This is actually a continuation of this issue:
It happened again, on a different cluster. Then, when I tried to reindex the data from the "faulty" index, I noticed that reindex itself created less documents than the "total" of the original index reported, so I don't think it is possible to retrieve the "missing" documents, as Elasticsearch simply doesn't return them - only counts them in the "total" for search/scroll requests.
The index request was a straight-forward reindex, with only source and dest specified, no pipelines were used.