I got very excited running across this post. The first part of your analysis is a good demonstration of how biomotor abilities are actually latent variables which are dimensionally reduced from specific manifest contextual measures (although you are telling the manifest measure seeds to strongly correlate in R, so the model is already mostly pre-determined). This is true and important. The problem here is that the second part, the operative part in refuting the primacy of biomotor abilities in programming and periodization, is based on assumptions that are incorrect. Assumptions inherent to your script: 1) All training adaptations are random, 2) training adaptations occur coterminously in time and magnitude, and 3) all training adaptations can be made equally well concurrently. With those estimation parameters, the FA model was guaranteed to get shuffled around. This is a real problem with modeling. Actual training would likely work on individual or related sets of biomotor abilities (periodization/planning/practical realities), would tend to improve attributes that are already related (specificity), would produce non-random adaptations (SAID), and would happen at differential rates depending on training goal (adaptation time horizons). For all of these reasons, in a real pre-post, clusters in the FA model would almost certainly be preserved about the latent biomotor variables.