Issue discovered by @shrekshao in gpuweb/cts#2115 (comment):

I saw when @builtin(frag_depth) being written in fragment shader with no DepthStencilState at createRenderPipeline on Metal backend is seeing the following Metal error

Error creating pipeline state depthAttachmentPixelFormat is not valid and shader writes to depth
    at Initialize (../../third_party/dawn/src/dawn/native/metal/RenderPipelineMTL.mm:396)
    at CreateRenderPipeline (../../third_party/dawn/src/dawn/native/Device.cpp:1617)

Same situation is accepted on D3D12 backend.

We have this topic investigated on Vulkan at #1013 but not on Metal? Shall we revisit this topic (maybe validate it or simply make Metal backend to force set depthStencilState?)

My reply there:

This is a very useful find, and I think it needs to be raised to the group. Probably worth raising even if we have a workaround.

The past discussion on #1013 seemed to be specifically about whether depthWriteEnabled was true, but this error looks like it's about whether the pipeline has a depth attachment at all. If I'm understanding correctly, I don't think it's possible to workaround this by setting Metal's depth stencil state: in order to silence this error we'd have to specify a depthAttachmentPixelFormat, and that would break compatibility with the Metal render pass (which wouldn't have a matching depth attachment).

The other workaround I can think of would be to remove the frag_depth from the generated shader at pipeline creation time, but we avoid doing things that force shader generation to happen later than createShaderModule.