fix(auth): render gpt account login method selector

[eddieparc]

Summary

• render OAuth select prompts in /gpt-account add with the extension selector
• map selected labels back to browser and device_code
• cover both mappings with a regression test

Fixes #1483

Verification

• bun run --cwd packages/coding-agent test test/suite/account-extension.test.ts
• bun run check
• isolated tmux Senpi TUI smoke (tui-smoke.mjs --self-test --driver tmux)

No real OAuth, device code, or token was used.

---

Summary by cubic

Fixes /gpt-account add so OAuth login method selection works correctly (fixes #1483), and adjusts credential pooling when adding a second account.

Changes

• Renders select prompts with the extension UI selector instead of the text-input adapter, which previously showed an empty field and returned an invalid method.
• Maps the chosen label to the corresponding OAuth option id (browser or device_code).
• Promotes a flat legacy credential to a pool when adding a new login, so the existing credential remains the default slot.
• Adds the /gpt-account extension command for managing OpenAI Codex OAuth accounts (add, remove, pin, unpin, list).
• Adds regression tests covering the method mapping and credential promotion.

^{Written for commit 2a088da. Summary will update on new commits.}

[code-yeongyu]

Closing as superseded by #1486 (merged as a1ac401, closes #1485), which covers the same select → ctx.ui.select fix and additionally:

• passes placeholder and honours the per-prompt AuthPrompt.signal so the paste-the-code dialog closes when the local callback wins;
• prints the device-code user code (Enter code: <code>) next to the verification URL;
• opens the browser for auth_url in the TUI;
• shares the relay with /claude-account add (builtin/oauth-login-interaction.ts).

This branch is also based 777 commits behind main (merge-base 69d9437cd): the LAB-109 appendLoginSlot promotion and the LAB-110 gpt-account.ts command it re-adds are already on main, which is why it shows as conflicting. Thanks for the report and the reproduction — the tmux/Linux repro in #1483 is what confirmed it was not a terminal issue.