fix: preserve terminal buffer content across float/dock moves and grid rebuilds

[yodobrin]

Problem

When a terminal is moved between tiled ↔ floating mode (dock/undock), or when a new terminal is added in grid mode, the terminal content disappears — the pane goes blank. The PTY process is still alive, but the user loses all visible output and scrollback.

Root Cause

TerminalPanel creates an xterm.js Terminal in a useEffect. When moved:

1. Float ↔ Dock: Terminal removed from tiling tree → added to floatingPanels (or vice versa) → React unmounts old TerminalPanel, mounts new one.
2. Grid rebuild: buildGridTree() creates new split node IDs → React keys change → existing components remount.

On unmount, term.dispose() destroys the entire xterm buffer.

Fix

Use @xterm/addon-serialize to snapshot the buffer before dispose() and restore it on remount.

• On cleanup: flush pending PTY data, serialize buffer to cache
• On init: restore cached buffer at original dimensions, then fit to new container

Cache safety

Protection	Detail
Auto-expiry	Entries deleted after 10s (configurable). Remounts happen in ~1ms.
Size cap	Buffers > 2MB skipped to prevent memory bloat.
Pop semantics	Consumed on first read, preventing stale reuse.

Files Changed

• package.json — added @xterm/addon-serialize
• src/renderer/terminal-buffer-cache.ts — new module for buffer snapshots
• src/renderer/components/TerminalPanel.tsx — serialize on cleanup, restore on init

[InbarR]

Pushed two follow-ups onto this branch (commit 7f41459):

1. Stale-timer fix in terminal-buffer-cache.ts: the original schedules a setTimeout per save without tracking it. If the same id is saved twice within EXPIRY_MS (e.g. a rapid float → dock → float in <10 s), the first timer would still fire and delete the fresher snapshot, blanking the next remount. Now timers are tracked per id and cleared on every save / pop.

2. e2e regression spec (tests/e2e/float-buffer-preserved.spec.ts): writes a sentinel into the focused pane's xterm buffer, moves it tiled → floating → tiled, and asserts the sentinel survives both transitions. Locks in the fix so we can't silently regress it later.

Otherwise the original implementation reads cleanly - good catch on flushing pendingData before serialize(). Thanks for the fix!