys.path.insert(0, str(SCRIPT_DIR))

--- a/.claude/skills/score-content-quality/scripts/test_writer_rubric.py
+++ b/.claude/skills/score-content-quality/scripts/test_writer_rubric.py
@ -7,7 +7,10 @@ import pathlib
 import sys
 import unittest

-sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "score-content-quality" / "scripts"
+if str(SCRIPT_DIR) not in sys.path:
+    sys.path.insert(0, str(SCRIPT_DIR))

 from writer_rubric import (  # noqa: E402
    COMMON_DIMENSIONS,
--- a/.claude/skills/search-knowledge/scripts/test_search.py
+++ b/.claude/skills/search-knowledge/scripts/test_search.py
@ -3,8 +3,16 @@

 from __future__ import annotations

+import pathlib
+import sys
 import unittest

+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "search-knowledge" / "scripts"
+EMBED_SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "embed-knowledge" / "scripts"
+sys.path.insert(0, str(SCRIPT_DIR))
+sys.path.insert(0, str(EMBED_SCRIPT_DIR))
+
 from search import search_cards


--- a/.claude/skills/write-next-chapter/scripts/test_candidate_cas.py
+++ b/.claude/skills/write-next-chapter/scripts/test_candidate_cas.py
@ -6,7 +6,7 @@
 - transition/start_next 的守卫与 token 递增；
 - 该存储可直接替换 InMemoryCasStateStore 驱动 run_writer_pipeline 全链。

-跑法（仓库根目录）：.venv/bin/python .claude/skills/write-next-chapter/scripts/test_candidate_cas.py
+跑法（仓库根目录）：.venv/bin/python tests/skills/write-next-chapter/test_candidate_cas.py
 """

 from __future__ import annotations
@ -16,11 +16,14 @@ import pathlib
 import sys
 import unittest

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+TEST_DIR = PROJECT_ROOT / "tests" / "skills" / "write-next-chapter"
+CHECK_TEST_DIR = PROJECT_ROOT / "tests" / "skills" / "check-content-consistency"
+SCRIPT_DIR = SKILLS_DIR / "write-next-chapter" / "scripts"
 DETECT_DIR = SKILLS_DIR / "check-content-consistency" / "scripts"
 READ_CONTEXT_DIR = SKILLS_DIR / "assemble-context" / "scripts"
-for path in (SCRIPT_DIR, DETECT_DIR, READ_CONTEXT_DIR):
+for path in (READ_CONTEXT_DIR, DETECT_DIR, SCRIPT_DIR, TEST_DIR, CHECK_TEST_DIR):
    if str(path) not in sys.path:
        sys.path.insert(0, str(path))

--- a/.claude/skills/write-next-chapter/scripts/test_candidate_cas_db.py
+++ b/.claude/skills/write-next-chapter/scripts/test_candidate_cas_db.py
@ -6,7 +6,7 @@ PASSED 终态、身份不可变，以及条件 UPDATE 在真实并发语义下
 测试行用 unittest-cas- 前缀的 run_id，结束前物理清理（本表是可变注册表，无 append-only 约束）。

 跑法（需 Tailscale 内网可达 muse-example）：
-.venv/bin/python .claude/skills/write-next-chapter/scripts/test_candidate_cas_db.py
+.venv/bin/python tests/skills/write-next-chapter/test_candidate_cas_db.py
 """

 from __future__ import annotations
@ -15,8 +15,11 @@ import pathlib
 import sys
 import uuid

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
+from psycopg.errors import RaiseException
+
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+SCRIPT_DIR = SKILLS_DIR / "write-next-chapter" / "scripts"
 for path in (SCRIPT_DIR, SKILLS_DIR / "access-database" / "scripts"):
    if str(path) not in sys.path:
        sys.path.insert(0, str(path))
@ -107,7 +110,7 @@ def test_db_trigger_rejects_malformed_direct_updates() -> None:
            raise AssertionError(f"触发器必须拒绝: {sql}")
        except AssertionError:
            raise
-        except Exception:
+        except RaiseException:
            pass
    latest = store.latest(run_id)
    assert (latest.state, latest.revision) == ("DRAFT", 1), "非法 UPDATE 不得改变链"
@ -130,7 +133,7 @@ def test_start_next_monotonic_at_db_level() -> None:
        raise AssertionError("REJECTED 开新轮必须递增 attempt/candidate_version")
    except AssertionError:
        raise
-    except Exception:
+    except RaiseException:
        pass
    assert store.latest(run_id).state == "REJECTED", "非法开新轮不得改变链"

@ -146,7 +149,7 @@ def test_insert_must_start_draft_revision_one() -> None:
        raise AssertionError("初始行必须是 DRAFT")
    except AssertionError:
        raise
-    except Exception:
+    except RaiseException:
        pass
    try:
        with connect() as conn:
@ -157,7 +160,7 @@ def test_insert_must_start_draft_revision_one() -> None:
        raise AssertionError("初始 revision 必须是 1")
    except AssertionError:
        raise
-    except Exception:
+    except RaiseException:
        pass


--- a/.claude/skills/write-next-chapter/scripts/test_persist_writer_run.py
+++ b/.claude/skills/write-next-chapter/scripts/test_persist_writer_run.py
@ -5,7 +5,8 @@ import sys
 import unittest


-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "write-next-chapter" / "scripts"
 sys.path.insert(0, str(SCRIPT_DIR))

 import persist_writer_run as writer_persist  # noqa: E402
--- a/.claude/skills/write-next-chapter/scripts/test_run_writer.py
+++ b/.claude/skills/write-next-chapter/scripts/test_run_writer.py
@ -12,10 +12,13 @@ import sys
 import unittest
 from decimal import Decimal

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-READ_CONTEXT_DIR = SCRIPT_DIR.parents[1] / "assemble-context" / "scripts"
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "write-next-chapter" / "scripts"
+READ_CONTEXT_DIR = PROJECT_ROOT / ".claude" / "skills" / "assemble-context" / "scripts"
+TEST_CONTEXT_DIR = PROJECT_ROOT / "tests" / "skills" / "assemble-context"
 sys.path.insert(0, str(SCRIPT_DIR))
 sys.path.insert(0, str(READ_CONTEXT_DIR))
+sys.path.insert(0, str(TEST_CONTEXT_DIR))

 from test_writer_contract import valid_context, valid_draft  # noqa: E402
 from writer_contract import build_writer_creative_input, canonical_json, retrieval_identity  # noqa: E402
--- a/.claude/skills/write-next-chapter/scripts/test_run_writer_pipeline.py
+++ b/.claude/skills/write-next-chapter/scripts/test_run_writer_pipeline.py
@ -12,11 +12,14 @@ import tempfile
 import unittest
 from unittest import mock

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
-SKILLS_DIR = SCRIPT_DIR.parents[1]
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SKILLS_DIR = PROJECT_ROOT / ".claude" / "skills"
+TEST_DIR = PROJECT_ROOT / "tests" / "skills" / "write-next-chapter"
+CHECK_TEST_DIR = PROJECT_ROOT / "tests" / "skills" / "check-content-consistency"
+SCRIPT_DIR = SKILLS_DIR / "write-next-chapter" / "scripts"
 DETECT_DIR = SKILLS_DIR / "check-content-consistency" / "scripts"
 READ_CONTEXT_DIR = SKILLS_DIR / "assemble-context" / "scripts"
-for path in (SCRIPT_DIR, DETECT_DIR, READ_CONTEXT_DIR):
+for path in (READ_CONTEXT_DIR, DETECT_DIR, SCRIPT_DIR, TEST_DIR, CHECK_TEST_DIR):
    sys.path.insert(0, str(path))

 from test_check_writer_candidate import _candidate_body, _requirements, _valid_pair  # noqa: E402
--- a/.claude/skills/write-next-chapter/scripts/test_semantic_verdict.py
+++ b/.claude/skills/write-next-chapter/scripts/test_semantic_verdict.py
@ -4,7 +4,7 @@
 语义状态固化到候选行是接受通道 DB 兜底的依据，绑定/哈希校验必须失败关闭：
 报告版本、候选哈希、候选版本、上下文哈希、运行 ID 任一不一致都拒绝落库。

-跑法：.venv/bin/python .claude/skills/write-next-chapter/scripts/test_semantic_verdict.py
+跑法：.venv/bin/python tests/skills/write-next-chapter/test_semantic_verdict.py
 """

 from __future__ import annotations
@ -14,7 +14,8 @@ import pathlib
 import sys
 import unittest

-SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
+SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "write-next-chapter" / "scripts"
 if str(SCRIPT_DIR) not in sys.path:
    sys.path.insert(0, str(SCRIPT_DIR))