Struct VmfPhaseWeights 

pub struct VmfPhaseWeights {
    pub thq: QTensor,
    pub thk: QTensor,
    pub v_proj: QTensor,
    pub out_proj: QTensor,
    pub decay: Vec<f64>,
    pub conv: Option<Vec<f32>>,
    pub k_gate: Option<(QTensor, Vec<f32>)>,
}

Weights of one vmf_phase layer (model.layers.{i}.vmf_attn.*).

Fields

thq: QTensor

[nh·nphase, hidden] — query phase projection

thk: QTensor

[nh·nphase, hidden] — key phase projection

v_proj: QTensor

[nh·dv, hidden]

out_proj: QTensor

[hidden, nh·dv]

decay: Vec<f64>

Per-component decay exp(−exp(A_log)), len nh·2·nphase (precomputed).

conv: Option<Vec<f32>>

Short causal depthwise conv before the projections (embryo genomes born with --conv-k): flat [hidden·k] taps, identity-initialised at training start. The projections were TRAINED on the conv output — running without it scrambles the layer (measured: train-val ppl 60 exported to runtime ppl 93 539). None = pre-conv genomes, untouched.

k_gate: Option<(QTensor, Vec<f32>)>

Selective-write input gate κ (hybrid_k core, stage 71): weight [nh, hidden] + bias [nh]; κ_h = σ(W_k·x + b)_h multiplies the state WRITE (S = decay·S + κ·φk⊗v). None = classic phase core, bit-identical to the pre-κ kernel. Measured at mechanism level: knee ×2–6 earlier, restores correlated-noise robustness, LM crossover vs softmax at SEQ 512 (experiments/lc_final_merged.json).