The mechanisms underlying the contribution of interlayer bonding to friction in defected twisted graphene interfaces are revealed using fully atomistic machine-learning molecular dynamics simulations. This involves stochastic events of consecutive bond formation and rupture, that are spatially separated but not necessarily independent. A unique interlayer atomic transfer healing mechanism is discovered that can be harnessed to design a run-in procedure to restore superlubric sliding. A phenomenological model is developed that enables the extrapolation of the simulation results towards experimentally relevant sliding velocity regimes. This allows us to identify a distinct transition between logarithmic increase and logarithmic decrease of frictional stress with increasing sliding velocity. While demonstrated for homogeneous graphene interfaces, a similar mechanism is expected to occur in other homogeneous or heterogeneous defected two-dimensional material interfaces.