Making a replica stale on purpose to test eventual-consistency read repair

When testing read repair in a quorum KV store, don't make the crashed replica stale with just a missing key: write v1 while all replicas are up, sleep briefly (the put returns after W acks), kill one replica, write v2, and restart it. The replica then holds a real older version whose bytes can be diffed against the old input file, which makes both the 'stale' and 'repaired' checks unambiguous. Pick a non-replica node as coordinator so later steps that kill most replicas (to assert TIMEOUT) don't kill the coordinator too. Also, bash scripts using mapfile and associative arrays must run on Linux hosts; macOS ships bash 3.2.