Incorrect matrix for ZXW circuit depending on W-spider orientation and vertex color

[semiclassical]

Consider the following ZXW circuit and its matrix as reported by ZXLive:

This diagram should correspond to the operator Z1+Z2=diag(2,0,0,-2), which the above matrix corroborates (up to a global scalar).

Now suppose we drag the 'head' of the W-spider so that it points to the left. This should not affect the spider's behavior, but the resulting matrix is different:

The boundary between these two cases seems to be whether the spider's head is on the left or right. In particular, the behavior of a 'vertical' W-spider is sensitive to small adjustments in the head's orientation.

The head being on the left also interacts oddly with the color change rule:

• If both vertices are color-changed to be red within a Proof, then the matrix returns to the first (correct) example

• If only the upper vertex is color-changed, the matrix is the second (incorrect) example

• If only the lower vertex is color changed, a new incorrect matrix occurs:

[RazinShaikh]

It's easier to see the problem with the following one. We should get the matrix for Hadamard (it's actually Had@X@X = Had)

But flipping the W orientation gives us X@Had@X:

The problem seems to be in the following block of code in pyzx but I can't figure out what the bug is.
https://github.com/zxcalc/pyzx/blob/a2775a2fec311485ec482e2a9c48eeb7b0922c80/pyzx/tensor.py#L173-L180