Aurora Serverless v2 doesn't go back to 0.5 ACUs even when there is no activity on the database

We have already read this and performance monitoring / performance insight is already disabled from our end.

There could be other features that consume memory. They might not set a lower bound on the ACUs when used individually, but in combination might take up enough memory to cause scaling. Are there any options you turned on when creating this cluster (e.g. IAM) or in PostgreSQL after setting up the cluster (e.g. logical replication)? Global database (the "Add Region" action in the console) is another one that's documented as requiring a minimum number of ACUs (8).