v0.11.0

Added

• MCP server for the taxonomy. datapitfalls mcp serves the Avoiding Data
  Pitfalls rules to Claude over stdio (Claude Code, Claude Desktop), and the web
  app serves the same server over Streamable HTTP at /api/mcp (a claude.ai
  custom connector). Claude does the checking itself, so the server makes no
  Anthropic API calls and needs no key. Tools: list_pitfalls, get_pitfall
  (with near-match suggestions for an unknown id), get_checklist (for sql,
  code, chart, image, prose, text, dashboard or document, with an
  optional domain focus) and search_pitfalls; a review_for_pitfalls
  prompt that reports findings in the engine's shape; and the resources
  datapitfalls://rules/{id} and datapitfalls://domains. Every response cites
  the book and avoidingdatapitfalls.com.
• datapitfalls/mcp export: createDataPitfallsServer(),
  handleMcpRequest(req) (a Web-standard handler for any host) and
  runStdioServer(). It loads only the taxonomy, never the engine or its SDKs.
• rulesForInput(kind, domains?) in the taxonomy API: the rules that apply
  to code, text, an image or a document. The engine and the MCP server both use
  it, so they choose rules the same way.

Changed

• The package's executable is now dist/bin.js, which starts the MCP server
  without loading the engine and hands every other command to the CLI as
  before. Engine behavior is unchanged.