Doubt regarding apply_similarity_pose function

I have a doubt regarding the function "apply_similarity_pose" in align.py
So given a transform from Camera world to similarity world, this function computes the reverse.
So the existing inverse function is calculated as ,

Rp = R.dot(A.T)
tp = -Rp.dot(b) + s * t

But according to derivation, the inverse function looks like

Rp = (R/s).dot(A.T)
tp = -Rp.dot(b) + t

An example for verification:
Lets assume,

R = A = I(3x3)
b = t = [0, 0, 0]
s = 2

Basically, there is only a scale difference between similarity and camera world. So coordinates in similarity world is obtained by just multiplying 2 with camera coordinates

Now according to the formula in align.py
Rp becomes I(3x3).dot(I(3x3)) = I(3x3)
tp becomes -[0, 0, 0] + s * [0, 0, 0] = [0, 0, 0]
With this Rp, tp the coordinates in similarity world coordinates remains same during Similarity->Camera world conversion and is not scaled down. With the formula that I have posted the similarity world coordinates gets scaled down appropriately.

So Am I missing something in this calculation such that the existing calculation in align.py is correct?. Or is the scale neglected ? since down the pipeline we usually use a normalised unit scale point to do triangulation?

Start in align.py at apply_similarity_pose and trace how the camera-world to similarity-world transform is represented and consumed downstream. Reproduce the identity-and-scale example from the issue, compare both inverse formulas, and verify the expected coordinate scale before documenting or correcting the behavior.