Logits Don't Lie: Pathway Specificity, Phonetic Scatter, and the Geometry of What Language Models Know | Synapse