As for a metafeatures that you don't have.
Christophe and I developed one new metafeature called "hardness"
It measures the degree of difficulty of dataset.
This features is highly (negatively) proposal to the performance of
many algorithms.
The basic idea is to consider the number of instances whose target
label is different from one of its 1-NN.
It is quite similar to 1-NN.
It is easy to implement.
I attached the paper ( it was in ICMLA 2008) and look at Section 4.
Thanks