Decide between a tuned GBM, a Temporal Fusion Transformer and a foundation model

A food-delivery company needs 7-day, hourly order forecasts per kitchen for 6,000 kitchens (some new, many with strong weather and promotion effects), plus the 90th percentile for rider staffing. A vendor demos a time-series foundation model that "beats everything zero-shot", an ML engineer proposes a Temporal Fusion Transformer, and the incumbent is a LightGBM global model.

1. What properties of this problem favour each of the three candidates?
2. How do you test the vendor claim fairly, including the evaluation-leakage concern?
3. Define the experiment and the decision rule: what do you compare, on which metrics, with what cost term?
4. What do you do about the 90th percentile and about cold kitchens under each candidate?
5. State clearly what in your answer is solid and what is from memory.