A CV that runs onto a second page when it was meant to be one is a small failure that a language model cannot see, because the model deals in words and the constraint lives in millimetres. The page-fit gate closes that gap. It renders the pack to an actual PDF, asks pdfinfo for the real page count, and compares it to the budget the run was given. Over budget fails; within budget passes.
The important word is real. The gate does not estimate length from a word count or a token guess, because a guess is another opinion and opinions are exactly what the drafting loop is already full of. It measures the same artefact the reader will open. A page count is a fact you cannot argue with, which is what makes it worth gating on.
Like the other gates it returns a plain pass or fail with the numbers attached, so the loop that redrafts an over-long pack knows precisely what it overshot, and by how much.