The autosplit functionality is best effort. In fact if we run two insert that should both trigger trigger an autosplit on the same shard, it is not guarantee that the second insert will actually cause a chunk split.
This is the case when the second insert call a ChunkSplitStateDriver::tryInitiateSplit while a previous chunk split operation is still running. The second attempt will be simply abandoned without any retry logic.

This is similar to SERVER-43960, we need to call waitForOngoingChunkSplits() after every insertion.