Add native agent plugin for Grok / xAI CLI

[djbclark]

Summary

There's no built-in agent plugin for xAI's Grok, despite the official
grok CLI (github.com/xai-org, or via the plugin marketplace at
github.com/xai-org/plugin-marketplace) supporting exactly the
non-interactive shape ralph-tui's other agent plugins rely on.

What the grok CLI already supports (v0.2.118, confirmed locally)

• Non-interactive invocation: grok -p "<prompt>" --always-approve --output-format <format>
• Multiple structured output formats, including streaming-messages-json,
  which is literally the Anthropic Messages wire format — i.e. directly
  reusable against the same event-parsing logic the claude plugin already
  has, with minimal translation.
• Auth via standard OAuth/OIDC user login (~/.grok/auth.json, issuer
  https://auth.x.ai) tied to a SuperGrok subscription — not a separate
  pay-per-token API key, so driving it from ralph-tui rides the existing
  consumer subscription rather than opening new billing.
• grok agent headless/stdio subcommands exist too, for deeper
  streaming control if the simple -p/--output-format path isn't enough.

Why this matters

Every other major coding-agent CLI with a comparable non-interactive mode
(Claude Code, Codex, Gemini CLI, Cursor Agent, OpenCode, Kimi, Droid, Kiro,
Pi) already has a built-in plugin. Grok is a notable gap given the CLI's
capabilities line up closely with the existing plugin shape.

Rough scope estimate

Comparable to the existing pi.ts/kimi.ts built-in plugins (~300-400
lines): extend BaseAgentPlugin, override meta (id grok,
defaultCommand: 'grok'), detect() (find binary + parse grok --version),
buildArgs() (-p, --always-approve/equivalent flag, --output-format streaming-messages-json, --model passthrough), and reuse/adapt the
existing Claude-wire-format JSONL parser for the streaming-messages-json
output.

Happy to contribute a PR for this if there's interest — wanted to check
whether it's already under consideration first.