Near-infrared spectroscopy (NIR and Vis–NIR) is widely used for rapid, non-destructive analysis in food, agriculture, pharmaceutical, process analytical technology (PAT), and bioprocess-monitoring applications. Yet, deep-learning studies in NIR chemometrics often report conflicting conclusions about convolutional neural network (CNN) design: small versus large kernels, shallow versus deep architectures, raw spectra versus preprocessing, compact models versus multi-scale networks, and random-split performance versus transfer robustness. This review argues that many of these apparent contradictions arise from incomplete conditioning rather than from inherently incompatible results. CNN performance in NIR chemometrics depends on the interaction between spectral physics, dataset regime, acquisition protocol, validation design, and deployment scenario. We therefore organize the literature around three central moderators. First, NIR signals are indirect, highly collinear, and often dominated by broad overlapping bands, scattering, temperature, and matrix effects. Second, convolutional design choices should be interpreted through receptive-field reasoning: kernel size, depth, dilation, and multi-scale branches determine which wavelength spans are available to the model, whereas the effective receptive field determines which parts of that span are actually used. Third, validation design can act as a hidden hyperparameter, because random splits may favour architectures that exploit shared batch, instrument, season, or process-run structure rather than transferable chemical information. Building on these points, we propose a conditional design framework in which preprocessing, architecture, hyperparameter tuning, transfer evaluation, interpretability, and reproducibility are treated as coupled components of the modelling pipeline. The goal is not to identify a universally optimal CNN for NIR spectra, but to move CNN–NIR chemometrics toward physics-aware, shift-aware, and reproducible model comparison.