It somehow had not occurred before, but the new split_adapter tests occasionally fail with slab_sort and no other sorter. I have to investigate further, but so far I observed that, after the first split_adapter pass, the left part of the collection is correctly sorter so it seems unlikely that the issue comes from split_adapter.

Next step is: isolate a failing split_sorter(slab_sort) input, check that everything is normal with the right part of the input after the first split_adapter pass, then run slab_sort alone on it to confirm that the issue really comes from there. Once it is confirmed, find the bug in slab_sort and fix it.