Information Retrieval (IR) increasingly relies on subjective, graded, and natural-language notions of relevance, motivating the development of reproducible test collections for ranking and recommendation. In affect-sensitive music domains, however, such resources with human-validated relevance signals remain scarce. We introduce CalmSet, a test collection for emotion-tagged music retrieval and recommendation in a therapeutic context for children with Autism Spectrum Disorder (ASD). CalmSet contains 432 modular music tracks instantiated from four purposefully composed base songs with controlled provenance, each formed by distinct combinations of seven active musical layers. Each track is annotated with ranked top-3 therapeutic intent labels and natural-language descriptions. Annotations are produced via a hybrid human-in-the-loop pipeline: CLAP proposes candidate intent labels, a large language model generates auxiliary semantic descriptions, and crowd workers provide ranked judgments without exposure to model outputs; final labels are aggregated using a Borda-based procedure. As initial baselines, we evaluate five one-vs-rest multi-label classifiers over CLAP audio embeddings, observing moderate micro-F1 scores (up to 0.60) but low exact-match accuracy (<0.10), while top-3 label overlap is substantially higher (Jaccard@3 up to 0.48), motivating graded-relevance evaluation. CalmSet supports both sparse (e.g., BM25) and dense audio–text retrieval models using therapeutic labels or natural-language descriptions as queries.