ShanireZ · ShanireZ · Aug 27, 2026 · Aug 26, 2026 · Aug 26, 2026 · Aug 26, 2026
diff --git a/deploy.md b/deploy.md
@@ -252,7 +252,7 @@ wrangler tail dpmaster
 
 ## Sitemap 与搜索平台提交
 
-构建把 sitemap 写进发布目录，线上地址固定为 `https://dp.round1.cc/sitemap.xml`，列出 47 个页面，每条带 Git 派生的 `lastmod`。单域站点不输出 hreflang 互指，`xmlns:xhtml` 命名空间也不再出现。
+构建把 sitemap 写进发布目录，线上地址固定为 `https://dp.round1.cc/sitemap.xml`，列出 47 个页面，每条带源码证据派生的 `lastmod`：已提交源码取 Git 提交时间，工作树编辑取文件修改时间，绝不使用模板或构建日期。单域站点不输出 hreflang 互指，`xmlns:xhtml` 命名空间也不再出现。
 
 ★ 域名迁移后必须重新提交，旧域名的验证与 sitemap 记录不会自动继承：
 

diff --git a/docs/engineering/architecture.md b/docs/engineering/architecture.md
@@ -69,8 +69,9 @@ The manifest intentionally keeps only dependencies imported by current source. D
 | `site/src/components/dp-engine/` | Shared visualization, playback, controls, and safe-caption rendering. |
 | `site/src/components/games/` | One game per family; games consume public result Interfaces instead of teaching frames. |
 | `site/src/components/games/runtime/` | Shared deterministic random source, round statistics, lazy audio, and viewport gate for the seven games. |
-| `site/src/config/site.ts` | DP大师 brand, copyright holders, and the single production host with its same-origin API endpoints. |
+| `site/src/config/site.ts` | Public representation identity only: DP大师 brand, canonical origin, and language. Client endpoints and sidebar-only copy live in separate modules that the semantic route graph cannot reach. |
 | `site/src/lib/publicRoutes.ts` | Catalog-derived authority for 47 public routes plus one separately declared internal prerender specimen. |
+| `site/src/lib/publicWebContract.ts` | Shared internal representation paths and public `Content-Signal` contract consumed by generation, discovery, and the Worker. |
 | `site/src/lib/pageMeta.ts` | Pure route metadata authority for titles, summaries, canonicals, breadcrumbs, and indexing policy. |
 | `site/src/lib/seoHead.ts` | Server/client shared head and JSON-LD generation. |
 | `site/src/lib/discovery.ts` | Sitemap, robots (including the AI-crawler group), llms.txt, and route-summary generation. |
@@ -79,7 +80,7 @@ The manifest intentionally keeps only dependencies imported by current source. D
 | `site/src/entry-server.tsx` | React 19 static rendering entry used by the prerender build. |
 | `site/scripts/build.mjs` | Builds the client plus an isolated production SSR bundle, then prerenders. |
 | `site/scripts/prerender.mjs` | Writes route HTML, one Markdown representation per public route, 404.html, and discovery files. |
-| `site/scripts/markdown-representation.mjs` | DOM-based projection from locally prerendered semantic HTML to compact Markdown; it is not a second hand-written content source. |
+| `site/scripts/markdown-representation.mjs` | Deterministic DOM-based projection from locally prerendered semantic HTML to compact Markdown with a normalized legal heading outline; it is not a second hand-written content source. |
 | `site/worker/` | Shared feedback, privacy-bounded first-party analytics, and egress-probe cores. |
 | `site/worker.js` + `site/worker/` | The single deployment adapter: public-page content negotiation plus same-origin feedback, analytics, and egress diagnostics. |
 
@@ -99,15 +100,15 @@ All 29 teaching solver surfaces are Adapters over the algorithm boundary: public
 
 Known deep links are physical prerendered HTML assets. Cloudflare Workers Static Assets serves them and uses `404-page` for anything unmatched, returning the themed `404.html` with a real HTTP 404. See root [deploy.md](../../deploy.md) for the platform contract.
 
-For a registered public `GET` or `HEAD`, `worker/content-negotiation.js` selects HTML or Markdown from `Accept`; HTML wins equal preferences and the Worker returns 406 when both supported media types are explicitly unacceptable. Markdown is fetched only through a distinct internal asset URL, so its ETag and static-asset cache key cannot collide with HTML. Both representations add `Vary: Accept`, alternate links, `nosniff`, and the approved public `Content-Signal`; direct requests to the internal prefix return 404. APIs, ordinary assets, the internal specimen, and unknown routes bypass negotiation.
+For a registered public `GET` or `HEAD`, `worker/content-negotiation.js` selects HTML or Markdown from `Accept`; HTML wins equal preferences and the Worker returns 406 when both supported media types are explicitly unacceptable. Markdown is fetched only through a distinct internal asset URL, so its ETag and static-asset cache key cannot collide with HTML. Both representations add `Vary: Accept`, alternate links, `nosniff`, and the approved public `Content-Signal`; direct requests to the internal prefix return 404. `HEAD` executes the same matrix and headers but never returns a body. APIs, ordinary assets, the internal specimen, and unknown routes bypass negotiation. The internal path mapper and public signal live in `publicWebContract.ts`, preventing the generator, discovery files, and Worker from silently diverging.
 
 # SEO And Accessibility
 
 `pageMeta.ts` is the shared route metadata authority. The site has a single origin, so every public page canonicalizes to itself and publishes no language alternates at all. Unknown routes are non-indexable and deliberately have no canonical link.
 
 `seoHead.ts` renders the same head contract at build time that `RouteMeta` maintains after client navigation: title, description, citable abstract, robots, canonical, Open Graph, Twitter, and a Schema.org graph. Every indexable page includes Organization and WebSite nodes and states `isAccessibleForFree`; the WebSite node declares a `SearchAction` pointing at the real `?q=` problem search. Lessons add Course/LearningResource/TechArticle with `educationalUse` and an `EducationalAudience`, family pages add CollectionPage plus an ItemList of their ready lessons, other pages add WebPage, and routes with hierarchy add BreadcrumbList. Course duration is deliberately not declared: there is no trustworthy source for it, so `hasCourseInstance` is omitted rather than fabricated.
 
-`site/scripts/generate-seo.mjs` writes the checked-in baseline for `sitemap.xml`, `robots.txt`, `llms.txt`, `route-summaries.json`, and the home-route head inside `index.html`; the prerender step regenerates the four discovery files into the build output. All of them derive from the same 47-path catalog contract: home, seven families, 37 completed lessons, method, and problem index. `llms.txt` is grouped by family and declares the same-URL Markdown interface and approved public content signal. `robots.txt` explicitly allows both retrieval and training crawlers; a release is incomplete until Cloudflare AI Crawl Control is confirmed not to inject a conflicting managed no-train block. `pnpm check:seo` rejects generated-artifact drift, including a hand-edited `index.html` head.
+`site/scripts/generate-seo.mjs` writes the checked-in baseline for `sitemap.xml`, `robots.txt`, `llms.txt`, `route-summaries.json`, and the home-route head inside `index.html`; the prerender step regenerates the four discovery files into the build output. All of them derive from the same 47-path catalog contract: home, seven families, 37 completed lessons, method, and problem index. Sitemap entries are sorted by canonical URL. `semantic-source-graph.mjs` follows the real SSR injection chain from `entry-server.tsx` through the static application/providers and route tree, together with metadata roots and the exact route page, lesson, and family-art modules. Mixed shell modules use an AST-rooted projection: the route-rendering JSX plus the closure of local declarations and render-phase assignments reached by lexical binding identity are hashed. Control predicates, loop/switch scopes, shadowed names, class helpers, route/outlet bindings, and class-field/member identity remain precise; effect-only callbacks and JSX attribute names do not leak in. A nested write outside the render or declaration execution scope is accepted only for a recognized client-effect callback; any other ambiguous nested write fails generation and must be refactored instead of silently producing incomplete evidence. The graph deliberately prunes surrounding navigation/feedback, unrelated route pages, client-only analytics/theme/motion controllers, and the generated evidence module. Public representation identity is structurally isolated in `config/site.ts`; client endpoints and sidebar-only copy are in modules reachable only through pruned branches. Each route records a versioned, line-ending-neutral SHA-256 digest of that graph. Clean sources use the latest Git commit timestamp, dirty sources use the newest source-file modification timestamp, an unchanged digest preserves the previous value, and generation fails when changed content lacks strictly advancing source evidence; changing the digest schema preserves the prior content timestamp only when no newer source evidence exists, while a simultaneous real content update adopts its later timestamp. Version 4 was reseeded from the trusted `2026-08-22` history: the home route uses the visible hero change at `23:46:11+08:00`, and the other 46 routes use the single-origin metadata/content change at `11:47:40+08:00`. Earlier rejected schemas had admitted a shared client effect, sidebar-only copyright copy, or incomplete execution/scope context. `llms.txt` is grouped by family and declares the same-URL Markdown interface and approved public content signal. `robots.txt` explicitly allows both retrieval and training crawlers; a release is incomplete until Cloudflare AI Crawl Control is confirmed not to inject a conflicting managed no-train block. `pnpm check:seo` rejects generated-artifact drift, including a hand-edited `index.html` head.
 
 # Analytics
 

diff --git a/docs/operations/verification.md b/docs/operations/verification.md
@@ -44,7 +44,7 @@ Pure TypeScript Modules imported by Node tests must use explicit `.ts` extension
 
 # SEO And Accessibility Checks
 
-`pnpm check:seo` verifies the 47-path catalog, route metadata, checked-in single-origin sitemap/robots baseline, `llms.txt`, 47-entry route summaries, and Git-derived `lastmod` data. It also checks the source contracts for the one Cloudflare output, React prerender/hydration, route CSS injection, canonical metadata, and real 404 behavior. `pnpm check:html` then requires the sitemap/public-route/Markdown sets to agree, rejects missing or extra Markdown assets, browser-only markup and internal-path leakage, and requires each Markdown representation to remain less than half the size of its corresponding HTML. When routes, lesson readiness, or route-owned source files change, run `pnpm seo:generate` and review the generated files before committing them.
+`pnpm check:seo` verifies the 47-path catalog, route metadata, checked-in single-origin sitemap/robots baseline, `llms.txt`, 47-entry route summaries, and source-backed `lastmod` data. A line-ending-neutral SHA-256 digest follows each route's real SSR injection chain and exact semantic content graph while pruning navigation, feedback, analytics, theme, and motion-only branches. Mixed shell modules contribute an AST-rooted closure of route-rendering JSX, local declarations, and render-phase assignments resolved by lexical binding identity. Regression probes cover control predicates, function/loop/switch shadowing, class fields, later assignments, route/outlet bindings, effect-only writes, and non-reference JSX names; ambiguous nested writes outside known client-effect callbacks fail closed. Client endpoints and sidebar-only copy are structurally isolated from the public representation configuration and are unreachable from all 47 semantic graphs. Clean sources use their latest Git commit timestamp; dirty sources use their newest filesystem modification timestamp; and generation fails if the digest changes without that evidence advancing. A digest-schema migration preserves the prior timestamp only when it has no newer source evidence; a simultaneous content edit must still advance. Version 4 was explicitly reseeded from trusted `2026-08-22` Git history: `23:46:11+08:00` for the home hero change and `11:47:40+08:00` for the other 46 routes' latest single-origin metadata/content change. Earlier rejected schemas admitted a shared client effect, sidebar-only copyright copy, or incomplete execution/scope context. This keeps same-day edits precise while unchanged or cross-platform regeneration remains deterministic. The gate also checks the source contracts for the one Cloudflare output, React prerender/hydration, route CSS injection, canonical metadata, and real 404 behavior. `pnpm check:html` then requires the sitemap/public-route/Markdown sets to agree; rejects missing or extra Markdown assets, browser-only markup, internal-path leakage, illegal heading hierarchies, and unresolved same-origin links; regenerates every Markdown document twice from its matching semantic HTML to prove exact same-source parity and determinism; and requires each representation to remain less than half the size of its HTML. When routes, lesson readiness, or route-owned source files change, run `pnpm seo:generate` and review the generated files before committing them.
 
 The browser gate runs Chromium against built `dist/` through a strict, non-reused custom preview server. Route tests directly open `/`, `/part/a`, `/part/a/01`, `/method`, and `/part/g/plug`, then exercise live client navigation and keyboard focus. They check HTTP status, prerendered HTML, absence of pending streamed-Suspense placeholders, route CSS at first paint, hydration without errors, CLS below `0.05`, title/description/abstract/canonical/zero-hreflang/Open Graph/JSON-LD metadata, one visible `h1`, route announcements, current-page semantics, initial-load focus, changed-route focus, and skip-link focus. Navigation-positioning tests click a representative lesson outline in each of the seven families, restore a direct lesson fragment, verify the mobile outline and an ordinary family-page fragment, and require a changed route to reset old-page scroll immediately. An unknown path must return 404 with `noindex,nofollow`, no canonical, and the themed not-found heading. Worker-level Node tests separately cover the Accept matrix, 406 behavior, HEAD parity, representation-specific ETags, hidden internal assets, and unaffected API/static/unknown paths.
 
@@ -95,8 +95,11 @@ When changing lesson content:
 
 Deployment checks live in root [deploy.md](../../deploy.md). At minimum after deployment:
 
-* Open a normal route and a direct deep link on both domains.
-* Confirm each domain self-canonicalizes and links to both regional variants.
+* Open a normal route and a direct deep link on `https://dp.round1.cc`; confirm both self-canonicalize and emit zero hreflang links.
+* Run GET and HEAD against a public route with `Accept: text/html` and `Accept: text/markdown`, then request an unacceptable media type and confirm 406.
+* Confirm both successful representations carry distinct ETag values, `Vary: Accept`, the appropriate alternate/canonical `Link`, and `Content-Signal: ai-train=yes, search=yes, ai-input=yes`.
+* Alternate HTML, Markdown, and HTML requests through the final edge to prove the cache never crosses representations; direct requests under `/_representations/` must return 404.
+* Inspect the final `robots.txt` and confirm Cloudflare has not injected `ai-train=no` or crawler `Disallow` rules that conflict with the repository policy.
 * Request a made-up path and confirm HTTP 404 plus `noindex,nofollow`.
 * Submit a feedback test.
 * Confirm `POST /api/analytics` returns 204 for a valid same-origin event and inspect the platform log.

diff --git a/site/index.html b/site/index.html
@@ -51,7 +51,7 @@
     <meta name="twitter:title" content="DP大师 · DP Master" />
     <meta name="twitter:description" content="DP大师面向算法学习者，用精讲、逐帧可视化、题目索引和小游戏讲清状态定义、转移顺序与模型迁移。" />
     <meta name="twitter:image" content="https://dp.round1.cc/og/dpmaster-social.jpg" />
-    <script id="dp-structured-data" type="application/ld+json">{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://dp.round1.cc/#publisher","name":"AzureL蔚澜算法","url":"https://dp.round1.cc","logo":"https://dp.round1.cc/favicon.svg","brand":{"@type":"Brand","name":"DP大师"}},{"@type":"WebSite","@id":"https://dp.round1.cc/#website","name":"DP大师","alternateName":"DP Master","url":"https://dp.round1.cc/","description":"DP大师用精讲、逐帧可视化、题目索引和小游戏讲清动态规划。","inLanguage":"zh-Hans","publisher":{"@id":"https://dp.round1.cc/#publisher"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://dp.round1.cc/problems?q={search_term_string}"},"query-input":"required name=search_term_string"}},{"@type":"WebPage","@id":"https://dp.round1.cc/#webpage","url":"https://dp.round1.cc/","name":"DP大师 · DP Master","description":"DP大师面向算法学习者，用精讲、逐帧可视化、题目索引和小游戏讲清状态定义、转移顺序与模型迁移。","abstract":"DP大师面向算法学习者，用精讲、逐帧可视化、题目索引和小游戏讲清状态定义、转移顺序与模型迁移。","inLanguage":"zh-Hans","isPartOf":{"@id":"https://dp.round1.cc/#website"},"publisher":{"@id":"https://dp.round1.cc/#publisher"},"isAccessibleForFree":true,"dateModified":"2026-08-22","image":"https://dp.round1.cc/og/dpmaster-social.jpg"}]}</script>
+    <script id="dp-structured-data" type="application/ld+json">{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://dp.round1.cc/#publisher","name":"AzureL蔚澜算法","url":"https://dp.round1.cc","logo":"https://dp.round1.cc/favicon.svg","brand":{"@type":"Brand","name":"DP大师"}},{"@type":"WebSite","@id":"https://dp.round1.cc/#website","name":"DP大师","alternateName":"DP Master","url":"https://dp.round1.cc/","description":"DP大师用精讲、逐帧可视化、题目索引和小游戏讲清动态规划。","inLanguage":"zh-Hans","publisher":{"@id":"https://dp.round1.cc/#publisher"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://dp.round1.cc/problems?q={search_term_string}"},"query-input":"required name=search_term_string"}},{"@type":"WebPage","@id":"https://dp.round1.cc/#webpage","url":"https://dp.round1.cc/","name":"DP大师 · DP Master","description":"DP大师面向算法学习者，用精讲、逐帧可视化、题目索引和小游戏讲清状态定义、转移顺序与模型迁移。","abstract":"DP大师面向算法学习者，用精讲、逐帧可视化、题目索引和小游戏讲清状态定义、转移顺序与模型迁移。","inLanguage":"zh-Hans","isPartOf":{"@id":"https://dp.round1.cc/#website"},"publisher":{"@id":"https://dp.round1.cc/#publisher"},"isAccessibleForFree":true,"dateModified":"2026-08-22T23:46:11+08:00","image":"https://dp.round1.cc/og/dpmaster-social.jpg"}]}</script>
 <!-- dp-route-head:end -->
   </head>
   <body>