` content while excluding navigation and interactive controls. Documentation representations start from maintained Markdown, expand controlled in-tree includes, retain headings, prose, lists, tables, code, and supported containers, normalize supported raw HTML, remove links to the private app, and resolve public links and assets against the canonical page. Browser-tool source marks the canonical explanation with `data-agent-include` and the unusable interactive region with `data-agent-exclude`; other pages may use the exclusion annotation for illustrative or duplicate responsive content. Unknown meaning-bearing markup, unresolved or unsafe includes, metadata drift, duplicate outputs, unknown public page targets, invalid rendered fragments, private routes, and representations over 256 KiB without a reviewed catalogue exception fail the build. External targets remain the source owner's responsibility. Both build outputs include a Cloudflare Pages `_headers` contract. Explicit `.md` files declare `text/markdown; charset=utf-8`; discovery and corpus text files declare `text/plain; charset=utf-8`. The build adds `Vary: Accept` to every catalogue-owned canonical HTML path and explicit Markdown artifact so an enabled Cache Rule can separate the two representations. These origin headers do not enable canonical-URL content selection. Optional edge selection remains a separate operator action. The documentation build also publishes `llms-full.txt`, an OpenPost convenience artifact rather than part of the llms.txt v2 proposal. Catalogue metadata selects each page, records exclusions with reasons, and assigns included pages to stable sections. The corpus preserves one source link per page while removing repeated representation metadata and generated warnings. It excludes OpenAPI bodies, legal notices, generated CLI reference repetition, and every explicit catalogue exclusion. The build warns when the corpus exceeds 1 MiB and fails when it exceeds 2 MiB. Each site invokes only its own projection, and generated files stay in ignored build output. The production-artifact contract enumerates every manifest and catalogue route. It requires one canonical HTML file, one explicit Markdown file, useful prerendered content, identity and discovery metadata, absolute links, bounded output, no stale aliases, and no exposed framework state. It also proves canonical plan values, provider facts, legal text, browser-tool explanations, HTML-only sitemaps, curated indexes, corpus policy, and repeatable generation. Shared generator and catalogue changes plan both public builds. Ordinary documentation prose plans only documentation, and each public build cache includes the shared generator, catalogue dependency, and complete output directory. Run `bun run check -- public-routes` for the production-artifact contract and `bun run check -- release-version` for selective CI planning. Then use `bun run build -- marketing` and `bun run build -- docs` to inspect the exact production artifacts. This content covers public product and operating information only. Do not add authenticated application state, Workspace data, OpenAPI conversions, or MCP protocol conversions to these outputs. ## Verification order 1. Update behavior, generated contracts, tests, and public docs together. 2. Run `bun run doctor` before broad or release work. 3. Run scoped root checks while editing, then `bun run verify`. 4. For visible changes, run the related app, docs, or marketing browser suite at desktop and phone widths. 5. Before a public campaign, complete the [Launch Verification Matrix](https://docs.openpo.st/operations/provider-launch-matrix) for each account and format. 6. For a production release, follow [Releases and Versioning](https://docs.openpo.st/development/releases) and verify the workflow, release, deployed revision, and public readiness.