Slang is widespread in informal communication, yet its fluidity poses challenges for natural language processing (NLP), especially when words alternate between slang and non-slang senses. While prior work has examined slang through dictionaries, sentiment analysis, and lexicon building, little attention has been given to detecting slang usage in context. We address this gap by reframing slang detection as distinguishing slang from non-slang senses of the same lexical item. To support this task, we introduce SlangTrack (ST), a diachronically structured dataset of dual-meaning words annotated at the sentence level with high inter-annotator agreement. We benchmark (1) deep learning models with static and contextual embeddings, (2) transformer-based models, and (3) large language models evaluated in zero-shot, few-shot, and fine-tuned settings. Fine-tuned transformers, especially BERT-large enriched with sentiment and emotion features, achieve the strongest performance, reaching an F1-score of 72% for slang and 92% for non-slang usage. Our findings highlight both the difficulty of contextual slang detection and the value of affective cues for improving model robustness.