Avoid canary deploy timeouts when proxy autostop overlaps an update

[tomr-a5]

Change Summary

What and Why:

When proxy autostop overlaps a canary Machine update, flyd can intentionally leave the replacement stopped while flyctl waits for it to start. The deploy then times out even though the update completed in the state selected by the Machine lifecycle.

How:

For canary updates of existing Machines whose request intended a launch, flyctl now honors the optional stopped target returned by Flaps. It waits for the exact returned Machine instance to settle stopped, then skips only the smoke, test-Machine, and health checks that require a started Machine. Explicit skip_launch, new creates, rolling deployments, unknown target states, and older servers retain their existing behavior, and the recovery planner's pre-deploy launch intent remains stable across retries.

Focused normal and race tests cover the legacy and recovery deployment lanes, the excluded modes, and the exact instance_id sent as the Flaps wait-version query.

Related to:

• superfly/machines-meta#8
• superfly/nomad-firecracker#4755
• Expose the target state selected for Machine updates fly-go#283

---

Documentation

• Fresh Produce
• In superfly/docs, or asked for help from docs team
• n/a