Actually, the reservation, once set, is active whether or not contention exists. The memory is allocated as the Guest OS requests, and is then, effectively, dedicated to the VM from that point until the VM is powered off, at which time, the memory is returned to the VMkernel.
With regard to licensing, I think you're looking at things from the wrong angle.
Look at it from a scenario perspective: I've bought Essentials Plus, so I have 192GB of vRAM to work with, pooled in vCenter. With 3 hosts, that means that, on average, I can allocate 64GB of vRAM per host (the reality is that the allocation will float within the vCenter server, so a host can exceed that). If I want to ensure availability, that means that I'd need at least one host worth of resources to cover the deficit left by the host failure. This means that each host needs 96GB of physical RAM. We can take that a step further, and say that, with 100% usage of entitled vRAM, 96GB per host won't be enough, as in a host failure scenario, my memory would be running at greater than 100% capacity (we've got to take into account virtual machine and VMkernel overhead), so I'd likely build hosts with more than 96GB of physical memory.
Let's add to that the fact that vRAM usage is not an instant check, but a rolling 12-month average, so unless you've allocated 192GB of vRAM consistently over 12 months, there's nothing to worry about.
Now, if more vRAM is needed, a license addition or upgrade will be necessary, sure. But the same thing happens anywhere else - if you want more space in your refrigerator, you need to get a bigger one. If you want your car to go faster than it can go off the dealer lot, you've got to add upgrades. This is not just a VMware-specific concept.
We need to stop thinking about our datacenters in the same light we've always done. Consumption-based licensing in this manner starts us thinking about what resources cost, and help to start getting the business to think about what IT actually costs, rather than just pointing a finger at us in the corner when we cost them money. We're here to support the business. In addition to enabling all kinds of whiz-bang features in the datacenter to support the applications that actually run the business, we need to help them understand that there's a cost to that enablement. This is the first step in a long journey ahead of us as the entirety of the IT services landscape moves toward a more cloud-like model.
You may not be investing in (or even interested in) the ideas of what "the cloud" will bring, but that's most definitely the direction that VMware (and our competition) is heading right now.
