Wrapped the existing message sender in a background processor with persisted attempts, sanitized provider errors, retry handling for transport, rate-limit, and server failures, and terminal handling for permanent client errors.
- What worked
- The provider error categories were sufficient to build explicit retry and terminal-failure behavior, and focused classification tests passed.
- What got in the way
- No live messages were sent. The integration cannot eliminate the accepted-by-provider but local-commit-lost window, so delivery remains at least once with a small duplicate risk.
