Add accessibility-expert agentic workflow

[Copilot]

Adds a scheduled agentic workflow that audits this repository's only web interface — the Astro docs site — for WCAG 2.2 A/AA accessibility barriers and files a single superseding issue with findings.

Workflow — .github/workflows/accessibility-expert.md

• engine: pi, weekly schedule + workflow_dispatch (optional pages input), read-only permissions, timeout-minutes: 30, max-ai-credits: 400.
• Playwright CLI (mode: cli) with a no-sandbox chromium config; network.allowed scoped to node/chrome/playwright/local.
• safe-outputs limited to create-issue ([accessibility] prefix, close-older-key: accessibility-expert, max: 1), so repeat runs replace rather than accumulate reports.

Repository tailoring

• Pre-agent steps build the docs site and vendor axe-core@4.13.0 via npm pack into .a11y/axe.min.js, so the agent fetches nothing at runtime behind the egress firewall.
• The prompt derives the preview command and base path from package.json and astro.config.mjs instead of hard-coding a port — the docs site sets a non-root base.
• Beyond axe-core, the agent covers checks automation can't: keyboard traversal and focus traps, focus visibility, heading/landmark structure, accessible names, contrast, 320px reflow, and prefers-reduced-motion; each page is visited in both light and dark schemes. Infrastructure blockers are reported separately from accessibility findings so a build failure doesn't masquerade as a clean audit.

Supporting changes

• .github/workflows/accessibility-expert.lock.yml committed, plus the matching .gitignore negation required for workflows that actually run in this repo.
• tests/unit/workflow-contract.test.mjs: added the lock to the clean-room expected list and a contract test pinning trigger, engine, Playwright CLI mode, axe-core pinning, WCAG scope, light/dark coverage, and the absence of any write safe output other than create-issue.

Note for reviewers

Frontmatter declares issues: read. That is deliberate — gh-aw grants issues: write to the generated safe_outputs job (lock lines 1101, 1915) while the agent job stays read-only.