This study investigates large-scale coherence dynamics during human sleep using paired polysomnography (PSG) and cEEGrid EEG recordings from the OpenNeuro ds005207 dataset. Sleep EEG signals are mapped into a discrete coherence-state space and modeled as session-level Markov chains. After correcting known failure modes of time-shift surrogate nulls via surrogate-variance regularization and validating results using Markov-based surrogate families, we identify two reproducible session regimes: Normal and Coherence-shifted. We show that coherence regimes exhibit modest but structured stage dependence, dissociate by recording modality, and display transition-aligned dynamical signatures, particularly in cEEGrid recordings. State-occupancy analyses further reveal regime-specific reconfiguration around sleep-stage transitions. These findings demonstrate that sleep coherence is not static within conventional stage labels, but reorganizes dynamically in a stage-conditional, modality-dependent, and transition-coupled manner. All analysis scripts and derived figures are included to support full reproducibility.