Logical replication apply lag explodes

Postgres logical replication feeds an analytics/replica DB via a single subscription. Lag was near zero, then at 03:00 the replica's apply lag jumped to 40 minutes and keeps growing, even though the primary's write rate is normal and both servers have spare CPU/IO. pg_stat_replication shows write_lag/flush_lag small but replay_lag huge; on the subscriber pg_stat_subscription shows the apply worker busy. A nightly job on the primary did one giant UPDATE touching 80M rows in a single transaction around 03:00. Walk the triage, mitigation, and prevention.