The mechanisms linking speech production and perception remain underspecified, particularly in how segmental and suprasegmental features are processed across different contextual variations. This study investigated whether perceptual cue weighting could be predicted by distributional reliability of acoustic cues in production, focusing on the Mandarin Tone 2-Tone 3 contrast across both gradient coarticulatory (T1, T2, T4) and categorical tone sandhi (T3) contexts. We quantified production distributional reliability using the Bhattacharyya coefficient and assessed perceptual cue weighting through relative weight analysis. Bayesian mixed-effects modeling showed strong evidence for context-dependent acoustic distributions in production (BF₁₀ = 9.87 × 10²⁸) and perception (BF₁₀ = 4.56 × 10153). Critically, production-perception coupling emerged selectively. In gradient contexts, higher production reliability strongly predicted perceptual weighting (BF₁₀ = 12.48), with robust negative correlations for critical cues in T2 (Cohen’s d = -2.51, 95% CI [-2.93, -2.09]) and T4 contexts (d = -1.76, 95% CI [-2.28, -1.26]), but not in T1 context (d = -0.30, 95% CI [-1.02, 0.43]). No such coupling was observed for secondary cues across contexts (|d| < 0.8). In contrast, in the categorical T3 sandhi context, production statistics did not predict perceptual weights. These findings reveal a context-sensitive production-perception relationship: tightly coupled in gradient coarticulatory contexts, but dissociated in categorical rule-governed environments. This pattern supports a dual-route model for tone processing involving a statistical-auditory stream for phonetic variations and a symbolic-phonological stream for abstract alternations.