If possible, since the 4500 can act as a layer 3 switch, I would try to use it as the layer 3 between all three sites, assuming that you can terminate the links on that switch as well as leaving enough room for all the servers.
If you cannot do layer 3 routing then vlans are a moot point since there is no routing device to connect the vlans together. So try to use routing if possible.
As far as STP is concerned, regardless of whether routing is present or not, for all endstation connections - and servers are considered endstations from the network point of view - you can use the portfast keyword on switch port connections to enable the port to go into forwarding mode right away. Only if BPDU frames are seen will the normal span tree protocol go into effect.
As far as fault-tolerance is concerned, consider using etherchannel on the switch; however your servers will need to be configured accordingly. What etherchannel brings is that the multiple ports participating in it are seen a one logical port by STP, thus a failure of one will not cause STP recalcs. Also, if there is more than one phy link direct between one site and another (say site 1 and site2), then you can use etherchannel on the switches at those sites to bundle them into one logical STP port.
The issues that I see with one flat network are these:
1. All stations are in the same mgmt vlan as the switches which most likely degrade the overall network impact during a broadcast strorm so much that there won't be anyway to determine the cause without a lot of guesswork.
2. Future growth patterns can be haphazard enough to the point where you are forced to use layer3. Planning for it now will address a these and other concerns later on.
Let me know if I can help any further.
