Json inference of multiple files is brittle

Currently, we run the inference on a per file basis, and limit the number of records over all files. This means that if the first file has 1000 entries and the second 1000, and we run the inference with a max of 1000 rows, the whole inference will be based on the first file alone.

IMO we should distribute the rows at least evenly with the number of files we are inferring. In the case above, this would correspond to 500 lines for each file.

No file or test is named in the issue. Start by locating the per-file JSON inference entry point and the tests covering the maximum row limit; done means the inference distributes the limit across multiple files instead of consuming it from the first file alone.

Written by the indexing model from the issue text.