First of all the cause of the problem is not multicollinearity, as I first thought, but low sample size. Since you have only 8 samples, you can only do a CCA with 7 variables, for lack of degrees of freedom. 

Your CCA algorithm clearly ackowledges this, and only uses the first 7 variables it is provided with, and disregards the rest. That explains why the ordering of the variables matters.

Solutions are to get more samples, or to only include chemical variables you suspect to be most strongly related to microbiome composition. A more advanced (and in my opinion better) solution is presented in https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3590923/, which comes with automatic model selection.


  [1]: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3590923/