docs(pr-review): align verification wording with installer smoke

[LIHUA919]

install-local-smoke.py requires the installed PR-review skill to say never infer verified from metadata or CI, while the skill on main says never verified from metadata or CI. The exact-text assertion fails even though the guidance has the same intent. Clarify that sentence so the shipped skill and installer smoke agree; review execution behavior is unchanged.

Validation on signed head c2fde1555: the full examples/install-local-smoke.py passed with Node 26; risk-based premerge passed 3 direct checks and 2 selected canaries; loopx check found 0 errors and 0 warnings. A DCO-signed merge incorporates main at 07eb7a1, and the PR diff against that base remains the one-sentence skill change. This is a separate upstream repair for the premerge failure observed while qualifying #5280.

Historical CI at head e0fbbe8 failed Stage2c e2e 2 in 14 parameterizations because a subprocess loaded the installed wheel rather than the checked-out source (test_cold_source_disposition_e2e.py:121). This test assertion was independent of the skill edit. PR #6040 preserves the installed-wheel original-producer path while correcting it; its exact head ac2af0f passed stage2c (e2e 2) and merged into main as 07eb7a1. The overlapping PR #6063 was closed in favor of #6040. New head c2fde1555 includes that merge; its fresh CI is pending, so the prior red run is not counted as current validation.

The old head's test-shard (2) job also failed four Todo/Lark cases. Its JUnit artifact has the same four test names and failure messages as closed PR #6063's shard 2 artifact. The separate PR #6068 fixes the shard 2 Todo-display KeyError: 'todo_id' assertion; its old CI run still shows the now-merged Stage2c baseline failure. The old test-shard (4) failed six Goal-acceptance/Lark cases; its JUnit test names and failure messages likewise match #6063's shard 4 artifact. A separate PR #6075 addresses the shared Lark private-return fixture; its required CI is pending. This PR changes only skills/loopx-pr-review/SKILL.md, which those tests do not read, so those failures are not attributed to the skill sentence. PR #6062 remains unmerged; the current head's required checks and exact-head review still need readback.

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=xhigh

动机

安装或升级 LoopX 的维护者需要安装验证准确读回已交付的评审指令。 旧版本已安装的 skill 缺少 smoke 所要求的 infer 一词，完整安装验证在这句文本断言处失败；新版本澄清该句，让验证继续完成，同时保留证据不足必须标为 unverified 的要求。 独立的旧版安装复现了该失败，新版完整安装 smoke 通过；两版实际安装后的文件均与各自源码一致，重复安装无改写。 本 PR 只修复 skill 与安装验证的措辞一致性，不改变评审资格、证据门禁、CI 策略、合并授权或其他能力默认，也不宣称模型准确率提高。

改动思路

复用现有 canonical skill，在同一句增加明确的“infer”并调整标点，比放宽安装 smoke 的独立断言更合适。不修改能力的执行合同、状态检查或安装器；安装器仍只负责把当前源码交付给 host，证据判断仍属于既有 review capability。

具体改动

精确 head e0fbbe8ec19bb96354b7468979589bde3ed0afd4，完整 base→head 只有 skills/loopx-pr-review/SKILL.md:72 的 +1/-1。前后均要求填当前 review_plan.result_template，缺失证据保持 unverified，不得由 metadata 或 CI 判为 verified；新的句子把这个禁止推断明确说出来。其他全部 skill 字节、host metadata、生产代码和测试均未改动。

按改动前的 skills/loopx-pr-review/SKILL.md、不可变版本 2643d4ff79acfb5ae8730b11c329d7c6287e3262 核验：Execute One Review Plan 的证据来源、顺序及完成检查保持；Publish And Read Back 的精确 head 发布/读回及批准与合并权限分离保持。实际安装路径是既有 scripts/install-local.sh 的 skill 物化；examples/install-local-smoke.py:453/458 独立检查安装结果。另经公开 loopx workflow-skills --install --skills-dir 安装、单独读取 manifest/文件、重复安装，两版文件均等于各自源码，重复结果均为 unchanged。

对主干的风险

最强风险是为迎合字符串断言而删掉证据义务，或只验证源码却实际安装别的版本。本轮对完整指令逐条比较，并核验真实安装后的字节和源码来源，未发现该漂移。

相同原始完整 smoke 在 baseline 返回1，确实于 line458 抛出 never infer verified 文本断言；head 返回0，initial/fresh/stale 三次安装及后续全部检查完成。当前230项 review 合同/result-check 测试通过，含缺失证据及丢失指令语义不能批准的反例。验证环境为 CPython3.14.8 / Node24.21.0；source check 对改动 skill 为0 errors/0 warnings，diff-check干净，语义 advisory无新支持的词汇 carrier；advisory不证明所有动态语义。

初始附加探针误用仓库不支持的 python -m loopx 入口，在写入前失败；保留该 harness 失败后改用已交付的 uv run --extra test loopx，实际安装、独立读回和重复路径均通过。没有改断言或产品规则来取绿。此次没有 opt-in/default-off 声明、新协议/状态、默认切换或权限扩大；CI按当前合同未查询。确定的安装/指令等价证据不等于模型遵循度或准确率已经提升，也不等于本机升级或跨平台 release 资格。

我的整体评价

无阻塞发现，goal_achieved 指本句澄清和安装验证恢复；长期行为与用户路径 preserved。未来重构检查考察了 literal smoke 的维护成本：本次保留原独立保护并修正原句已经足够，无需新增抽象或重写验证框架。交付完整、可逆、现有 owner 清楚，可批准此精确版本；合并与安装采用分别按既有维护者流程处理。

English verdict: APPROVE - head e0fbbe8. This one-sentence clarification preserves the existing missing-evidence and metadata/CI prohibition. The unchanged full installer smoke reproduces the baseline phrase failure and passes at the head;230 contract/result checks and real workflow installation, separate readback and unchanged repeat installation pass. No new activation, authority, runtime state or model-quality claim is introduced. Merge and machine adoption remain separate.