perf: reduce conversation viewer render time by up to 99%

[ChakornK]

Summary

The conversation viewer rebuilt its entire transcript on every frame. buildContentLines() re-wrapped and re-truncated every line of every message on every render, and again on every scroll event. The per-message Markdown cache from #259 skips parsing, but nothing cached the assembled lines. At thousands of lines, it becomes extremely costly to render (~200ms/frame) and is guaranteed to go over the frame time budget. The viewer subscribes to the session, so a running agent redraws on every token and pays this cost continuously.

What changed

Cache the render, not just the parse.

ConversationViewer gains a contentCache, a WeakMap keyed by the message object holding that message's rendered block (header + body, lines pre-truncated). Three things invalidate an entry: terminal width, markdown mode, or the message text. A streaming append rebuilds only the streaming message. buildContentLines() is now an assembly loop over cached blocks plus separators. The per-role rendering moved into messageLines(), which runs only on a cache miss.

Performance

Results from npm run bench:ab -- master:

Comparing working tree against master (4f572ea), 3 round(s).

benchmark                                                                       4f572ea       working      delta
-------------------------------------------------------------------------  ------------  ------------  ---------
ConversationViewer.render — cold cache (first frame) > 50 messages              3.220ms       3.090ms      -4.0%
ConversationViewer.render — cold cache (first frame) > 500 messages            33.967ms      33.075ms      -2.6%
ConversationViewer.render — markdown: assistant (default) > 50 messages         2.119ms      184.57us     -91.3%
ConversationViewer.render — markdown: assistant (default) > 500 messages       20.070ms      266.64us     -98.7%
ConversationViewer.render — markdown: assistant (default) > 5000 messages     216.221ms       1.416ms     -99.3%
ConversationViewer.render — markdown: off (raw wrap) > 50 messages              2.331ms      175.50us     -92.5%
ConversationViewer.render — markdown: off (raw wrap) > 500 messages            24.684ms      260.85us     -98.9%
ConversationViewer.render — markdown: off (raw wrap) > 5000 messages          258.652ms       1.193ms     -99.5%

Testing

• npm run lint: clean
• npm run typecheck: clean
• npm run test: 2129 passed, 7 skipped
• npm run build: clean