The paper introduces the Anti-CrossTalk (ACT) framework to address crosstalk in cross-sectional stock ranking, where unintended information interference occurs across predictive factors. ACT identifies two forms of crosstalk: temporal-scale crosstalk, where trends, fluctuations, and shocks are entangled, and structural crosstalk, where heterogeneous relations are indiscriminately fused. To mitigate these, ACT first decomposes each stock sequence into trend, fluctuation, and shock components using Temporal Component Decomposition (TCD). The trend component is processed by a Progressive Structural Purification Encoder (PSPE) that sequentially purifies structural crosstalk via bidirectional residual subtraction and dynamic graph refinement. Fluctuation and shock components are routed to dedicated isolation branches (FCI and SCI) to prevent non-transferable local patterns from contaminating cross-stock learning. Finally, an Adaptive Component Fusion (ACF) module integrates all branch representations for ranking. Experiments on CSI300 and CSI500 datasets demonstrate that ACT achieves state-of-the-art performance in both ranking accuracy and portfolio returns, with ablation studies confirming the necessity of each anti-crosstalk design.