A few points to consider:

(1) the semantic and metrics layers are different concepts and may involve different architectures (see Microsoft Fabric) ;

(2) the semantic layer minimizes the impact of the challenges you mention but doesn’t automatically avoids them;

(3) the semantic layer is entity-based where the entities are built usually with the help of views and other database objects

(4) there are ways to start with metadata-based implementations of a semantic layer, though it’s not the only way;

(5) centralized metrics don’t ensure per se accuracy, clarity, and/or granularity, but are enablers for such;

(6) whether you can build a semantic layer depends on the existence of the means to do it – data objects that can be used to encapsulate, reuse and maintain the logic – which are the building blocks of any modern DBMS;

(7) the use of notebooks and data pipelines for data processing can hinder upon case the creation of a semantic layer, though the layer can still be built with a proper design and the effort is still well-spent compared with the alternatives;

(8) semantic and metrics layers make sense from the reasons you mentioned, and compared with the alternatives this approach is still the best way to go;

(9) semantic layers have been used for the past 20-25 years, though the metrics layer is relatively new, being popularized in 2021, even if the term might have been used in organizations earlier.