fix(site): align LHTB publication contract

[huangruiteng]

Summary

• distinguish the five-arm LHTB dataset/selector contract from the three-arm main comparison narrative
• simplify the English and Chinese core-result headings while keeping the 7/7/4 strict-solve evidence in the body
• lock both localized headings and body evidence into the publication-scope regression test

Validation

• baseline origin/main: 1 failed, 3 passed in benchmark/tests/test_publication_scope.py
• fixed head: 4 passed
• Ruff: passed
• public-site TypeScript/Vite build: passed
• public presentation design baseline: passed
• bilingual Blog index smoke: passed (3 paired articles)
• frontstage share-bundle smoke: passed
• desktop/mobile browser navigation matrix: passed
• public/private boundary scan: clean
• LoopX premerge canary: 6/6 selected checks passed; its benchmark-sensitive manual-review hold remains and is satisfied only through the explicit maintainer review and authorization recorded for this PR

Scope and risk

This changes only public LHTB copy and its focused publication-scope test. It does not change benchmark data, scoring, task semantics, runners, permissions, or control-plane behavior. The first-screen English and Chinese previews were reviewed and approved before commit.

[huangruiteng]

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

Reviewed exact head: 2ac8a4224e285e0b9ef73148d4722f387cf0730a

No blocking findings.

动机

本 PR 修复 #4709 合入后已复现的发布范围红测，并落实经首屏预览确认的文案层级：核心标题只突出 LoopX 相对 Plain 与原生 Goal 的均分优势，完整通过数 7/7/4 继续保留在正文。基线 main 的 focused pytest 为 1 failed, 3 passed；本 head 为 4 passed，两个请求均完整交付。

改动思路

改动直接落在现有两个 owner，没有新增抽象或第二份数据源。data.json 与 armLabels / armKinds 继续表达五臂实验全集；mechanismRows 明确表达 plain、native_goal、new_heartbeat 三个主比较臂。lhtb-copy.json 继续同时承载标题和正文，标题负责主结论，正文保留均分、7/7/4 与完整验收边界。

正向路径是：选择中英文 locale → 读取 lhtb-copy.json → 渲染新标题和原正文 → focused pytest、Vite build、share-bundle 与浏览器矩阵验证。负向路径是：任一 locale 的标题、正文证据、五臂全集或三臂主叙事漂移时，focused pytest 直接失败。

具体改动

• apps/presentation/site/src/lhtb-copy.json：英文标题改为 “Higher mean reward than Plain and native Goal.”，中文标题改为“均分高于 Plain 和原生 Goal。”；正文的 7/7/4 与验收边界原样保留。
• benchmark/tests/test_publication_scope.py：把原先过载的 arms 拆成 all_arms 与 main_comparison_arms，分别校验五臂数据/标签和三臂主叙事；新增中英文标题精确断言与正文 7/7/4 证据断言。

关键代码讲解

1. en.resultTitle / zh.resultTitle 与相邻 resultBody：只调整信息层级，不改变任何实验数值；真实中英文首屏已在提交前由维护者确认。
2. test_lhtb_published_data_and_bilingual_copy_share_scope：用两个显式闭集表达两个不同合同，避免再次用五臂数据范围覆盖三臂主叙事；同时保证标题精简不会删除正文证据。

对主干的风险

风险较低且局限于 LHTB 公共页面和 focused test。没有修改 benchmark 数据、评分、task semantics、runner、权限、持久化或控制面行为。最强回归场景是未来编辑把历史臂重新误当主叙事，或精简标题时删掉正文证据；新增断言会直接拦截。

验证：

• baseline focused pytest：1 failed, 3 passed（成功复现已知问题）
• exact-head focused pytest：4 passed
• Ruff、git diff --check、DCO：通过
• public-site TypeScript/Vite build、design baseline：通过
• bilingual Blog index smoke：通过，3 组中英文章
• frontstage share-bundle smoke：通过
• desktop/mobile browser navigation matrix：通过
• public/private boundary scan：通过
• premerge canary：6/6 selected checks 通过；仅保留 benchmark-sensitive 的人工维护者复核标记。本 PR 不改 benchmark 执行或评分，且用户已明确授权本次自合并。

远端 CI 在审查时仍运行中；按维护者明确指示，本次不等待它结束，未将其表述为已通过。

语义与 CI 对齐

该 head 复用当前五臂实验 vocabulary，同时明确三臂展示子集，没有创建新协议或改变既有评分语义。基线失败与修复后通过构成直接的前后对照；标题和正文的中英文语义也由真实页面截图验证。

我的整体评价

这是一个单目的、可回滚的小型公共文案/测试修复，直接修复主干红测并改善首屏信息层级。代码体量与问题成比例，测试覆盖了正向结果与关键负向漂移，未发现阻断项。未来相关 refactor 无需在本 PR 扩张；当前 owner 与显式集合已经足够清晰。

English verdict: APPROVE - Exact head 2ac8a4224e285e0b9ef73148d4722f387cf0730a fixes the main regression by separating five-arm source coverage from the three-arm primary narrative, preserves the 7/7/4 evidence below the approved bilingual headings, and passes focused, build, bundle, browser, boundary, and 6/6 selected canary validation; remote CI was explicitly not awaited.