A Practical Multi-Source Pipeline for Conversational Music Recommendation in the RecSys Challenge 2026

Our submission under the team name niwatori ranked third overall in the RecSys Challenge 2026 Music-CRS task. The pipeline combines candidates from 14 retrieval sources while retaining, for each candidate, which sources returned it and their ranks and scores. It uses these signals together with context–track features in a LightGBM LambdaRank reranker. A Qwen3.6-27B responder verbalizes the top ranked tracks. The submission also ranked third in nDCG@20 on the final Blind-B leaderboard. We complement the leaderboard result with a diagnostic that fits on Train and evaluates the official Devset. In this diagnostic, we compare the full pipeline with reciprocal-rank fusion, ablate per-retriever features, and decompose ranking error into retrieval misses, reranking exclusions, and ranks-2–20 ordering loss. Code is available at https://github.com/ryowk/recsys2026-niwatori.