Do AI coding tools actually solve the structured enterprise context problem or do they just demo well on clean repos

Stale embeddings are the part nobody talks about when evaluating AI coding tools. The demo always shows a clean repo. Nobody shows the tool six months into production on a codebase that's been actively developed, where the index hasn't kept up, where deprecated patterns are still in the retrieval layer, and where the model has no idea which shared library already handles the thing it just generated. At the Global 2000 level this is the whole problem. The AI produces code that looks fine in isolation and quietly creates technical debt in systems it was never shown. A new service gets added, the index doesn't know. A shared library API changes, the index doesn't know. A pattern gets deliberately phased out eighteen months ago, the index definitely doesn't know. Any tool actually addresses repo graph drift at the organizational level or do they all assume a clean project directory and call that context?