cb.fabric-backend-boundaries

To enforce Fabric backend boundaries, apply architecture guardrails that separate Fabric cell semantics from backend execution. This prevents cross-domain ownership issues when modifying CUDA lowering, planning, or runtime components.

The separation between Fabric cell semantics and backend execution is an architecture design contract. It ensures changes to physical operators and CUDA backend code remain well-scoped and traceable across ownership domains.

Yes, you need a design contract before editing fabric.cuda.nn declarations. The contract guides required review paths and verification steps to ensure low-level backend code changes do not cross ownership boundaries.

You discover live owners for Fabric CUDA backend components by following the discover-live-owners steps provided by the guardrails. This ensures changes to physical operators are properly scoped and reviewed before implementation.

You should use Fabric backend boundary guardrails whenever changes risk crossing ownership boundaries in CUDA lowering, planning, or runtime. This includes modifying Fabric CUDA backend execution and physical operators.

Modifying Fabric CUDA backend execution without guardrails risks mixing cell-level semantics with backend scheduling ownership. This creates cross-domain ownership issues, making changes poorly scoped and difficult to trace during reviews.