My experience was you have to put some reservation to the VMs when you want use the percentage model. It's correct that the cluster reserve some resources for a failover scenario but the commitment stays still up to you which neans in case of a host failure HA may restart all failed VMs but you'll have less resources available.
To solve this issue you have to put reservation on VMs. Once the VMs is powered on admission control reserves the entire reserved VM capacity.
I wrote the automation script that run all five minutes, calculates the proper cluster percentage and put resource reservation to all VM that follow a specific naming convention. This allows you to different between jit, prod, dev, test etc. environments.
Here the link: http://wp.me/p37XPc-4
Cheers
Rafael
http://rafaelcamison.worldpress.com