The puzzle continues...

I found some reference for computing "multicoherence," which should properly estimate the potential MISO subtraction potential in situations where the witness channels themselves have nontrivial coherence. Specifically, I followed the derivations in LIGO-P990002. The underlying math is related to principal component analysis (PCA) or gram-schmidt orthogonalization. 

This produced the following results, wherein the Wiener subtraction is still below what the coherences predict. 

I've attached the data and code that produced this plot.