They aren't saying that the feature gating code adds to app size, but rather that there's a lot of code behind feature gates (for tests/staged rollout/locale-specific features/etc) that most users won't see which still add to app size.
I'm sure a lot of developers on here try and minimize their use of integration branches in the day-to-day (they are necessary for some things but keep them short and sweet) and try and get in-progress features into master ASAP - that's largely due to the fact that maintaining multiple copies of the same basic logic can quickly become extremely difficult to manage.
Localization is a really big exception to this but that's why, whenever possible, you'll see game companies limit localization to strings only - including logical statements in the realm of information to be localized can make security issues extremely fun to track down along with causing frequent usability breaks in less used localizations.
I don't know - whatever the reasons for it and no matter the resources FB has - this stuff increases in cost exponentially and if they do have a really fragmented codebase it's likely that the majority of their labour goes into process definition and QA to make sure that they don't break the Swahili language version of the landing page for China when they change their contact us link.
It works poorly on mobile, where users are not keen to reinstall an app every few days, and some do not update for months and years, because of lack of space, scarce bandwidth, old hardware, or just neglect.
Additionally, at least where Facebook is concerned and IIRC, they actually do heavily utilize out-of-app data in their mobile app. There is a good deal of code, but a lot of the UI ends up being tweaked by data that's being served to the client.