Hi @thermokarst! Thank you very much for your reply
Some days ago, after some detective work, I found the answer for my question in another place. Apparently, there wasn't read loss in denoising step in earlier dada2 versions, including the one implemented in qiime2-2019.1. Since dada2 version 1.8, a quite irrelevant number of reads might be excluded during denoising.
Previous behavior ("correct everything") can be recovered in R implementation of dada2. I am afraid we can't force current QIIME2 implementation of dada2 to "correct everything", right?
Thanks again