Long query blocking transaction pool

Your API talks to Postgres through PgBouncer in transaction pooling mode, with a small server-connection pool. Most queries are fast (1-5ms). A new reporting endpoint runs a heavy 2-8 second analytical query against the same database and the same PgBouncer pool. After it ships, you see correlated multi-second latency spikes on completely unrelated fast endpoints, and PgBouncer's 'client waiting' (cl_waiting) count climbs during the spikes. The database itself isn't CPU-bound and the fast queries are still individually fast when they finally run. How do you triage and mitigate?