Traditional non-reversible acceleration often relies on operations on the gradient of the potential function, while in non-smooth and heavy-tailed settings the gradient either does not exist or is numerically unstable.