This is a good point, but a separate issue (and kudos to PDF.js if accessibility was a purposeful design point of their implementation, not just a happy coincidence... they could certainly have chosen other avenues, such as generating a WebGL scene that would be opaque to screen readers). And Chrome's viewer should indeed be accessible, I'll raise this question internally.
For that to have been a happy coincidence is basically impossible with PDF. You're told a run of positioned glyphs to render, but then you have to figure out what text those really represent by looking at the font tables (even the simplest documents tend to contain ligatures)...and that's before you assemble the runs into sentences, paragraphs, columns etc, generally without help from the PDF, although it can optionally contain this info. The first version of PDF.js rendered positioned images of text runs, text extraction came later and was definitely deliberate.
(I was paying attention, as I wrote some of the code to do multicolumn text selection in poppler)
