Ah, is that a reference to Feynman? You criticism is valid but the work being done in the area is certainly not of the phlogiston type. In fact, your criticism applies to much of modern medicine where the systems are so complicated and little understood that we can only catalogue observations, look for statistical correlations, construct simple models, study animal models and back hypothesis with little more than logic and statistical arguments/tests. But things are improving thanks to computers, the ability to solve complex non-linear systems numerically is allowing more mathematical models to be developed for fundamentals. But a unified theory of biological systems is a very long ways away. As more fundamental understanding is developed, the best that can be done is inductive inference. As long as there is good design of experiments, meaningful data being collected and studied and peer review, it is not cargo cult science.

As for your criticism. Spaced repetition is well backed by experiments and centuries of observation and fits with our growing understanding of memory - see Long term potentiation and PKMZeta.

Contextual interference is more tricky as it is not fully understood but it is also long studied. Unlike cargo cult, various hypotheses are being offered, tested and refined. For example, the short-term long term memory observation is based on the observation that in tests of recall, stroke patients with impaired short term memory (regardless of method) and those using the varied examples method performed better after 24h+ than those using blocked practice with functional working memory. The hypothesis being that the constructs active in long term recall are more exercised using the varied examples method, leading to better long term performance. Another hypothesis is that increasing the complexity by interleaving examples increases the challenge level and hence engagement and motivational/reward circuits parts of the brain important for learning. In addition, interleaving examples forces a more general model to be learned, encouraging generalization instead of just short term memorization.

The comparison to machine learning (they use ANNs) is: when the examples do not reflect the distribution from which they are coming from, then the learned model is less general. And in particular, when a new set of examples with different properties is trained on it interferes negatively with already learned material. In the paper they give examples of similar occurrences in rats and humans.

Finally, it is important to note that randomly interleaving examples is not useful when trying to figure something out - where short term memory is key. The observation is that contextual interference disappears for complex tasks, this is because the information requires more complex processing and integration in working memory. The ability to hold more information and more complex models in working memory correlates strongly with the ability to understand things. Interfering with this process leads to poor performance short and long term. As the material is familiarized, compressed and basic models are composed, difficulty reduces. It is here that interleaving becomes important.

The practical implications are when practicing drill exercises, rather than blocking the material according to similarity, interleave the material so a less biased model is formed and the material is more challenging to keep focus. When designing exercises break them up into the core simple, defining examples and interleave these while gradually increasing the difficulty by composing more difficult models from the simpler material and continuing to interleaving these with the simpler basis examples set.