[Performance]: Reduce live Markdown re-render churn while streaming

[shauryagangrade]

Problem

RichUI re-renders the streaming Markdown every 80 new characters
(_TRUNCATE_STEP = 80) under rich.live.Live. Long model responses cause
hundreds of full-screen re-renders, which is CPU-heavy and can visibly lag on
slower terminals (and is brutal for the vhs demo recordings).

Proposed approach

• Benchmark re-render cost; consider: larger/adaptive step, rendering the
  final block only on stream end, or throttling to N fps.
• Keep the "current partial output visible while streaming" behavior.

Where to look

• gcode/ui.py:20 (_TRUNCATE_STEP), :132 (the Live refresh handler),
  :164-191 (assistant_start/token/assistant_end).

Acceptance criteria

• A long model response streams with a measurable reduction in re-renders/Live updates.
• No visual regression in partial-token display.

Difficulty

Medium.