Cassandra compaction backlog from tombstones

Your Cassandra cluster (write-heavy time-series telemetry) starts showing rising write latency and growing dropped-mutation counts at 04:00. Per-node metrics show pending compactions climbing into the thousands, SSTable counts per read exploding, disk IO at 100%, and tombstone warnings in the logs. A retention change last week switched a high-volume table from TTL-based expiry to periodic DELETE batches. How do you triage and stabilize this, and what's the durable fix?