Document the local MCP server for every agent

[mando]

Summary

Adds docs/mcp-server.md (sidebar: Reference → MCP Server), a setup and tool
reference for the local MCP endpoint shipped in #299.

• Setup per agent: Claude Code, Claude Cowork and Claude Desktop (through an
  mcp-remote bridge, since the desktop app starts connectors as a command
  rather than calling an address), Codex, Cursor, Gemini CLI, VS Code, plus a
  note that Claude Chat can't reach a local address at all.
• Checking the connection: a tools/list probe, and what 404 and 405 mean.
• Tool reference: a question-to-tool table, the shared
  since/until/source/project filters, then all six tools with their
  parameters, caps and a request and response pair each.
• Session text: what the two Settings toggles gate, and what search and
  detail results leave out while text access is off.
• Without MCP: the CLI and HTTP fallbacks, moved here from Connect Your Agent.

Two client snippets on the existing page were wrong and are fixed here. Gemini
CLI needs httpUrl (it reads url as SSE, which this endpoint refuses with 405),
and Codex needs experimental_use_rmcp_client = true or it only starts
command-based servers.

connect-your-agent.md keeps its own job (what agent access is, the two toggles,
the reasoning) and links to the new page rather than repeating setup, tool detail
or the fallbacks.

Test plan

• bun run docs:build passes, and every anchor the two pages link resolves
  against the generated HTML.
• Every example response is real output, captured by seeding the synthetic
  demo store (scripts/demo.ts) and querying /mcp against it. No real
  session data or local paths.
• Verified the documented gates: agent access off returns 404, text access
  off makes get_session_transcript return its explanatory refusal and drops
  the opening prompt from search rows.

🤖 Generated with Claude Code