This paper develops a unified framework for event coreference and document-level event graph construction. It introduces a contrastive clustering objective that groups event mentions into canonical event nodes, a temporal anchoring module that enforces consistent timeline placement, and a document event graph (DEG) representation that supports downstream reasoning tasks. The approach combines neural mention encoders, contrastive losses with cluster prototypes, and a lightweight temporal consistency regularizer. The manuscript provides formal definitions, algorithmic pseudocode, theoretical discussion of clustering stability, and an experimental plan including datasets, baselines, and ablations. This work builds on the alignment primitives from P1 and the calibration methods from P2 to produce robust, calibrated event graphs for document understanding.