At least one instance is always running, regardless of the number of queued jobs.

The is a limitation of the AWS API. It can be worked around (e.g. cancelling and creating fleets as necessary); otherwise, this surprising limitation should at least be documented.