Maintain documentation discovery and local search

The website offers Start, Learn, How-to, Reference, Understand and Contribute entry points plus a task catalogue. The catalogue uses the existing website/content/coverage.json rows and the explicit integration tasks in website/plugins/discovery.mjs. Both resolve only published pages. The original 27-topic review contract is not a limit on discoverable tasks. Catalogue entries do not create new guide acceptance claims: a reference remains labelled as a reference until its owner supplies a reviewed walkthrough.

Search ownership and scope

LSF maintains the small Docusaurus integration in website/src/theme/SearchBar/, website/src/pages/search.tsx and website/lib/search-build.mjs. The reviewed engine is MiniSearch 7.2.0, a browser/Node JavaScript library with no runtime dependencies. The website lock pins its registry integrity. This is an LSF-owned local integration, not an official Docusaurus search service. The website maintainer owns updates and advisory review; consult the upstream Docusaurus search options when changing it.

The production build indexes the rendered Markdown/MDX content of the explicit public documentation and decision inventory. It retains headings, section anchors and all published language-tab text. Navigation, scripts, styles and pages with a noindex robots directive are excluded. Hidden source files, unreferenced snippet regions, Wiki trees, component review pages and test-only build directories are never added to the corpus. An archived page can remain readable while opting out of search using Docusaurus noIndex: true front matter.

The index is limited to 6,000 records and 16 MiB. This byte budget accommodates the current guides and three retained release snapshots, including alpha.5, without silently omitting older searchable guides. Individual rendered sections are limited to 256 KiB and split into 6,000-character search records without dropping their remaining text. Exceeding these limits fails the build and requires review. The maintained build check recomputes the expected index from the actual HTML and publication identity, rejecting altered or stale content. Browser loading also checks the revision, base path, channel inventory and byte limit.

Queries execute in the reader's browser. Only the static same-origin index is downloaded; there is no search backend, third-party query collection, account or secret. Query URLs use search/?q=publication&version=0.1.0-alpha.3. Results are restricted to that selected channel and labelled with its version and profile. Unknown versions and empty results have explicit recovery states. The navbar preserves the selected documentation channel; optional local storage failure does not prevent a version-labelled URL from working.

Add or revise a discoverable guide

1. Add current prose under docs/, following the existing coverage contract and authoring ownership. Use registered source examples instead of copied SDK programs.
2. Update the relevant coverage row's page path and role, or register a new integration task in the discovery plugin without rewriting historical review receipts. Missing pages and duplicate task IDs fail the build. Audience, SDK language and topic filters only expose values backed by existing catalogue entries. Catalogue filters have stable audience, language and topic query fields.
3. Keep each heading meaningful: search results link directly to its rendered anchor. Preserve historical snapshot bytes; follow the version procedure for attributable corrections.
4. Extend the existing journey in website/scripts/test-discovery.mjs if the new guide changes a reader task. Reuse the ordinary website workflow and compact receipts; do not add a separate benchmark or issue-numbered workflow.

Verification

Run website type/unit checks, both production builds, test:build, test:examples, test:versions and test:discovery after committing the source. The discovery journey uses the pinned Playwright browser and @axe-core/playwright 4.13.0. It fails on serious/critical WCAG A/AA findings, incorrect-version results, broken required targets, external browser requests, hydration errors and failed navigation. It covers landing-to-guide, query URLs, released search, no results, missing versions, filters, previous/next links, direct reload, light/dark, keyboard, mobile and reduced motion. The theme suite and manual review supplement automated accessibility checks; a 640 CSS-pixel reflow check alone is not a claim of native browser zoom verification.

Review the actual production pages manually for sensible task wording, reading order, visible keyboard focus, native 200% browser zoom and diagram readability. Retain only screenshots that show a useful review state. Compact automated receipts go to website/.generated/discovery-review/. These checks qualify site behavior; runtime and SDK execution evidence retain their original owners.