L2 Normalisation Layer and Its Gradient

Embedding models normalise vectors to unit length before computing cosine similarity: $y = \frac{x}{\|x\|}$ (with $\|x\|$ clamped below by eps).