I am not following your line of reasoning correctly.
The watermarking is done using characters that look different from each other. Printing the book will not change those characters, and therefore it is still possible to extract the original information. Harder, but not impossible.
For this information to be lost, Evince would have to print the book in a crappy font that uses a single symbol for both. Which is possible, but I think it's not what you were going for.
The hash won't be the same, which is what matters the most.
They won't have people manually reviewing the punctuation to catch infringement, and it's highly probable that Evince printing to PDF will mess the whole internal structure, so it's hard to automate it.
If you just print it then that defeats the point. They don't care about individual copies, they want to find who is stripping DRM and posting books on torrent trackers.
How likely are you to notice if your resulting PDF had a specific unique pattern of zero-width spaces embedded with it? Or if a few curly quotes aren't curly? Let alone scenarios like the cover JPEG having stenographic info in it...
The way to check this is pretty easy though, get 2 users to bit for bit compare their books to work out if they are identical.