fix: harden Codex automation lifecycle and error handling

[mastertyko]

Summary

• Handle app-server stdin failures without crashing the host, reject outstanding requests, and preserve replacement processes.
• Retry native actions only when failure is known to precede dispatch; never replay a possibly executed action because an RPC error mentions a stale thread.
• Honor explicit cancellation signals when the host context supplies a different signal.
• Reject Chrome executions that were waiting on preparation invalidated by a reset.
• Bind Chrome trust probes to the exact app-server version being considered for trust.

Each fix includes regression coverage. The changes preserve stopped-session errors, bounded pre-dispatch retries, Chrome tab ownership, uncertain-outcome handling, and existing permission behavior.

Verification

• bun run check: TypeScript passed; 244 tests passed across 22 files.
• bun run qa:host: passed.
• npm pack --dry-run: passed.
• git diff --check: passed.
• Live native smoke through OMO: computer_use_list_apps returned an application successfully.
• Live Chrome probe on app-server 0.153.4: open, observe, reload, close, and cleanup passed.
• Independent final review approved the changes with no blocking findings and reproduced the automated checks.

Notes

The Chrome live probe used a process-local trust override and did not alter persisted trust. Normal use of a newly installed app-server version still requires /codex-computer trust.

The native smoke run also emitted an existing OMO memory-reflection warning about a stale extension context. The same warning occurred before the changes and did not prevent the tool result.