Your 'Add' nodes, between them, although they can add up to more than 1, are tantamount to a logical OR of the summands.
Your first node tree says:
If the incoming ray is (a Shadow ray OR a Diffuse ray OR a Glossy ray):
evaluate the surface as Transparent (i.e. having no effect on anything seen through it).
Otherwise: evaluate the surface as the Principled BSDF,( including an IOR of 1.45)
Your second node tree says:
If the incoming ray is (a Shadow ray OR a Reflection ray):
evaluate the surface as Transparent.
evaluate the surface as the Principled BSDF.
Let's follow a ray from the camera through a pixel on the screen to the window (a). It strikes the window's front surface. This is a Camera ray, so the that point is evaluated as the BSDF.
Because the BSDF has a specular strength, but no roughness, it spawns a glossy reflection ray, (not shown) which finds the objects in the room, and contributes to the color of the point on the front surface of the window. so far, no difference between your trees: both of them show reflections.
(Note that the 'Reflection ray' condition you have in the second tree will affect how other objects see the window when they reflect it, not how the window sees other objects.)
The BSDF also spawns a transmission ray, (b), which is bent, because the BSDF has an IOR of 1.45. So far, there is no difference between the trees. But now, the ray hits the back face of the window.
(b) is Glossy. So in your first tree the backface of the window is evaluated as transparent, and the ray which will bring back a sample of the background shoots off in a straight line (c'). But in your second tree, the surface will evaluate as a backfacing BSDF, with an IOR of 1.45, so the outgoing transmission ray will bend back to its original direction (c).
That's why the window transmits such radically different views of your backgound. The second tree behaves more like real refraction. As far as I can see, the 'Glossy' condition is the only one making a significant difference to the appearance of the window.