Autovacuum blocked by long transaction

Your primary Postgres (16) has been degrading all afternoon. Write latency on a hot orders table has crept from 4ms to 90ms p99, the table's on-disk size grew 40% since morning, and pg_stat_user_tables shows n_dead_tup climbing past 12M with last_autovacuum stuck at 09:14 this morning. Autovacuum workers are running (you see them in pg_stat_activity) but never finishing on this table. A nightly analytics export job started at 09:10 and is still active, holding a transaction open. Nothing was deployed today. Walk through how you triage this and what you do to stop the bleeding versus fix it durably.