Fix Codex held-key input batching

[Jinwoo-H]

Summary

• shorten latency-sensitive synchronized-output coalescing to a single frame so held key repeats do not batch behind the fallback
• add a scheduler regression test for repeated latency-sensitive synchronized Codex redraw frames

Repro / validation

• Reproduced in packaged Orca with Codex TUI: held \x\ previously had steady PTY output but xterm writes batched two redraws and hit ~95ms gaps; Backspace stayed under ~66ms.
• Fresh packaged build after fix: 46 held-key keydowns -> 46 xterm writes, no double-redraw writes, max xterm write gap ~62ms, no >80ms gaps.
• pnpm exec vitest run --config config/vitest.config.ts src/renderer/src/lib/pane-manager/pane-terminal-output-scheduler.test.ts
• pnpm exec vitest run --config config/vitest.config.ts src/renderer/src/components/terminal-pane/pty-connection.test.ts src/renderer/src/lib/pane-manager/windows-pty-compatibility.test.ts src/renderer/src/components/terminal-pane/terminal-conpty-device-attributes.test.ts src/renderer/src/lib/pane-manager/pane-terminal-output-scheduler.test.ts
• pnpm exec oxlint src/renderer/src/lib/pane-manager/pane-terminal-output-scheduler.ts src/renderer/src/lib/pane-manager/pane-terminal-output-scheduler.test.ts --quiet
• pnpm run typecheck:web
• pnpm run build:electron-vite

Note: local electron-builder --dir produced dist-held-key-fix/win-unpacked and then hit the known Windows winCodeSign symlink extraction failure.

[pullfrog]

✅ No new issues found.

Reviewed changes — this PR shortens the latency-sensitive foreground coalesce delay from 32ms to 16ms so held key repeats (ticking every 30–50ms) don't batch two redraw frames behind the fallback timer.

• Reduce LATENCY_SENSITIVE_FOREGROUND_COALESCE_DELAY_MS from 32 to 16 — the new value is one frame at 60fps, shorter than the minimum key-repeat interval, so each repeat gets its own xterm.write instead of coalescing two into one batch.
• Update two existing test time advances — 31ms assertions changed to 15ms to stay just below the new 16ms threshold.
• Add regression test — "does not batch repeated latency-sensitive synchronized frames across key-repeat ticks" enqueues two Codex-style frame pairs and verifies two separate writes with the correct data.

  ^{｜ View workflow run ｜ Using DeepSeek Pro (free via Pullfrog for OSS) ｜ 𝕏}