Support Android automation with Codex-managed ChatGPT sign-in

ARTEMIS currently supports IDE integration through MCP, but starting an autonomous Flash/Pro task still needs separately configured model-provider credentials. A user already signed in to Codex with ChatGPT cannot use that managed session to run an Android task directly from ARTEMIS.

Could ARTEMIS support an optional Codex execution path using the official Codex app-server?

Proposed initial scope:
- `artemis codex login` / `status`: reuse Codex-managed ChatGPT authentication, or initiate its browser/device-code flow.
- `artemis codex run --serial ...`: let Codex invoke a small set of validated Android tools backed by the existing Accessibility Helper.
- Preserve ARTEMIS device locking and local trace status, screenshots, UI trees, and action records.
- Leave Flash/Pro and their provider settings unchanged; clearly document the separate runner's limitations.

Codex should own credential persistence and refresh. This should not extract OAuth tokens, treat subscription credentials as API keys, or call private ChatGPT endpoints. Official protocol: https://developers.openai.com/codex/app-server

I have a working experimental implementation and will link a PR. A live test with Codex CLI 0.153.4 and an Android 16 Samsung device reused an existing ChatGPT login, navigated to Settings > Battery, and returned the correct percentage without a model API key configured in ARTEMIS. This is one smoke test, not a reliability benchmark.

Start with the proposed `artemis codex login`, `status`, and `run --serial` entry points, then read the official Codex app-server protocol. Define completion around Codex-owned authentication, validated Android tool execution, preserved device locking and trace artifacts, and documentation of runner limitations without extracting tokens or changing Flash/Pro settings.

Written by the indexing model from the issue text.