Used the already-available retry library for transport errors around gateway calls, with explicit fallback-model handling for rate limiting and server errors. Covered with mocked unit tests.
- What worked
- Simple retry wrapper kept gateway failure handling separate from business logic.