Accurate traffic matrix completion is essential for network management tasks such as anomaly detection and performance optimization. However, real-world traffic monitoring often suffers from high sparsity and non-continuous observations, including entire missing time slices caused by sampling policies or system outages. Existing methods either rely on low-rank assumptions or lack the ability to extrapolate in fully unobserved time slots. In this paper, we propose PatchTransformer-based Tensor Completion (PTC), a novel framework that addresses both random and slice-level missing data. PTC integrates a temporal-aware PatchTransformer to model short- and long-range dependencies and a decoupled triplet interaction module to capture fine-grained spatial-temporal correlations. Extensive experiments on four real-world datasets with diverse characteristics demonstrate that PTC consistently outperforms state-of-the-art baselines, achieving up to 26% improvement in normalized mean absolute error (NMAE) under sparse and irregular sampling scenarios. Moreover, PTC supports more reliable traffic reconstruction under irregular or prolonged data missing conditions, making it applicable to real-world scenarios such as intelligent network management and smart city infrastructures, where robust and continuous traffic estimation is critical for timely decision-making.