When sort or classic's group spill to disk, they will also spill to disk the last batch when they are done:

https://github.com/mongodb/mongo/blob/a3b93f233e6d6742b783a92c64697dd08e7a80ba/src/mongo/db/pipeline/group_processor.cpp#L167

https://github.com/mongodb/mongo/blob/a3b93f233e6d6742b783a92c64697dd08e7a80ba/src/mongo/db/sorter/sorter.cpp#L850

And in the PR for TEXT_OR stage.

This makes reading the results pretty straight-forward, but we pay the price of spilling an extra chunk, which can pretty big and even if it is small, we pay a lot of conversion overheads.

However sort merge iterator accepts a span of abstract iterators:

https://github.com/mongodb/mongo/blob/a3b93f233e6d6742b783a92c64697dd08e7a80ba/src/mongo/db/sorter/sorter.cpp#L450

So it should work perfectly well if we pass a bunch of file iterators as well as an in-memory iterators for the last batch that was not spilled to disk.

It may complicate the code a little bit, but should come with some performance improvements. Especially for the cases where the stage spilled only once or twice.