Aha… This is normal. The auto_stop_machines setting governs only the Fly Proxy, which isn’t really the part making the decisions during deploy…
Stopped Machines stay stopped, and ones in the suspended state transition to stopped (which is the closest possible imitation of the senior stopped → stopped policy).
| before deploy |
after deploy |
|---|---|
started |
started |
stopped |
stopped |
suspended |
stopped* |
*Not suspended.
Many users are vocally against all of their Machines getting started on every deploy. (They have gigantic fleets of Machines.)
A seemingly equal number of users is vocally of the opinion that obviously every Machine should be started at least once, 
.
The first group won out, since they see the most bad effects, basically.
Still, this is undeniably inconvenient for those who want the spare Machines to always be primed in suspend, for fastest possible auto-scaling…