"Online 97% of the time for two CPU-minutes of work" is the same arithmetic that bites on plain VMs: with a 12-hour weekday window a machine is used 60 of the 168 hours in a week and could be off for 108, 64%. The part worth underlining for anyone copying the fix is the dependency: stopping the thing that polls does nothing for the thing it kept awake. Outside Fabric the "does not stop when you stop the workers" list is short: managed databases (a stopped RDS instance restarts itself after seven days), load balancers and NAT gateways billed per hour of existence, disks, public IPs. Sorting the bill by item, as you did, is the only way those show up.
Gonzalo Gjuka