Algorithm 1 Two-Stage Self-Supervised Training Pipeline
Input: Normal training set $\mathcal{D}={\left\{{\mathcal{X}}_i\right\}}_{i=1}^N$; hyperparameters m, $\beta$, ${\lambda }_r$, ${\lambda }_d$, $E_w$
% — Stage 1a: CSD Transformer Pre-training —
1: for epoch $=1$ to $E_{1a}$ do
2:  Construct positive pairs from temporally adjacent CSD matrices $({\widehat{\mathbf{M}}}_t,{\widehat{\mathbf{M}}}_{t+1})$
3:  Generate hard negatives by injecting orthogonality noise into P–V coupling blocks
4:  Update CSD Transformer via ${\mathcal{L}}_{\mathrm{triplet}}=max(0,\parallel {\mathbf{z}}_a-{\mathbf{z}}_p\parallel -\parallel {\mathbf{z}}_a-{\mathbf{z}}_n\parallel +m)$
5: end for
% — Stage 1b: Temporal VQ-VAE Pre-training —
6: for epoch $=1$ to $E_{1b}$ do	▹ Early stopping, patience = 15
7:  Update VQ-VAE via ${\mathcal{L}}_{\mathrm{VQ-VAE}}={\mathcal{L}}_{\mathrm{VQ}}+\beta {\mathcal{L}}_{\mathrm{commit}}+{\lambda }_r{\mathcal{L}}_{\mathrm{recon}}+{\lambda }_d{\mathcal{L}}_{\mathrm{div}}$
8: end for
	▹ StatFusion receives no gradient in Stage 1
% — Stage 2: Conditional Joint Fine-tuning —
9: Load pre-trained CSD Transformer and VQ-VAE weights; randomly initialise fusion, gating, StatFusion
10: for epoch $=1$ to $E_2$ do
11:  if epoch $\le E_w$ then	▹ Gating warm-up
12:    $\mathbf{w}\leftarrow {\mathbf{w}}_{\mathrm{base}}+{\mathbf{w}}_{\min }$	▹ Fixed weights
13:  else
14:    $\mathbf{w}\leftarrow$ gating network output (Equation (22))
15:  end if
16:  ${\mathcal{L}}_{\mathrm{total}}=w_1·{\parallel {\mathbf{Z}}_{\mathrm{expected}}-{\mathbf{Z}}_{\mathrm{CSD}}\parallel }_2^2+w_2·D_{\mathrm{synergy}}$
17:  Update all parameters end-to-end	▹ Gradients flow to StatFusion via MHCA
18: end for
Output: Trained model; discard VQ-VAE decoder for inference