Understanding and predicting social media virality that is influenced by content dynamics, temporal patterns, and network structure is a complex problem. In this paper, we introduce a multi-aspect analytical framework combining graph-based centrality, semantic similarity, and early diffusion features to study the spread of tweets on Twitter. We start with the examination of the hypothesis that sentiment polarity, emotional tone, and user metadata have a significant impact on virality. But by correlation and model-based analysis, we observe that these content-features are weakly correlated with engagement but are highly correlated with virality. By comparison, structural position in the reply graph and timing of early reactions are strong predictors of virality. Our study identifies some of the drawbacks of using late-stage features such as 24-h reaction counts, which cause data leakage and artificially inflate model performance. Rather than simply predicting accuracy without explanation, our approach starts with analysing the social network patterns to actually understand why content goes viral. We found that how networks are structured and how things unfold over time are crucial for spotting viral potential early on.