If I have a client on a 128k MPLS circuit going through a Cisco 2900 router and using QoS that's splitting it up into two 64k tunnels, is it possible for traffic to spike up to 256k for a short period of time or is this physically impossible even if data is showing it's not?
It's impossible to say for sure without having more information about what type of link you have.
Still, keep in mind how the tools usually measure this bandwidth. The router doesn't maintain a "bandwidth" metric that you can read from SNMP. Rather, it has counters for the number of bytes that have been input and output. So your monitoring tool will poll that number at set intervals and calculate
(ByteCount(SecondPoll) - ByteCount(FirstPoll)) / (SecondPoll - FirstPoll)
The problem with this is that you have the router providing the byte count when it receives the poll request (and it may be counting bytes as they are queued, not necessarily as they are emitted), and the tool counting time when it sends the poll request. This allows for all sorts of fun problems with buffering in various places that can lead to unexpected spikes in the data because you end up violating the unstated assumption that time delta between byte count values is the same as between the polling requests. You can also get issues with counter resets and rollovers, though most tools are good at detecting that.