Rebalance stops offset commits

A loyalty-points Kafka consumer (manual offset commit, 12 partitions, 6 instances) calls an external points API per record, then commits offsets in a batch every 30s. At 23:10 customers report points credited multiple times; finance sees inflated balances. Dashboards: no lag, no consumer crashes, throughput looks normal. But the consumer group's commit-rate JMX metric has dropped to near zero on several instances since 23:05, while records continue to be processed. Logs show occasional CommitFailedException: ... rebalance warnings the handler catches and ignores. How do you triage and mitigate?