We study contextual bandits with a global budget constraint and delayed reward feedback. We consider a simple primal–dual variant of Disjoint LinUCB in which an adaptive dual variable acts as the resource price, while model updates follow a delayed-feedback heuristic. For reproducible evaluation, we build a scalable semi-synthetic benchmark on the Criteo Attribution dataset and use a stop-at-budget protocol that guarantees zero budget violations at execution time. In the benchmark, rewards are generated from an arm-specific clipped-linear model fitted on a temporal training split. We compare the method against LinUCB and fixed cost-penalty heuristics of the form UCB-γc. The experiments show that the fixed penalty is highly sensitive to both γ and the budget regime, whereas the adaptive primal–dual mechanism is more robust across tested budgets. These conclusions should be interpreted in the context of a delayed-feedback simulator with important modeling limitations.