Function kda_recurrent 

pub fn kda_recurrent(
    q: &[f32],
    k: &[f32],
    v: &[f32],
    gate: &[f32],
    beta: &[f32],
    dims: KdaDims,
    state: &mut [f32],
) -> Vec<f32>

2026-09-25: Decode formulation: one token at a time, carrying state.

S <- S * diag(exp(g_t))          decay along the key axis, per channel
delta <- (v_t - S^T k_t) * beta_t
S <- S + k_t (x) delta
o_t <- S^T q_t

q/k are L2-normalised and q scaled by 1/sqrt(head_dim) inside this function, as the golden’s use_qk_l2norm_in_kernel=True run does. Pass raw post-conv q/k/v.