feat: add MCP server for coding agents (Claude Code, Cursor, Windsurf)

[rajarshidattapy]

Closes #1

What this adds

An MCP (Model Context Protocol) server at mcp/index.ts that exposes webpull's crawling engine as tools callable by any MCP-compatible coding agent.

Why it belongs here

webpull's primary use case is feeding docs to LLMs. The CLI forces a manual two-step — pull first, then point the agent at the directory. This collapses it into a single agent-native call:

# Before
$ webpull https://docs.anthropic.com
$ claude "read ./docs.anthropic.com and explain streaming"

# After (agent does it all)
> "pull the Anthropic docs and show me all streaming examples"

Tools exposed

Tool	Description
pull_docs	Crawl a URL using webpull, save as local markdown
search_docs	Full-text search across pulled docs
read_doc	Read a specific file by path
list_docs	List all pulled sites and file counts
run_code	Execute JS/TS/Python/Bash in a sandboxed subprocess

Setup (one command)

claude mcp add webpull bun run /path/to/mcp/index.ts

Implementation notes

• Zero new dependencies — hand-rolled stdio JSON-RPC loop, no @modelcontextprotocol/sdk needed. Runs on Bun, same runtime webpull already requires.
• Single file — everything lives in mcp/index.ts. No new packages, no build step.
• Shells out to the installed webpull CLI — doesn't import from src/, so the MCP server stays decoupled from internal refactors.
• Docs persist across sessions — stored in ~/.webpull-docs/<hostname>/, only pulled once per site.
• run_code is scoped to developer machines — inherits the user's env, not a multi-tenant sandbox. Documented clearly.

Files changed

• mcp/index.ts — MCP server (new file)
• mcp/README.md — setup instructions for Claude Code, Cursor, Windsurf
• package.json — adds webpull-mcp bin entry