toolpath-desktop: benchmark Preview at N > 2k steps

[eliothedeman]

Follow-up from #34 review.

The Preview was exercised during #34 on a real 1737-step Claude session and felt fine. Several re-render hot paths were noted that won't matter until someone opens a longer session; worth a synthetic benchmark before they become a user complaint.

Known hot paths

• buildTree(preview.doc) — full O(N) walk on every preview mutation (see toolpath-desktop: buildTree re-normalizes on any preview mutation #39).
• renderMarkdown(t.text) per turn per render (see toolpath-desktop: memoize markdown rendering per chat turn #38).
• diff.raw.split("\n") per tool turn per render.
• StepTree.svelte's built.nodes.filter(matchesFilter) — O(N) per keystroke in the search box.
• The tree sidebar renders every matching row; at 10k steps that's DOM-heavy. Probably wants windowed rendering (e.g. svelte-virtual-list) before we hit that scale.

Ask

1. Generate synthetic fixtures: 1k, 5k, 10k-step paths. Realistic-ish mix of text turns + Edit/Write/MultiEdit tool steps.
2. Measure:
   • Time to first paint after derive
   • Median render time on a treeQuery keystroke
   • Memory-resident size of the Document + parsed DOM
3. Record the baseline; re-run after toolpath-desktop: memoize markdown rendering per chat turn #38 and toolpath-desktop: buildTree re-normalizes on any preview mutation #39 land to quantify the win.
4. If any measurement exceeds a threshold (say 200 ms on a keystroke), open targeted follow-up issues.