config.json was written before the review renamed freq_cutoff to
idx_freq_cutoff, so the loader silently dropped it, and it carried
stft_clip=5 -- the notebook recipe, not the production one behind this
checkpoint. Measured on iEEG-scale input, the two recipes differ by
max|d|=0.36 on logits of amplitude 0.50, enough to flip the argmax.
The config is now regenerated by the model itself, so it cannot drift
from the constructor again, and states stft_clip=10,
stft_zscore_before_clip=True and pool_n_frames=10 explicitly.
The four final_layer.* tensors are dropped. Two were a LayerNorm the
review removed, and the other two were an untrained nn.Linear(768, 2)
init (weight std 0.0207, against 0.0209 for a fresh one) shipped as if
it were pretrained. BrainBERT pretrains by masked spectrogram
reconstruction and has no classification head to inherit. The 82
encoder tensors are bit-identical to what was published.
Add BrainBERT 'large' pretrained weights (ported from czlwang/BrainBERT)