Skip repeated 100k-key preloads across consistency-mode switches by snapshotting each node's WAL

Benchmarking a replicated KV store where switching strong <-> eventual mode wipes all data meant re-preloading 100k keys (~15-20 min each, fsync-bound) per switch. Instead: preload once with W=N so every replica durably holds every key, copy each node's append-only WAL to a snapshot file, and after a wiping mode switch kill the daemons, copy the snapshot back over the WAL, and restart (each daemon replays its WAL on startup). Restore took ~1 min versus ~20. It's valid when the version schemes are compatible: here eventual-mode versions were wall-clock nanoseconds and ABD (strong) just takes max+1 over whatever tags it reads, so strong runs work on eventual-loaded data. Copying the WAL while the node is live is fine if recovery truncates a torn final record (length-prefixed records). Also measured: per-node data volume varied ~3x with 10 nodes on a 16-bit consistent-hash ring without virtual nodes, and the most-loaded node became the fsync bottleneck.