Adequate nutrition, patient acceptance and consistent dietary adherence are required for optimal chronic glycaemic control in T2DM. Customised and clinically adapted meal planning may best support these factors. This work is situated in the field of Medical Informatics, combining explainable sequence modelling with clinically meaningful dietary restrictions, ensuring both prediction accuracy and health-oriented decision support for Type 2 Diabetes Mellitus (T2DM). In contrast to the traditional RNN based recommenders, the proposed approach explicitly incorporates adherence-aware assessment, interpretability and nutritional safety. In this paper, we propose CA-Seq2Seq-LSTM-Attn, a Constraint-Aware Seq2Seq-LSTM with Dot-Product Attention framework for generating customised seven-day meal plans (21 meals) from nutritional constraints and previous dietary behaviour. We use an LSTM-based Seq2Seq architecture in place of Transformer-based models because of its strong inductive bias for modelling short, structured temporal sequences (21 meals over 7 days), better generalisation on small high-variance dietary datasets, and lower computational complexity than data-hungry Transformer models. The model is a sequence-to-sequence Long Short-Term Memory (LSTM) architecture with attention for capturing temporal eating patterns in daily meal sequences. The constraint-aware decoding strategy ensures that the calorie, carbohydrate, sugar, protein limits, and the Glycaemic Index (GI) and Glycaemic Load (GL) thresholds are satisfied, which is crucial in the context of diabetes management. The attention mechanism makes the respective forecast more interpretable by pointing to important previous meals. We evaluate the system using enriched user-recipe interactions and comprehensive nutritional data from the Food.com dataset. The experimental results show a good ranking performance with NDCG@10 of 0.5026 and Recall@10 of 0.5909. The model also ensures a better diet consistency, as indicated by the lower variance in carbs (826.6690), weekday adherence (74.39%), intra-plan variety (0.7180), and high explainability (minimum perturbation-based faithfulness confidence reduction: −0.0288). Incorporating temporal sequence modelling, attention-based interpretability and constraint-aware decoding substantially improve clinical nutritional feasibility and recommendation accuracy. The proposed framework outperforms the traditional Collaborative Filtering (BPR-MF, NCF) and sequential deep learning baselines (GRU4Rec, SASRec) on all ranking and nutrition metrics. The proposed CA-Seq2Seq-LSTM-Attn framework is a promising decision-support tool for personalised dietary control in T2DM as it generates interpretable, nutritionally balanced and clinically matched meal recommendations.