As artificial intelligence rapidly develops, time-delay reservoir computing (TDRC) has emerged as a highly efficient and hardware-friendly framework for processing complex temporal tasks. Driven by the inevitable trend toward full hardware integration, recent advancements have focused on directly generating input mask signals using physical optoelectronic devices. However, this physical generation process unavoidably introduces mask noise. Despite extensive previous studies exploring diverse input mask designs, the specific impact of hardware-induced mask noise on TDRC computational performance has remained unexplored. To address this gap, we systematically investigate the system's maximum tolerable mask noise (STMN) within an optoelectronic TDRC system, utilizing a Mach-Zehnder modulator as the central nonlinear node. We evaluate the noise robustness of three primary mask categories: digital binary mask, digital six-level mask, and analog chaos mask. Gaussian white noise is injected to simulate hardware generation imperfections, and the system's performance is rigorously assessed via the Santa Fe chaotic time series prediction task using a normalized mean square error (NMSE) threshold of 0.1. Our findings reveal a critical trade-off between mask encoding complexity and noise tolerance. In noise-free environments, the analog chaos mask achieves optimal predictive performance. However, under noisy conditions, its performance degrades precipitously, exhibiting severe vulnerability to noise. In stark contrast, digital masks demonstrate vastly superior stability over a broad noise range. Crucially, we discovered that even within the exact same mask type, the topological complexity of the waveform pattern decisively dictates the STMN. By categorizing binary masks into simple patterns (dominated by regular, equal-duration high-low alternating sections) and complex patterns (characterized by highly irregular, multi-scale duration distributions), we observe a profound performance divergence. The system utilizing a complex-pattern binary mask can tolerate a severe noise level down to a signal-to-noise ratio (SNR) of 7.539 dB, expanding the noise tolerance margin by nearly 11 dB compared to the simple-pattern configuration (18.417 dB). An in-depth mechanistic analysis from a nonlinear dynamics perspective elucidates these phenomena. By introducing the metric of cross-period mask consistency, we reveal that digital masks prevent feature mapping distortion by preserving the topological stability of their discrete levels. Furthermore, the complex-pattern mask's irregular temporal topology covers a broader spectrum of scales, disrupting the synchrony of short-term regular fluctuations (forced periodic excitations). This multi-scale, irregular driving force induces divergent dynamical evolution across distinct virtual nodes, effectively suppressing coherent noise amplification and significantly elevating the effective dimensionality of the system. Ultimately, this research establishes a robust design principle for hardware-integrated TDRC: prioritizing digital masks combined with highly complex waveform patterns maximizes noise resilience, providing critical theoretical guidance for the future physical synthesis of TDRC input masks.