One Window, Four Edits

A convolutional layer hands you this single $4 \times 4$ feature map. The dividing lines mark the four pooling windows: the window is $2 \times 2$ with stride $2$, so the windows do not overlap and the pooled output is $2 \times 2$.

1 9  1 | 2  0
2 0  3 | 1  8
3------+------
4 4  0 | 7  1
5 1  2 | 0  5

The map is summarised twice over: once by max pooling, which keeps the largest value in each window, and once by average pooling, which takes the mean of the four values in each window.

Someone now proposes a single edit to the top-left window — the one holding $9, 1, 0, 3$. Exactly one of the four edits below leaves the max-pooled output completely unchanged while changing the average-pooled output. Which one?