I’m new to RNA sequencing so this may be a naive question but I’m wondering why there doesn’t seem to be a filtering step (or option) in the nf-core/RNAseq workflow after Picard markDuplicates. I though the idea of the process was to add a SAM flag to the alignments that came from PCR duplicates or Optical Duplicates, and then you can see these in the final MultiQC report. But then it never seems like these are filtered out? Shouldn’t these be marked, and then filtered out first (or enable an option to automatically filter alignments marked with this FLAG) before quantification? I don’t understand why quantify with Salmon before marking duplicates?

It's just for QC purposes. It's not safe to dedup rnaseq data, by its nature you get massive pile-ups of reads at certain locations (highly expressed genes), trying to remove duplicates would introduce more bias than it addresses.

Yes, "duplicates" are detected as having the same start (and fragment end for read pairs) and orientation. Some of these are actual PCR duplicates, others are just coincidence, i.e. apparent duplicates, but really from independent molecules. You get more apparent duplicates the deeper you sequence (because if you put 1000 reads into a 100-nt exon, a lot will map to the same places). Ideally you would use UMIs to distinguish (and filter out) PCR duplicates, but sadly those kits are not commonly used. The reason we run MarkDuplicates is that we still want to know the number, because if a sample has 95% (apparent) duplicates, you may want to look at the results with some skepticism.