across_documents reads all files as one source
across_documents now joins the files of a run into one source and extracts one record from it. Before, it extracted each file on its own and kept one file’s value for each field. With citations off, that was always the first file.
The files are joined in the order you list them, so list them in reading order. For the parts of one recording, list them in recording order. A list field now collects items from every file. Citations still point to the original files, with pages and timecodes counted from the start of each file.
Audio mixed with documents cannot be joined. Such a run still extracts each file on its own and merges the results.
Breaking change. A run that relied on the first file winning now gets the joint record. If you need a separate result for each file, use per_document.