A context window is a budget, not a box
I used to picture the context window as a container. If the document fit, the model had read it. What I actually see is a budget. The opening and the last few lines get used. The middle is where a constraint goes when I paste too much and then wonder why it was ignored.
A longer prompt is not a more careful one. It costs more, it is slower, and the part I cared about is easier to skip. I stopped attaching a whole thread when the decision lived in one paragraph. I put that paragraph at the end, after the material, because that is the part the next word is standing next to.
When a task needs the whole document, I would rather have the program select the pages than ask the model to "keep all of this in mind." The program can point at a page. The model will sound as if it did, either way.