Why a post-rollback failure transition needs its own serialization boundary. I found this while auditing a reliability claim in a small inbox processor. The processor used PostgreSQL, multiple worker instances, and FOR UPDATE SKIP LOCKED. On its no...