Add Grok Build as a coding-agent backend with automatic detection

**提交人**: Karl
**客户端版本**: 0.1.58

---

## 使用场景

Cindy can already run coding workers through agent runtimes such as Claude Code and Codex. A user who has Grok Build installed and signed in wants to use that existing local agent account in the same way.

## 当前痛点

Grok Build is not currently available as a worker/runtime option. Configuring an xAI model API is not an equivalent workaround because the requested capability is a complete coding-agent runtime that can inspect and modify a local project while reusing its own authenticated session.

## 诉求

Add Grok Build as a supported coding-agent backend. During installation or first-run setup, Cindy should detect a local Grok Build installation and its authenticated state. When available, users should be able to select Grok Build when creating a worker without separately entering an xAI API key.

## 建议方案

- Add a Grok Build runtime adapter alongside the existing Claude Code, Codex, and Pi adapters.
- Detect whether the Grok Build executable is installed and usable.
- Reuse Grok Build's existing supported login/session mechanism without reading or copying credentials directly.
- Expose Grok Build in worker creation, model discovery, runtime status, and settings UI.
- Handle missing installation, logged-out state, unsupported versions, startup failures, cancellation, and timeouts with actionable errors.
- Add integration and regression tests for detection, worker launch, message streaming, tool execution, cancellation, and session cleanup.
- Keep the integration optional so existing installations are unaffected when Grok Build is absent.
---
**OS**: darwin arm64 (25.5.0)
**界面语言**: zh-CN

Locate the existing Claude Code, Codex, and Pi adapters and the worker creation, model discovery, runtime status, and settings entry points. Read their launch, streaming, cancellation, and session-cleanup tests first, then map Grok Build detection and authenticated-state handling onto those paths. Done means optional installation detection, worker launch and tool/message flows, actionable failure handling, and regression coverage for the requested lifecycle cases.

Written by the indexing model from the issue text.