Modern software engineering increasingly involves data-centric workflows, where the purpose of a code cell lies not in its syntax but in how it transforms data. Existing automatic comment generation tools often fail to capture this intent, producing generic documentation. We propose a paradigm shift: capturing post-execution data transformations—semantic differences in the data that reveal what the code actually does—as the foundation for generating explanatory comments for data-wrangling code. To this end, we introduce a dual-encoder architecture that enriches code embeddings with execution-aware signals, pairing code with a new formal grammar of its effects on data. We evaluate the approach building a dataset of executed Python notebooks pairing code, its effect sequence, and human comments. The results show that, in instances without detected transformations, the full pipeline outperforms the baseline across metrics and human evaluation. When transformations are present, the baseline remains competitive, while the simplified data-diff variants model surpass the baseline on several metrics. We make available all software and data to encourage replication and further studies in this area. While our experiments focus on comment generation, our core contribution is broader: we introduce execution-aware embeddings — a novel modality to captures what code does to data — and argue for their applicability to a variety of downstream tasks.