Hey @kmw,
I'm not familiar with OBItools, but it does seem that with the --fastq-output option supplied it uses the Sanger PHRED 33 offset, so my guess is that it is just adding the quality scores where the overlap is (I could be entirely wrong here).
Of greater concern is I don't think using DADA2 on this data really makes sense as it is expecting raw sequence data which allows it to infer where sequencing error may have occurred. @benjjneb, is this a fundamental expectation of DADA2, or would you still get "reasonable", but not ideal, results if merging was done before-hand?
Perhaps we can focus on getting your paired-end data into QIIME 2 as SampleData[PairedEndSequencesWithQuality] (skipping OBItools)? Then dada2 denoise-paired can take you the rest of the way.
Otherwise another option might be to do your own quality filtering, and then import an OTU table and representative sequences; using QIIME 2 only for the "downstream" analysis.