Parallax: A Parameterized Local Linear Attention That Keeps Softmax and Adds a Learned Covariance Correction Branch Deja un comentario / Por / junio 1, 2026