Anchor sizes are fixed. YOLO uses k-means unsupervised learning to select a set. Detailed explanation here
Each neural net output (grid cell location + anchor box) makes its own predictions. Each prediction includes one whole boundary box. There is no sharing or merging or combining.