Two Numbers for Input Two

A hidden layer in the middle of a network has three inputs $x_1, x_2, x_3$ (the outputs of the layer below) and two neurons, $A$ and $B$. On one training example the inputs are

$x_1 = 2, \qquad x_2 = 0, \qquad x_3 = -1$

and the layer's weights are

Each neuron computes a weighted sum $z = w_1x_1 + w_2x_2 + w_3x_3 + b$ and passes it through its activation. Write $w_{A2}$ for the weight connecting $x_2$ to neuron $A$.

The backward pass has already worked its way back to this layer. Taking each neuron's activation into account, it has found the slope of the loss with respect to each neuron's weighted sum:

$\frac{\partial L}{\partial z_A} = 3, \qquad \frac{\partial L}{\partial z_B} = -0.5$

The layer now has two jobs. It gives each of its six weights a gradient, and it hands a share of the blame back to each of its three inputs, so that the layer below can carry the backward pass on. The blame handed to an input is the slope of the loss with respect to that input.

For input $x_2$: what is the gradient of $w_{A2}$, and how much blame does this layer hand back to $x_2$?