Beyond Next-Token Prediction: A Standards-Aligned Survey of Autoregressive LLM Failure Modes, Deployment Patterns, and the Potential Role of World Models | Synapse