variable_select_bottleneck: TFT-style input-dependent feature masks glu_combine/glu_backward: Gated Linear Unit activation (replaces ReLU) cross_branch_q_attention: 2-head attention on 12 Q-values selectivity_forward/backward: Mamba-2 learned replay priority gate Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>