Sarcasm is a type of sentiment employed by humans for comedic relief. The widespread use of sarcasm is a significant reason why native Bangla speakers often misunderstand humor-based comments. The increasing use of sarcasm in the Bangla language requires further natural language processing-based study, as Bangla sarcasm is particularly challenging to detect. We present BanSarc3, a ternary-class dataset (7,984 Facebook comments: sarcastic, non-sarcastic, neutral) addressing humor misinterpretation that fuels digital conflict. A hybrid RNN-BiLSTM model, leveraging bidirectional context for morphologically rich syntax, achieves state-of-the-art 89.6% accuracy (5.12–16.82% gain over prior work). Ternary classification reduced ambiguity-driven errors by 18% versus binary frameworks. Error analysis reveals generational lexical gaps and cultural hyperbole as key challenges. This work enables safer social media ecosystems for Bangla speakers and offers a blueprint for low-resource languages through open data/model release, advocating dialect adaptation and multimodal integration for equitable NLP.