count(*) per tenant in PostgreSQL 17, measured: why it got 5× slower for the largest tenant, when estimates are wrong, and what a counter row costs under load.