We use Gemma3-4B-it as the primary text encoder, conditioning on its penultimate-layer token hidden states. We also extract pooled text features from Jina CLIP v2, project them, and fuse them into the ...