Measure the bytes before the GPU does.

Matt Keeter posted a compact WebGPU failure that deserves a place in every graphics build. A Rust [[f32; 3]; 3] occupies 36 tightly packed bytes. WGSL gives mat3x3f a size of 48 bytes and an alignment of 16. The next field begins at byte 48 in the shader, not byte 36.

The compiler can accept both types. The host can write the buffer. The shader can read it. None of those facts proves that both sides agree on the map.

Keeter's repair parses the WGSL with Naga and compares its structure metadata with Rust offsets. Facet supplies runtime type information for the Rust side. The Facet derive is enabled only in tests, so the shipping struct does not need reflection support.

The WGSL specification defines alignment and size separately. A three-component float vector has a size of 12 bytes and an alignment of 16. An array rounds each element up to its alignment, so two vec3f values consume 32 bytes. Two Rust [f32; 3] values normally consume 24.

That is why one hand-written assertion is not enough. A generic checker should compare every field and the total span. It should fail when either side gains a field. A final runtime-sized array needs its own rule because the shader structure has no fixed final length.

Keep the test beside the buffer-writing code. Feed it the shader source used by the build. Run it in continuous integration. If generated bindings already own this contract, test the generated result instead of building a second source of truth.