Investor disagreement is central to theories of trading, yet existing proxies are either low-frequency or depend on platform-specific investor labels. We develop a daily, replicable measure of disagreement for unlabeled social media and examine its relation to trading activity in China's A-share market. Using 52.3 million investor comments on CSI 300 constituents from 2016 to 2025, we implement an LLM-based stance-simulation framework in which neutral, bullish, and bearish agents generate annotated training data that are then scaled to the full corpus using three parallel pre-trained language model classifiers. This design yields two distinct disagreement measures: within-group disagreement, capturing dispersion within a shared stance-conditioned interpretation, and cross-group disagreement, capturing divergence across stance-conditioned interpretations. We find that within-group disagreement is negatively associated with abnormal trading volume, whereas cross-group disagreement is positively associated with it. The negative within-group relation strengthens during periods of high policy uncertainty, consistent with a value-of-waiting channel, while the positive cross-group relation is stronger when disagreement involves bearish interpretations and under favorable market conditions, consistent with gains from trade under competing valuation models. The results are robust to alternative assumptions about latent group weights, before-market-open measurement, alternative LLM annotation, and a post-release annotation sample. These findings show that the trading implications of disagreement depend on whether heterogeneity reflects within-stance uncertainty or cross-stance conflict. More broadly, the approach provides a scalable framework for measuring belief heterogeneity in unlabeled text environments.