The coarse-filtering stage should ensure sufficient recall margin — it is better to over-recall and let the model eliminate candidates than to miss the correct option by saving tokens, because recall failures at the retrieval layer cannot be recovered by subsequent models.