Autovacuum blocked by idle transaction

Postgres 15 primary. Over the last 18 hours, write latency on a single high-churn table (events) has crept from 4 ms to 70 ms and the table's on-disk size grew 3x even though row count is flat. pg_stat_user_tables shows n_dead_tup at 240M and climbing, last_autovacuum is NULL for that table, and autovacuum_count = 0 since yesterday. Other tables vacuum fine. pg_stat_activity shows a idle in transaction connection from an analytics job opened 19 hours ago holding a transaction. CPU and IO are not saturated. Walk through how you triage and mitigate, then prevent recurrence.