Agree re noisy neighbours, but autoscaling depends on _requests_ rather than _limits_, so you could define requests for HPA scaling but leave out the limits and have both autoscaling and no throttling.
For such clusters, it's necessary to set up stuff like the LimitRanger (https://kubernetes.io/docs/concepts/policy/limit-range/) to put a hard constant bound between requests and limits.
In the meantime you can monitor rate of throttling and rate of CPU usage to limit ratio. Nothing stops you from doing this while also monitoring response latency.
On the other hand CPU request DOES potentially leave unused CPU cycles on the table since it's a reservation on the node whether you're using it or not.
Again needs may vary.
You can overprovision limit.
This is easy to test for yourself.
I've just checked Kubernetes' docs and I have to say you are absolutely correct. Resource limits are used to enforce resource quotas, but not autoscaling.