WebGPU/service front-end validation ignores a ShaderStage::None binding, doesn’t cap a binding count, or doesn’t check an identity collision. The backend, unaware of the exemption, still assigns that binding a GL SSBO index or an arrayLength-from-uniform slot, or trusts the identity.

The layers now disagree: a binding beyond GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS is bound, one buffer’s arrayLength is served another’s larger bound size (defeating the robustness clamp), a virtual capture device shadows a real camera, or an isochronous transfer is created against an unclaimed endpoint. The result is an OOB read/write or a spoofed resource in the GPU/device process.

Mirror the front-end’s exemption in the backend (skip None-visibility bindings in slot assignment and bind-group application too), or add the missing collision / range / claimed-interface check so both layers agree on the same set of records.