Function drift_problem 

pub fn drift_problem() -> ScheduleProblem

The fast-drift non-stationary problem the ablation runs on — the regime where PB2’s time-awareness genuinely pays (docs/design/09-implementation.md §8.3).

It is ScheduleProblem with the optimum drifting faster than PBT’s blind perturbation step can comfortably chase (shift ≈ 1.5× the default), over more segments (8) and with a sharper per-segment parabola (penalty = 16), so that missing the moving optimum costs more. On this problem a directed GP that places the next hyperparameter continuously toward the modelled optimum beats a coin-flipped local step; on the slow-drift default problem the two are indistinguishable (time-awareness buys nothing when the optimum barely moves), which is the honest boundary of the claim (measured — see tests/pb2_oracle.rs).