Exit when the Codex process dies

[jmeyers91]

Summary

• exit(1) when the spawned Codex app-server process dies (child exit or stdout EOF) instead of keeping the ACP connection open
• previously nothing was pushed to the client and in-flight requests (session/prompt, a login waiting on its completion) hung forever; only the next request failed, with a generic "Connection is disposed"
• exiting surfaces the death as stdio EOF, which rejects in-flight requests at the transport layer and lets the client restart the agent
• a diagnostic line (exit code, recent codex stderr, the VC++ hint for 3221225781) is written to stderr first, so clients that surface an exited agent's stderr still show the cause — the startup-crash diagnostics that runWithProcessCheck delivers lazily, including the missing-VC++-redistributable case, stay visible
• client-initiated shutdown (stdin close) keeps its existing behaviour; login/cli subcommands are untouched
• stdout end is watched as well as exit because the win32 spawn goes through a shell, so the child handle can outlive the real codex process
• happy to rework toward an in-band error or respawn design instead if that's preferred

Testing

• npm run typecheck
• npm test (340 passed, 29 skipped)
• new acp-e2e-backend-death.test.ts (POSIX-only): SIGKILLs the codex app-server under a spawned agent, asserts the agent exits within 5s — fails on main, passes with this change; run locally like the other e2e specs
• standalone reproduction (no test framework, sandboxed CODEX_HOME, fake key, local tarpit for the model call): https://github.com/jmeyers91/codex-acp-backend-death-repro — npm run repro shows the hang on published 1.1.9; ADAPTER_ENTRY=path/to/dist/index.js node repro.mjs shows the fix
• verified on macOS arm64 and Windows x64, on both the bundled backend and the CODEX_PATH route