Keep nmshd alive when a resize races a shell's exit (#199)

[raiseCatError]

Closes #199

Stacked PR. Base: feature/175-mux-interop (PR #204). Merge order: #204 first, then this PR. Once #204 is merged, retarget this PR to dev. The only code difference from #204 is this fix.

Root cause

• node-pty's UnixTerminal.resize() calls pty.resize(fd, …) synchronously with no state check.
• node-pty closes the fd when the PTY socket closes (_close()), which can happen before it emits exit. A frontend can also send a resize just before it hears of the exit.
• Resizing a shell that has exited therefore throws ioctl(2) failed, EBADF. This reproduces deterministically.

Production paths that could hit it:

1. Service resize messages, handled inside the socket data handler: the connection keeps its owned session after the shell exits, so a synchronous throw there is uncaught and ends nmshd with every live session.
2. The attach redraw nudge (immediate resize, then a 40 ms timer): sessions.has() stays true between the fd closing and exit.
3. In-process mode: TerminalApp resizing during teardown.

Conclusion: resize after close is an expected teardown race, not a separate ordering bug.

Fix

• ShellSession: it tracks exited (set on exit), and resizes after that are no-ops. Only EBADF (isClosedPtyError) is classified as the race; it marks the session closed and is ignored. Any other error still throws.
• SessionService: every resize (client messages and both redraw steps) goes through one guarded resize(). An unexpected failure is reported to the client that caused it as {type: 'error', code: 'resize'} instead of escaping and ending the service.
• Unchanged: session cleanup, exit reporting, attach and reattach redraw, fullscreen repaint, mouse and input mode restoration, and ordinary resizes.

Tests (tests/ptyResizeRace.test.ts)

• Resize after exit and after kill.
• A burst of resizes racing the exit, over five rounds.
• Only EBADF classified; other errors still thrown.
• The service survives late resizes before, during and after a shell exit, while another session keeps working and still resizes (stty size → 33x111).
• An unexpected resize error is reported to the client, and the session stays usable.
• On the old code, four behavioral tests fail with EBADF, deterministically rather than under load.
• Repeated 5/5. The related service, lifecycle, hardening, in-process and mux tests pass 3/3.
• Build, typecheck and git diff --check are clean. The full suite passed 604/604 twice.

CI

This repository's CI runs automatically only on PRs into dev/master, so a stacked PR has no automatic checks. It was run manually on this branch with workflow_dispatch: run 36518253918, Node 22 and Node 26 both passed.

[raiseCatError]

CI at current head d58503f (workflow_dispatch): run 36520050082. Node 22 and Node 26 both passed.