Windows: Scroll position jumps to old history when assistant outputs new text

Description

When the user scrolls up even slightly (1 line) to review previous conversation history while the assistant is working, the viewport jumps erratically to a much older position in the conversation the moment new output is rendered. This makes it impossible to review earlier context while the assistant is actively generating a response.

Environment

• OS: Windows 11 Pro for Workstations
• Terminal: Default (Windows Terminal)

Steps to Reproduce

1. Start a conversation with Claude Code and give it a task that produces output over time
2. While the assistant is working (outputting text), scroll up even 1 line to review earlier messages
3. Wait for the assistant to write new content

Result: The scroll position jumps far back into old conversation history, not to the new output or the user's current scroll position.

Expected Behavior

• If the user has scrolled away from the bottom, the viewport should stay at the user's current scroll position and not jump anywhere when new output arrives.
• Auto-scroll to new output should only occur when the user is already at (or near) the bottom of the conversation.

Additional Context

• This does NOT occur in Codex (OpenAI CLI), which handles scroll anchoring correctly during active output.
• This is a major usability issue — users frequently need to review prior context while the assistant works on a task, but this bug makes that workflow impossible.