fix: preserve custom Codex provider on fallback

[lucas-xu27]

Summary

• avoid Codex Desktop feature-disable flags for custom API endpoints, because the WebSocket-to-HTTPS fallback can discard the configured model provider and incorrectly call api.openai.com
• use the stable endpoint envelope for normal execution, stdin execution, resume, and the real model health probe
• add regression coverage for the endpoint command shape

Testing

• .venv/bin/pytest -q -p no:cacheprovider tests/test_cli_executor.py -k 'codex_endpoint'
• .venv/bin/pytest -q -p no:cacheprovider tests/test_cli_executor.py -k 'codex_execute_and_resume or driver_for_codex_keyed_profile_without_endpoint_still_injects_model'
• verified a real custom Responses endpoint request returns OK

[jiayuqi7813]

lgtm