There's a mistake in the PCA projection, your points are globally projected along the bottom left to upper right, which isn't good ... You know you're wrong since your need to transpose both Points and vectors! a frequent mistake is to take the two highest eigenvalued eigenvectors but confusing rows with columns within the eigenvector matrix returned by the eigenanalysis.. which leads to this kind of sloppy projection. (A simple fix would be to transpose your eigenvector matrix before you pick the two selected vectors) This python script should show you the appropriate pca projection, rather very similar to your MDS projection (altough mirrored)

Made a quick and dirty fork where you'll find a python version