In our first variant test, T238888: Variant tests: "initiation" test (A vs. B), we inferred group assignments based on user's preferences and their history of changing preferences. This was possible because the thing that distinguished group A from B is whether they were initially assigned as having their suggested edits module activated -- in other words, the Variant B users started out farther down the single branch of our workflow than the Variant A users.
But for future variant tests, it won't be a single branch -- users will have different experiences and won't be able to get to the other one. We think we should explicitly store variant assignments for analysis purposes.