It thinks the seq scan will stop 99.5% early, after finding 21 out of 4553 qualifying tuples. But instead it has to read the entire table to actually find only 1.
The selectivity estimate of the @> operator has been substantially improved in v13. It is still far from perfect, but should be good enough to solve this problem for this case and most similar cases. Turning off fastupdate on the index would probably also solve the problem, for a different reason.
