I guess the key is that the encoder knows exactly what the decoder will reconstruct at each step, so it can compensate for previous errors when choosing the next palette entry