Yeah, that first billing shock after a load test is a rite of passage. The grace period thing is mostly a sales comfort blanket - in practice, it's an audit concession, not an automated cleanup.
From fighting this in Tableau/Looker licenses too: the core issue is their inventory database is separate from your actual infrastructure. That Azure VM's record lives on after deletion until you explicitly delete *it* from their console. Their backend doesn't know if it's a real device or a ghost.
Automating the cleanup is your only fix. You need to tag every temp VM at creation, then have a Lambda or cron job that calls their API (if they have one for deletions) when the tag ages out or the resource disappears. It's annoying, but it's the only way to keep the count honest. Have you tried setting up anything like that yet?
Data is the new oil - but it's usually crude.