Heuristic data augmentation effectively mitigates data sparsity in sequential recommendation by perturbing the original user behavior sequences and generating new ones. Compared with augmentation methods that rely on external information, complex augmentation rules, or learnable generation modules, heuristic methods perform augmentation through simple operations based on prior human knowledge, rules, or intuition, offering high efficiency and generalizability. However, existing heuristic data augmentation methods typically face two challenges. First, the augmented data often exhibits semantic bias or loss of important user behaviors compared with the original data, leading models to learn inaccurate or even erroneous user preferences. Second, existing approaches focus on augmenting sparse input behaviors, yet the model training remains constrained by the sparsity and mismatch in labeled behaviors, limiting improvements in recommendation performance. To address this issue, this study proposes a framework for repairing and enriching heuristic data augmentation in sequential recommendation (REDRec). Specifically, this study first validates the semantic deviation in sequence preferences caused by heuristic recommendation methods through empirical studies and demonstrates that a simple Jensen-Shannon divergence filtering strategy can mitigate this issue. On this basis, a repair mechanism that is sensitive to the degree of semantic shift is designed. REDRec quantifies the semantic deviation of augmented sequences using Jensen-Shannon divergence and performs an embedding representation interpolation based on differentiated beta distributions for samples with varying levels of deviation to achieve semantic repair. To enrich the labels of augmented data, REDRec performs perturbation-based multi-round inference on repaired samples and generates new soft labels by adaptively aggregating these inference results. The new soft labels are added as additional data during model training. Experiments on multiple representative backbone networks and real-world datasets validate the effectiveness and generalization of REDRec.