Flow‑matching policies adjust inference cost by varying the number of solver steps, yet training with reinforcement learning at a single fixed budget can mask degradation at cheaper deployment budgets and even erase useful one‑step behavior from a pretrained policy.\ \ ArcFlow is a deployment‑aware algorithm that trains a single flow policy across a specified solver‑budget range. It integrates action‑aligned supervision, budget exposure, and worst‑budget checkpoint selection to achieve robust performance throughout the range.\ \ For reconstruction objectives, the bridge aggregates temporal probes into a surrogate ratio and a PPO‑style clip per action; the boundary residual equals the deployed one‑step reconstruction error, preserving essential single‑step information.\ \ Cycling exposes the shared field to the entire budget interval, while checkpoint selection optimizes the performance floor under the worst budget.\ \ For native velocity objectives, multi‑budget endpoint supervision applies the same principle without replacing the native update, keeping the architecture and deployment‑time cost unchanged.\ \ Across DM Control, manipulation, and locomotion benchmarks, ArcFlow attains the highest mean one‑step return among five flow‑policy baselines on 25 of 30 tasks under each method’s reported training settings.\ \ Ablation studies identify complementary gains from boundary supervision, the fixed‑grid bridge, budget exposure, and deployment selection.\ \ These findings establish deployment‑budget‑aware training as an effective route to strong low‑step control with a single policy serving multiple inference costs.\ \ Review