Thanks for sharing, @perdita.
This looks like it might be related to the bug reported here: Discarded results produced by q2-cutadapt's trim-paired when working with mixed-orientation reads
This was addressed in the 2020.11 release - can you re-run using 2020.11 to check?
If not, please share the entire log produced when running with --verbose and/or DM me links to the files to reproduce these results locally.
Bigger picture - the batch-size parameter is only there to assist with cases where you might have massive amounts of samples - unix systems have a limit on how many files can be open at a time. This parameter subdivides the samples and processes them in smaller units to get around that issue. It comes at the cost of increasing the runtime, so if you don't need it, I wouldn't use it (and it sounds like in the case above, you didn't need it).
Keep me posted.
:qiime2: