Fix hidden_size: 4096 -> 3584 to match Qwen2.5-Coder-7B-Instruct

Fix Phase 2: fusion layer processes text-only via learnable residual gate for gradient flow

Day 3 COMPLETE: Full model architecture