[BUG] Built-in edit/patch actions fail with gpt-5.4 (xhigh), but work with gpt-5.3-codex (xhigh)

**What's Wrong?
When running Codex CLI with model gpt-5.4 and reasoning effort xhigh, built-in CLI edit/patch actions are unreliable in real tasks. The same workflow succeeds after switching the model to gpt-5.3-codex with the same xhigh reasoning effort.

**What Should Happen?
Built-in edit/patch actions should behave consistently and reliably with gpt-5.4 (xhigh), matching the behavior observed with gpt-5.3-codex (xhigh).

Steps to Reproduce

1. Start Codex CLI 0.111.0 on Windows using PowerShell 7.
2. Configure model to gpt-5.4 and reasoning effort to xhigh.
3. Run an editing task that requires CLI built-in edit/patch actions.
4. Observe unstable or failed edit/patch behavior.
5. Switch model to gpt-5.3-codex with the same xhigh reasoning effort and rerun the same task.
6. Observe that the built-in edit/patch flow works normally.

Environment

• Codex CLI version: 0.111.0
• Model (problematic): gpt-5.4
• Model (works): gpt-5.3-codex
• Reasoning effort: xhigh
• OS: Windows 10 (10.0.19045)
• Shell: PowerShell 7

Additional Information

• Submission is sent via local GH proxy as requested (http://127.0.0.1:3213).