v0.3.0

BREAKING: Playwright is no longer installed by default. The default
install (pip install a11y-moda) now ships only lint + rules
subcommand dependencies (~30MB). To use scan / site / --render /
--render-crawl / --probe-modals, install the [scan] extra:

pip install 'a11y-moda[scan]'
playwright install chromium

Existing users who upgrade and then run --render will see a clear
install message printed to stderr; the CLI does not auto-install.

Reframing — knowledge service

a11y-moda repositions from "audit CLI" to "audit CLI + queryable MODA
rule knowledge service". The new rules subcommand exposes the
internal rule registry as a queryable API for AI agents (any IDE /
agent that can call CLI tools), so the agent can lookup MODA rules
before writing accessibility-sensitive code instead of only after
lint complains. Cross-platform by design — works with Cursor, GitHub
Copilot, Aider, Claude Code, Cline, Continue, custom agents.

Added

• a11y-moda rules subcommand group — query MODA rule metadata.
  Three subcommands: list, show, search. JSON / Markdown output.
  Filters: --level, --topic, --source (freego/extension),
  --scope (scan/lint), --search. English keyword aliases for
  search (button / form / image / dialog / etc. map to zh-TW
  desc substrings).
• a11y-moda explain <RULE_ID> — short alias for rules show.
• 9-field rule metadata — rule_id, guideline, level,
  level_name, desc, source, runtime_authoritative, wcag_url
  (WAI Quickref anchor), topic (codes/ subdir), scope (list of
  stages: scan/lint).
• examples/cursor/.cursorrules — Cursor integration.
• examples/copilot/.github/copilot-instructions.md — GitHub
  Copilot Chat integration.
• examples/aider/.aider.conf.yml — Aider integration with
  lint-cmd: hook.
• examples/generic-agent/AGENT.md — platform-agnostic
  instructions for any LLM agent (Cline, Continue, RooCode, custom).
• SKILL.md — argument-hint, knowledge-query section, MODA 編碼
  速查 (HM/GN/CS/AR/FA/SC + C/E suffix decoder).
• Workflow positioning expanded from T1–T7 to T0–T7 (T0 = pre-write
  rule lookup).

Changed (BREAKING)

• pyproject.toml dependencies no longer includes playwright.
  Moved to optional-dependencies.scan and optional-dependencies.all.
• CLI scan / site early-exit with friendly install message when
  --render / --render-crawl selected and Playwright not importable.
  lint and rules unaffected.

Notes

• Why no MCP server yet: rules subcommand + system-prompt
  examples (Layer 2 of cross-IDE strategy) ship first to validate the
  "edit-time query" workflow with the lowest-risk approach. MCP server
  (Layer 1, native tool integration) deferred to v0.4.0 pending real
  usage signal.
• Why no auto-install: industry consensus (pytest plugins, ruff,
  black, even Playwright itself) is to print a clear install command
  and let the user run it. Auto-install causes permission, network,
  and reproducibility issues. AI agents that read stderr will run the
  printed command on the user's behalf.
• Knowledge query is proactive, not reactive: the design intent
  is for agents to query rules BEFORE generating JSX, not just after
  lint reports an issue. Reactive lookup also works (explain <RULE_ID> after seeing a fail), but the bigger UX win is writing
  compliant code from the start.