Peri-urban agricultural districts must reconcile the grain area that food-security policy requires with the revenue that high-value crops can earn. Which constraint binds this trade-off is rarely identified. Prediction studies stop at accuracy; optimisation studies take the constraint set as given. Neither shows whether prediction quality or policy design governs the planting decision. We link XGBoost yield and price prediction with SHAP attribution and a revenue-maximising linear programme for five crops in Chongming District, Shanghai, China, over the ten modelled years 2015–2024, benchmark every prediction against a naive alternative, and solve the programme across a range of grain-area floors. The prediction layer is an input to the answer, not a contribution in itself. The yield model beats a per-crop historical mean for only one of five crops, and the price model is beaten by a last-observation rule on every metric. Yet the optimal structure is unchanged by ±40 % perturbations of vegetable price or ±30 % perturbations of vegetable yield, and responds only to the grain floor; the exception is maize, whose expansion rests on a single unsustained 2024 price observation. The district's official grain target of 18,667 ha is already met by its 2024 structure (18,703 ha), so food security costs nothing at the stated target. The 0.390 billion CNY (84,270 CNY per additional ha) is the cost of grain held above it. The revenue-maximising planting structure is set by the grain floor, not by market price.