fix(terminal): skip scroll restore on alt-buffer after split to avoid TUI cursor drift

[Jinwoo-H]

Summary

Targeted follow-up to #1298. When a full-screen TUI (Claude Code, vim, less) is drawing at the moment a split completes, scheduleSplitScrollRestore's 32ms and 200ms phases race the TUI's output — typed input lands one row below where the TUI expects and a residual fragment is left at the old cursor row.

The fix gates the scroll restore + refresh pair behind buffer.active.type. On alt-buffer we skip it (the TUI owns the screen, and alt-buffer has no scrollback, so scroll restore has nothing legitimate to do). WebGL reattach still fires on every path, so #1298's dead-canvas-after-split fix is preserved.

What stays intact from #1298

• Dead-terminal-after-split (blank WebGL canvas): ✅ WebGL reattach still runs on alt-buffer
• Worktree duplicate-key race: ✅ untouched
• Normal-buffer scroll preservation: ✅ untouched

Evidence

Bisected against origin/main. Bug reproduces at HEAD, persists at 8ae11955 (parent of #1363) and 5dc7fbda (parent of #1344), and is absent at be0e3c2f (parent of #1298). Introduction: b4f1f0ea (#1298).

Root cause in pane-split-scroll.ts::scheduleSplitScrollRestore:

• forceViewportScrollbarSync does scrollLines(-1); scrollLines(1) mid-draw
• refresh(0, rows-1) repaints rows from the xterm buffer, racing the TUI's next write

Test plan

• pnpm vitest run src/renderer/src/lib/pane-manager/ — 76/76 pass (including 3 new tests)
• pnpm typecheck — clean
• pnpm lint — clean on changed files
• Manual: Claude Code in a terminal pane, split + type — cursor lands on correct row, no rt-style residual fragment
• CI green
• Verify dead-terminal-after-split (the original fix(terminal): prevent dead terminal after split via WebGL lifecycle and worktree dedup #1298 symptom) still fixed by opening a worktree that previously triggered it

Made with Orca 🐋