We propose that human learning in high-density epistemic environments exhibits a phase transition structurally parallel to grokking in neural networks: a shift from content-level memorization to morphism-level processing, preceded by a measurable plateau in which internal progress measures improve while behavioral performance remains flat. The parallel is grounded in the mechanistic interpretability decomposition of Nanda et al. (2023) and generates five falsifiable empirical predictions with specified null models and change-point criteria. The paper introduces constellation pedagogy — a curriculum architecture organized as a typed morphism graph rather than a linear sequence — and presents a concrete 32-node, 91-edge instantiation spanning physics, mathematics, electrical engineering, and RF engineering. An interactive 3D visualization of the constellation graph is available as a companion artifact. This is a hypothesis and measurement-program paper. The phenomenology is motivational (N=1); the framework stands on its empirical predictions.