That is incorrect. Using biom add-metadata as described in that forum post adds the taxonomy labels of the reference OTU to which your OTUs clustered... but that does not necessarily mean it is a perfect match and that taxonomy label is correct. See here and here for details.
classify-consensus-vsearch performs the same global sequence alignment (using vsearch), but grabs the top N hits and then assigns the LCA consensus taxonomy among those (using separate code not associated with vsearch, and hence distinct from cluster-features-closed-reference). See here for a description of this process.
That post you linked to is specific for sending QIIME2-processed data off to picrust, where reference-based OTU clustering is essential for clustering sequences against ref seqs from fully sequenced genomes. It is not a recommendation for general practice.
No. It would assign the most likely taxonomy to each sequence, taking into account consensus among top hits (the idea being that you could have multiple close matches that have different taxonomy assignments! So a close match to the top hit does not make that the "correct" taxonomy).
I hope that helps!