Take various data collection analysis tasks as an example - there's many cases where you may not have a good idea of how everything fits together until you can

Take various data collection analysis tasks as an example - there's many cases where you may not have a good idea of how everything fits together until you can run queries on the data, with JSON being a very common format. You many may need to combine many JSON datasets, with different layouts and potentially deeply nested structures, as well as more traditional tabular data. Being able to query all of this by putting everything into a temporary sqlite file is incredibly powerful. Sure, you're moving much more into SQL territory once you're creating indexes on JSON data as you say, but I'm not sure that always means "sloppy hack". For data tasks at least, being able to combine the power of relational + noSQL databases is incredibly productive. Doesn't mean I would necessarily put something like this into production, but for figuring stuff out combining the two approaches has a lot of merit in my experience.