PDFs are a terrible, noisy filter for the raw data (see below)
Solutions exist for 2+3 above (for example, Mathematica or Jupiter notebooks) but PDFs are such a terrible core source of information that it’s hard to easily add them to this stack
Science is nuanced and any time you put things in spreadsheets it loses some important context. Papers capture some of this in an unstructured way.
My core issue with reason 3 is that you should be able to then add structure to your idea graph, and incorporate the primary evidence into the updated graph with notes on your uncertainties. This to me is the recursive cycle at the heart of this process. I think a more fluid tool would allow for many more layers of iteration here.