Rapid, product-based verification of white tea type and production year is important for label authentication and market supervision. Near-infrared (NIR) spectroscopy provides fast, reagent-saving chemical fingerprints and is well suited to this task. This study aimed to develop a unified NIR method for simultaneously identifying three white tea types and six production years, forming an 18-class joint discrimination task, and to clarify the spectral and physicochemical basis of classification. NIR spectra were collected from Shoumei, Baimudan, and Baihao Yinzhen produced in 2014, 2016, 2018, 2020, 2022, and 2024, forming 18 joint classes. To the best of our knowledge, this is the first NIR study to jointly resolve these type and production-year labels within a single classification task. To address this joint task, we developed a task-specific hierarchical dual-branch CNN-Transformer framework that integrated Min-Max and SNV-SG views from the same FT-NIR measurement through a shared convolutional backbone, a parallel CNN branch, and a six-segment local-global Transformer branch, followed by decision-level fusion. Unlike CNN-based tea models, Transformer-based spectral models for grain, drug, or protein spectra, NIRS-based origin traceability, transfer-learning tea classification, multispectral fusion, or SERS-based vintage/grade evaluation, this framework jointly addresses type and production year within one NIR framework and uses two views from the same measurement rather than fusing heterogeneous signals. Baseline comparisons, ablation experiments, spectral-segment occlusion, and physicochemical correlation analyses were conducted. The model achieved 99.17% test accuracy and 99.15% Macro-F1, with accuracies of 97.5% for Shoumei and 100% for Baimudan and Baihao Yinzhen. Macro-F1 exceeded the best baseline by 2.56 percentage points, and removing the CNN branch produced the largest performance loss. These results indicate that the joint task benefits from both local spectral detail and cross-segment integration, which single-branch or single-view models cannot fully capture. Spectral-segment importance and most physicochemical trajectories were tea-type-dependent, whereas gallic acid increased with nominal storage duration across all three types. These type-dependent patterns indicate that a shared chemical trend, such as increasing gallic acid, does not translate into a universal spectral marker, supporting the need for an integrated multisegment NIR fingerprint. These findings demonstrate that the proposed task-specific dual-view, dual-branch framework with integrated multisegment NIR fingerprints provides a promising approach for rapid verification of white tea type and production-year labels under controlled conditions, offering a product-based complement to production and distribution records for market supervision.