Beyond biographic details which are 99% disconnected from my old individual instagram/fb data (FB can just buy some correlation between email/phone/IP from my credit card or private postage company or whatever) they probably only profile IRL off the last 1-2yrs of your browsing behaviour in most cases.
The classic privacy invading ads usecase is a woman who is pregnant suddenly getting ads for kid stuff before other people they know they are pregnant. They probably won't be getting those ads a year later. At most they are now labelled as potential mother in some macro profile, but for day-to-day targetting it's all about the current thing she's looking at online.
I'm not defending this stuff and macro details matter for privacy at the nation-state and higher level, but my general point is that it's a living machine, constantly adapting with a short timeframe of data to maximize utility, not one that cares much about what you googled a few yrs ago 99% of the time.
Really requires a huge effort to pull together the right people and get fast feedback loops to stop inertia.
So when you have to split the two apart, it's likely a hard legal requirement that there's no way account A can know anything about account B anymore. This means, countless metrics and logs and whatever has to be thrown away. So I imagine, there's lots of auditing of systems and every single property that is stored somewhere was scrutinized in long meetings.
Less an engineering burden than a policy and legal burden.