I've never used fitrlinear (didn't know it existed, in fact), but it appears that it returns a composite object and that the required RegressionLinear object is contained in the .Trained property of the object. If I try your model from just some random data, I see
Partition: [1×1 cvpartition]
Properties for class classreg.learning.partition.RegressionPartitionedLinear:
So, because of the 'Kfold,5' factor, there are five linear models; you can evaluate any one of them by
I've not looked into what the 'Kfold,5' option does, either, so I don't know what that really means to have five returned models; I guess you will. :)
I agree the documentation could be a lot more informative about the content and using it it seems...I dug this out from just copying what an example had done by referencing a property; not by description of the functions.