fix: refresh stale MCP daemon for session CLI

[benvinegar]

Summary

• detect when hunk session is talking to an older live Hunk MCP daemon missing newer tools
• refresh that stale daemon from the current source tree before running newer session commands
• surface real MCP tool errors instead of vague empty-payload failures

Problem

A real live Hunk session could stay connected to an older daemon that only exposed list_sessions, get_session, and comment. In that state:

• hunk session list worked
• hunk session get worked
• hunk session comment add worked
• but hunk session context, comment list, and comment rm failed

This happened because the CLI was talking to a stale daemon rather than the current checkout's newer MCP tool surface.

Fix

• probe available MCP tools before running session commands
• if a command requires newer tools and the daemon looks like an older Hunk daemon, stop it and start a fresh daemon from the current source tree
• wait for the target live Hunk session to reconnect before proceeding
• throw actual MCP tool errors instead of silently treating them as missing payloads

Validation

• bun run typecheck
• bun test test/session-cli.test.ts test/cli.test.ts test/mcp-client.test.ts test/mcp-daemon.test.ts
• manual verification against a real live Hunk session:
  • session context works
  • session comment list works
  • session comment rm works