fix(core): quitting waits for every shell to be gone, on node-pty 1.2.0-beta.15 (#127)

[Maxaubert]

Closes #127.

What you saw: "Assertion failed! conpty.node ... remove_pty_baton(baton->id)" when the stable copy was closed for an update.

Two crashes at quit, both measured:

1. The assertion dialog: node-pty 1.1.0's per-shell exit threads erase from one shared list with no lock, so shells dying together race. Fixed upstream (Windows ConPTY backend data race in ptyHandles can crash in remove_pty_baton microsoft/node-pty#921, PR #922), released only in 1.2.0-beta.13 and later. node-pty is pinned to 1.2.0-beta.15, the owner's pick. Its conpty.node no longer contains the assert.
2. A silent abort (0xc0000409): about one quit in four with ten shells, on both versions. The WER dump stack is node::FreeEnvironment -> CleanupHandles -> ThreadSafeFunction::CallJS -> Napi::Error -> abort: a shell's exit callback landed while Node was tearing down.

Fix (core):

• Every kill goes through killPty, which remembers the shell until it has exited.
• will-quit holds the quit on shellsGone(3000), but only while a shell is actually dying. A hold that ended at once lost the follow-up app.quit() and left the app running; the opacityAlpha e2e caught that.
• It waits for the agent's exitCode, which the native callback sets, not the exit event. The event lags 1 to 2.7 s for a pwsh killed mid-start (a warm shell). Quit stays at about 0.15 s.

Prism: gets the quit wait through the core bump. Its own node-pty upgrade will be a separate Prism PR.

Gate:

• typecheck, lint (0 errors) and unit (1246) pass.
• Full headless e2e: all scenarios pass.
• New quitManyShells (ten shells, quit, three rounds, must exit 0 by itself): 24 of 24 rounds clean with the fix. Without it: 1 crash in 6 rounds on 1.1.0, 4 in 12 on the beta.

Core 0.23.1, app 0.28.3.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FHHaWKR4M5QtW7Wecyuk4t