I think CCA means in this context Canonical Correlation Analysis, and not Canonical Correspondence Analysis. This latter method is a different method mostly used in ecology, and it is only very loosely related to Canonical Correlation.

A comment on the natural orthogonality. There has been some debate on the difference between the Artic Oscillation and North Atlantic Oscillation, which is related to this question. It boils down to the situation in which one has 3 time series, A,B,C. A pathological example would be that A and B are correlated only for even timesteps, A and C are correlated only for odd time steps; B and C are uncorrelated when considering all timesteps. If one performes a PCA of these three series, one would obtain an eigenvector where all three series are contributing, and yet B and C are uncorrelated.

I have not looked into the data, but something similar could be happening here if, for instance, the correlation between T and RW and T and MXD is only present in subsegments of the record, RW and MXD being uncorrelated for the whole record.