Fix transcript flicker when switching to running sessions

[SmartAI]

Summary

Fix continuous transcript flicker when switching back to a running session with long, mixed-height history. The native macOS reproduction showed jumps of up to 4,011 logical pixels even while the provider was sending no new SSE content.

• Follow the measured last message plus the running-status height, rather than the estimated footer position. Avoid re-entering layout from geometry callbacks.
• Stop pooling variable-height message containers whose lazy content was being unloaded. The transcript remains virtualized.
• Apply Markdown styling synchronously so streamed updates do not expose unstyled paragraph heights for a frame.
• Add real backend/SSE regressions covering sidebar switching, frame stability, split Markdown delimiters, reading older output, and jumping back to live output. The down-arrow regression checks actual glyph pixels after the latest message leaves the viewport and receives more output. This catches the pre-fix blank destination that a transient atYEnd check missed. Also harden the native screenshot checkpoint.

Verification

• 24 focused offscreen desktop tests passed.
• All four native macOS regression cases passed across three consecutive runs: streaming at 5 ms and 25 ms with tail/paragraph stability checked within one logical pixel, plus down-arrow pixel checks for 60- and 600-paragraph messages.
• Ruff, QML lint with zero warnings, and git diff --check passed.

Qt behavior

Checked against Qt 6.11.2 documentation and source: setMarkdown() replaces the whole document, and variable-height ListView delegates affect content-size estimates. Keep Qt's native renderer and accumulated Markdown source; parsing arbitrary SSE fragments separately breaks syntax split across chunks. The rationale is documented in docs/desktop-design-system.md.