Robert Haas is the real deal, and will certainly have tested this feature.
Perhaps I missed it, but I don't think the post specified how many rows were in the source. Could have been more than a million. Having a relatively expensive filter (regex) eliminating all results keeps the performance picture clearer. Anything done with a tuple that passes the filter would apply to both the regular and the parallel scan cases, so eliminating the results highlights just the difference between regular and parallel scans