Failover testing with `kill -9` is manual labor and never runs in CI β yet from a driver's perspective, failover bugs are pure wire phenomena. So we moved the faults onto the wire: one TCP proxy per replica set node that can freeze, reset, or cleanly close connections, plus `hello` rewriting so the driver never sees the real addresses. The harness promptly found real bugs β including a read that spent 26 seconds hammering a dead primary.