Designing safety-critical Cyber-Physical Systems (CPS), such as self-driving cars, is inherently complex. A fundamental challenge arises from the separation between architecture modeling and safety analysis, which typically rely on different languages, tools, and modeling perspectives. As a result, engineering teams must maintain multiple, partially overlapping models of the same system. Keeping these models synchronized and consistent is labor-intensive, repetitive, and error-prone, leading to increased development cost and a higher risk of inconsistencies. We propose a unified modeling framework built on a deeply integrated metamodel foundation. Unlike Systems Modeling Language (SysML) profiles, which provide syntactic integration, our approach semantically unifies two EMF-based metamodels: a custom SysML implementation for system structure and a manually transformed Risk Analysis and Assessment Modeling Language (RAAML) metamodel for safety analysis. This integration enables a cohesive modeling environment in which architectural models and safety analyses are developed as synchronized views over a single, shared underlying system representation. As a result, consistency between system architecture and safety analysis is maintained implicitly by construction. We demonstrate that this approach provides a more robust, analyzable, and less error-prone foundation for model-based safety engineering than approaches based on dynamically applied profiles.