Fix new Windows terminals freezing after the first frame

[yutigi]

Description

On Windows, every newly spawned terminal painted its first frame and then went
silent. Typing reached the shell but never echoed, and the manual Refresh button
repainted once without restoring live updates.

The cause is not rendering. createWindow() handed the renderer its ptyHost
MessagePort from a did-finish-load listener registered after the awaited
page load:

await mainWindow.loadFile(indexPath);   // resolves ON did-finish-load
...  ~320 lines later ...
mainWindow.webContents.once('did-finish-load', () => {   // never fires
  ptyHostSupervisor.attachWindow(mainWindow.webContents);
});

loadURL/loadFile resolve on did-finish-load, so the listener was armed
after the event it waited for had already fired. attachWindow never ran and
electronAPI.ptyHost.onData had no source.

That is fatal because TerminalPanel.tsx permanently short-circuits the legacy
terminal:output IPC handler as soon as a ptyId arrives, making the port the
only byte source for a ptyHost-spawned panel. Main's headless emulator stayed
correct throughout, which is exactly why terminal:getState (Refresh) still
painted a single frame and the terminal looked frozen rather than broken.

Windows only because configManager.ts sets usePtyHost: process.platform === 'win32'.
macOS/Linux never fork the supervisor and stay on the working IPC path.

Changes

• main/src/index.ts — register the port hand-off before the awaited load, and
  use on rather than once so every load re-attaches.
• main/src/ptyHost/ptyHostSupervisor.ts — attachWindow now replaces a stale
  port pair instead of returning early. webContents.id survives a reload but the
  preload closure holding the renderer end does not, so the old guard would leave a
  reloaded window permanently without a data source. The destroy-cleanup listener is
  registered only on first attach so re-attaches don't accumulate listeners.
• main/src/preload.ts — warn once when a ptyHost frame is posted with no port.
  The previous silent no-op is what hid this bug.
• attachWindow now takes a narrow structural PtyHostPortTarget instead of the
  full WebContents, so it is testable without standing up a real renderer.

Type of Change

• Bug fix (non-breaking change which fixes an issue)

Checklist

• I have read the CONTRIBUTING.md guidelines
• My code follows the code style of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have run pnpm typecheck and pnpm lint locally
• I have tested the Electron app locally

Critical Areas Modified

• State management/IPC events

Additional Notes

Evidence the listener never ran

[ptyHost] attached window webContentsId= appears nowhere across two full days
of production logs, while [ptyHost] UtilityProcess exited (code=0) does — proving
the supervisor was running and only the attach was missing. After the fix,
[ptyHost] attached window webContentsId=1 appears on a live dev run.

End-to-end verification

Drove a real terminal through the RunPane CLI against a dev instance on an isolated
PANE_DIR and a throwaway repo: ~500KB streamed in 897ms with no flow-control
stall. Acks are only emitted from inside xterm's post-render callback, so that
round trip proves main → port → preload → xterm → ack → main. With a dead port
ptyHost.ack is a silent no-op, so the PTY would have paused at the 100KB
HIGH_WATERMARK and crawled on 5s PAUSE_SAFETY_TIMEOUT cycles.

Note on the "tested the Electron app locally" checkbox: verification was done via a
PANE_DIR=~/.pane_test pnpm dev instance driven by the RunPane CLI rather than by
hand through pnpm electron-dev.

Tests

main/src/ptyHost/ptyHostSupervisor.attachWindow.test.ts adds 4 tests covering the
port hand-off and the re-attach contract. Two of them were confirmed to fail against
the previous early-return guard, so they are not vacuous. main/src/test/setup.ts
gains a MessageChannelMain / MessagePortMain stand-in.

Main suite: 22 files / 47 tests fail identically before and after this change —
the pre-existing better-sqlite3-multiple-ciphers ABI mismatch documented in
AGENTS.md, all SQLite-dependent. The 4 new passing tests are the only delta.