It's sort of opt-in.

For websites, FLoC cohort computation only triggers if you call the document.interestCohort API or load ads - these actions are considered an opt-in. (https://github.com/WICG/floc/issues/103)

For users, it's sort of opt-in, too: You must be logged into a Google account, must have enabled Chrome history data sync, must not block third-party cookies, must have enabled Google web activity tracking and must have enabled ad personalization. (https://github.com/WICG/floc#qualifying-users-for-whom-a-coh...)

Also, you can disable FLoC via chrome://settings/privacy or chrome://flags. (https://github.com/WICG/floc/issues/103#issuecomment-8218146...)

It's not a perfect opt-in, but it's also not malware.

I believe what you are reading is ways that sync of floc classifier is disabled. That is, Chrome won't save the floc classifier in the cloud if sync is disabled, but, like history, will save it locally even if sync is disabled. (Assuming the setting isn't turned off.)

I'm not going to let anyone gaslight me. Features of a browser that aren't opt-in aren't definitionally malware.

User agents, for example, or even cookies, are not malware by any reasonable definition of the term. They present risks to the user and must be managed, but this is bounded.