Long streamed responses truncate earlier output (regression since 0.97.0)

Feedback thread ID: 019c2e80-cef5-7882-a6f3-5d064884852c (Feedback recorded, no logs)

Versions:

• 0.96.0: OK
• 0.97.0+: BUG
• 0.98.0: BUG

Environment:

• WSL2 Ubuntu 24.04.3 on Windows 10.0.26200.7705
• install: npm -g @openai/codex (also reproduced on Windows npm install)

Steps:
1) Run codex interactive TUI.
2) Ask for a very long novel-length response (streaming).
3) The beginning/top of the assistant output gets truncated/disappears; cannot scroll/copy full message.

Expected: full output remains scrollable/copyable.
Actual: early part is truncated in the TUI during/after streaming.