Read-only transaction after failover

A managed Postgres failover (planned maintenance) promoted a standby and updated the cluster's writer DNS endpoint. Most app instances recovered, but a subset kept erroring with 'cannot execute INSERT in a read-only transaction' for ~25 minutes, and a few writes from that subset later appeared to be lost. The healthy instances use the writer endpoint and are fine. The affected subset is an older service. Triage, mitigate, and prevent.