One of the key aspects of the immersion in the original Thief was that the visible geometry matched what you could interact with more or less directly. This was, admittedly, easier when the average level seldom went over 400 polygons on screen. But the things that the player can interact with are visibly delineated: you can see exactly which surfaces the rope arrow can be anchored to as a feature of the world properties, not as something the level designer had to customize or call out.
Which is not to say that collision volumes are bad–far from it! But the further that a virtual world gets from that one-to-one relationship, the more that an expert player will filter out. Watching me play Atelier Ryza, my friends can distinguish what is allegedly real (the sparkly collectable ingredients) from what isn’t (the non-sparkly decorative plants that you can’t interact with). I can ignore things that aren’t sparkly, and therefore I stop looking at what things look like and start concentrating on the interact-with-me glows.
Nier: Automata carefully positions its interactions to be as real as possible, within its limitations. Although its horizontal collision volumes are sometimes larger than the ruined building, its vertical collision planes are very precisely aligned so that the visible appearance of most jumps is completely honest about where you need to land to make the jump.
So each physical interaction in a game exists on a spectrum from more concrete to less concrete. Though the players’ perception of the interactions is shaped by the density of the concreteness of the surrounding interactions: a resource-bearing tree in Atelier Ryza is quite concrete, but since the other trees in the forest are only there to be immutable collision volumes it stands out. Minecraft on the other hand, has a world that is almost completely composed of concrete interactions.