Multiple-choice music benchmarks can reward two different behaviors: listening to the audio and exploiting cues in the question/options. We introduce ListenOnly-Music, a 947-item perception-required slice of MuChoMusic and MMAU-Pro, by filtering out items that two strong text-only reasoners can solve from the question and answer options alone. We audit five public audio-LLMs, from early music-specialist systems to recent 7B-8B chat models, under audio, text-only, noise, and mismatched-audio conditions. On ListenOnly-Music, earlier 2023 systems are statistically equivalent to no listening under a pre-specified TOST margin, whereas newer models show positive audio contribution. We then evaluate two inference-time interventions. Audio-contrastive decoding improves perception-required accuracy but reduces accuracy on text-solvable items, making it a diagnostic probe rather than a blanket decoding policy. Audio-first chain-of-thought helps caption-style models but hurts MCQ-tuned models. Overall, music-MCQ accuracy should be reported separately for text-solvable and perception-required items: recent audio-LLMs do listen more than earlier systems, but benchmark shortcuts still obscure how much progress is genuinely audio-grounded.