Improve main source detection (AutoTeX)

[dginev]

This issue contains a bulk report of 147 article in the no_problems category which come out with empty HTML. A cursory look at a handful showed that was due to picking up the wrong TeX source as the main input.

The file list is attached as a ZIP here:
no-content-no-problem.filelist.txt

With gratitude to Tianning Zhang, who provided the list of empty results, distilled from the ar5iv-04.2024 dataset.

A successful resolution could use them as a mini testbench where we at least check that each produced HTML file has HTML content with a document body.