When you call kmeans() then the rows of the matrix are each treated as independent points in space. So your 7125 x 312 matrix is treated as 7125 points in a 312 dimensional space.
Your call is then asking to cluster those 7125 different multi-dimensional points into 3 different clusters.
The result of the kmeans call is going to be a vector of length 7125, with each value being either 1, 2, or 3, reflecting which of the 312-dimensional centroids the row was found to be closest to.
You should not be expecting a 7125 x 312 matrix, because there is only one result per row because each row is part of a multidimensional point.
If what you are trying to do is treat the matrix like a grayscale image, and find 3 grayscale values that best cluster it, then you would have to pass in z(:) to kmeans, and reshape(idx, size(z)) afterwards.
But if that is what you want to do, consider instead using multithresh() https://www.mathworks.com/help/images/ref/multithresh.html
