Algorithm 1 Sensing-Aware Rollout Evaluation and Task-Conditioned Surrogate Selection

Input: Candidate surrogate set $\mathcal{F}$, planner-facing sensing–estimation interface, target planning task, rollout horizon H.

Output: Task-conditioned surrogate choice.

Step 1: Construct the planner-facing state using measured or estimated channels rather than simulator truth, as formalized by Equations (1) and (2). Step 2: Generate operationally relevant sensing conditions, including measurement noise, filtered estimated-state input, delay, dropout, asynchronous refresh, and, when available, an explicit estimator-in-the-loop interface. Step 3: Query each surrogate over the full planning horizon ($H=20$ s in this benchmark) through its implemented architecture-native interface: recursive for MLP/GRU/LSTM and direct sequence-style for TCN/Transformer. Step 4: Evaluate local accuracy (short-horizon aggregate RMSE), self-scaled long-horizon error growth (divergence rate), absolute fidelity (shared aggregate-threshold and physical native-unit views), and downstream task proxies (screening and tracking behavior). Step 5: Organize the metrics for the target task using the transparent benchmark-specific heuristic defined below: screening emphasizes observed task cost and long-horizon risk, whereas state-faithful prediction and feedback-rich control emphasize absolute-error fidelity and controller-facing assessment.