Anyone? Or point me in the direction of someone who can answer this.
HA does not use the metrics you use. HA is about powering on virtual machines, not guaranteeing resources... that is what DRS does. indeed this doesn't always, or probably does not reflect the real world resource usage. It is not recommended to use HA admission control as a resource management / indication tool.
Thanks for the response Duncan.
So I should ignore these metrics with respect to HA being in a position to failover a VM. In my example of 33% then my cluster would cope with the failure of one host (not taking into account fragmentation).