Time-series replica divergence repair

Design the replica-repair strategy for a distributed time-series/metrics store (N=3 replicas per partition, eventual consistency, tunable quorums) ingesting 2M writes/sec. Over time replicas drift: dropped writes, hinted-handoff replays that never landed, nodes that were down for hours. You need replicas to re-converge without re-streaming terabytes of data on every check, and without a foreground read having to fix everything. How do you detect and repair divergence efficiently?