$\mathbf{Q}$, $\mathbf{K}$, $\mathbf{V}$: all derived from the *same* input sequence, typically through learned linear projections: