One Cache, Two Houses

A hand-written network reads two numbers about a house, $x_1$ and $x_2$, and predicts its price. It has two hidden ReLU neurons and one output neuron with no activation:

$\text{prediction} = 1 \times (\text{hidden 1's output}) + 2 \times (\text{hidden 2's output}) - 1$

That makes nine parameters: four hidden weights, two hidden biases, two output weights and the output bias. The loss on one house is $(\text{prediction} - \text{price})^2$.

Like most hand-written networks, it keeps a cache. Every forward pass saves its inputs, each hidden neuron's weighted sum and output, and the prediction, overwriting whatever the previous forward pass saved. Whenever the backward pass needs a value from the forward pass, it reads it from the cache.

A training script does three things, in this order:

1. It runs the forward pass on house A, with $x_1 = 2$, $x_2 = 1$ and a price of $3$. It computes A's loss and A's output error signal, $2(\text{prediction} - \text{price})$, and holds on to that number.
2. It runs the forward pass on house B, with $x_1 = 1$ and $x_2 = 3$, just to print a sample prediction to the log.
3. It runs the backward pass for house A, starting from A's output error signal.

Nothing crashes, and nine gradients come back.

Which of them are correct for house A?