Cross-dataset generalization remains a major challenge in offensive language detection, especially for culturally sensitive languages such as Hebrew. A large Hebrew dataset introduced in prior work (citation omitted for double-blind review) was annotated via a taxonomy-grounded, prompt-guided LLM protocol and achieved strong in-domain results. However, performance degraded sharply on two external Hebrew corpora. We investigate whether this degradation reflects domain shift or annotation shift, i.e., differences in how offensiveness is operationalized across datasets. Using the same prompt framework and a dual-LLM agreement procedure, we re-annotate both external corpora and quantify label divergence. We observe substantial mismatch between the original and new annotations, consistent with the view that offensiveness is not objective but depends on cultural context, discourse conventions, political framing, and the interpretation of irony. Evaluating models against the new labels yields markedly improved performance, and fine-tuning with the new external labels further improves results. Overall, our findings suggest that cross-dataset failure in affective NLP tasks may often be driven by annotation mismatch rather than domain adaptation limitations, highlighting the importance of annotation validity and culturally grounded labeling protocols.