This paper revisits the idiosyncratic volatility (IVOL) puzzle in the Chinese stock market using a comparative analysis of alternative IVOL forecasting models. We show that deep learning approaches, such as the Temporal Fusion Transformer (TFT), and conventional time-series models like ARMA(1,1) substantially outperform the martingale benchmark in predictive accuracy. The IVOL puzzle proves fragile for total returns: while the Martingale forecast produces a significant negative relationship between IVOL and returns, this effect disappears under TFT-based forecasts in Fama-French regressions and is fully absorbed by the CH-4 factor model. Forecast error decomposition reveals that Martingale overestimation induces spurious negative correlations. Further decomposing returns show negative IVOL exposure overnight but positive intraday compensation, which largely offset in aggregate. The TFT model more accurately captures the true intraday risk premium, explaining the lack of robustness in the IVOL puzzle.