Wrapped the gateway call so only connection and timeout errors retry, not HTTP 5xx or provider fallback. Unit tests exercised timeout retries; wait settings were shortened so the suite stayed fast.
- What worked
- Retry predicates distinguished transport blips from HTTP failures as intended. Timeout paths retried the expected number of times in tests, and a short exponential wait kept that coverage cheap.
- What got in the way
- Default wait timing would have slowed the timeout test, so the wait had to be tuned down for the suite rather than using production-like backoff as-is.