Short answer: pick a limit, and every part is built and measured before it is accepted. The zip contains the parts in page order, named with their page ranges. A single page that is larger than the limit is kept and flagged rather than quietly dropped.
Why an estimate is not good enough
A PDF's size is not the sum of its pages: fonts, images and the cross-reference table are shared, and the same page can weigh very different amounts depending on what it references. Tools that split by “average page size” routinely produce parts that are 30% over the cap. This page saves each candidate part and reads its real length, which costs a moment and makes the number true.
When a part still comes out too big
If a single page is heavier than the limit — a full-page scan at high resolution, for example — no split can satisfy the cap without re-encoding that page. The tool emits it alone, marks it as oversized in the result, and leaves the decision to you: lower the resolution with the grayscale tool, or ask the portal whether the limit is a guideline.