Human-centric intelligent environments require reliable sensing mechanisms capable of identifying occupant cognitive states in real time. In adaptive building systems, differentiating between cognitive stress and boredom is particularly important, as these states require contrasting environmental responses. This study presents a systematic cross-modal comparison of facial infrared thermography (fITI) and eye-tracking features for contactless detection of these states. Eight participants completed stress (Montreal Imaging Stress Test) and boredom (sustained visual monitoring) tasks under controlled lighting conditions (150 lux artificial lighting and 300 lux daylight), while synchronized thermal and oculomotor data were recorded. Linear mixed-effects models were applied to 320 observations per modality. Results revealed three key findings. First, spatial thermal heterogeneity within facial regions of interest (ROI), quantified through pixel standard deviation, outperformed conventional mean temperature measures. The strongest thermal effect was observed in the Lower Left Cheek (|t| = 7.92), reflecting asynchronous discharge dynamics of facial arteriovenous anastomoses under sympathetic activation. A lateralized hemispheric asymmetry (|t| = 6.93) further indicated inter-hemispheric thermal divergence between states. Second, eye-tracking measures demonstrated high discriminative capability, with nine out of ten features reaching significance and a 7.2-fold median effect-size advantage over thermal metrics; baseline-corrected pupil diameter showed the strongest separation (|t| = 46.06) with immediate and stable differentiation between conditions. In contrast, thermal responses exhibited dual temporal dynamics, with immediate separation in heterogeneity features and progressive divergence in periorbital maxima over 5–10 minutes. Third, variance decomposition identified periorbital thermal regions (ICC = 0.06–0.09) as calibration-free indicators suitable for practical sensing systems, complementing low-ICC oculomotor features. Subjective validation through NASA-TLX (d = 1.63; per-participant ρ = 0.63) confirmed that physiological markers tracked experienced cognitive states rather than task labels. The resulting feature set is compatible with commercially available sensing hardware and provides a foundation for integrating human-centric physiological monitoring into intelligent adaptive environments.