Here's a simple fix. The issue was to make it work for both Hadoop1 and Hadoop2.
The test used to get the configuration from the cluster. The various Mini*Clusters mess around with the configuration enough to have this fail with Hadoop2.
Hadoop1 failed with UTIL configuration. After diff'ing the configurations I found the issue was that mapred.job.tracker was set to localhost:40462 un UTIL.
Forcing that to "local" makes it work again for Hadoop1.
So in total this patch works Hadoop1 and Hadoop2.
Planning to commit tomorrow (it's an uncontroversial test change, I think)