Automated hate speech detection relies on benchmark datasets for both training and evaluation, making benchmark composition a fairness concern that precedes classifier behavior. We audit dogwhistle coverage in two widely-used benchmarks (HateXplain and Measuring Hate Speech) using an existing glossary of 340 dogwhistle entries mapped onto a four-level coding sophistication taxonomy, ranging from explicit slurs (L1) to opaque persona signals (L4). Across 59.6k posts in the benchmark union, annotation failure is the dominant failure mode: at the stereotype-based level (L2), every racial and religious group with stable match counts is labeled non-hateful at rates exceeding 60%. Collection failure, where entire dogwhistle categories are absent from the union, becomes co-dominant at higher coding levels, with twenty group x level cells showing categorical gaps that re-annotation cannot fix. These failures are unevenly distributed: race:black vs. race:middle_eastern fails the 4/5 rule on annotation DI at L2 (DI=0.661), and lgbtq:Trans/NB posts one of the highest annotator failure rates (81.2%, second only to origin:undocumented) and the lowest presence rate of any comparably-sized group at every level. We distinguish collection failure from annotation failure as mechanisms with different remediation implications, and release our mapped glossary and pipeline publicly.