CNA-VGAP-025: No CI workflow selects WEBGPU or SDL_GPU; their evidence is local and scripted headless-Chrome runs, and SDL_GPU's Direct3D 12 driver route has no native execution evidence

CNA snapshot c1c316b9  ·  Known Issues › Verification gaps  ·  source links pinned to c1c316b9

Neither identity is configured by any workflow (WEBGPU is only linked into a browser bundle that is never run); their recorded runs are local (Linux, CNA's Mac mini M4) or scripted headless Chrome, and SDL_GPU's Direct3D 12 route has only Wine/vkd3d-proton runs.

Identifier

CNA-VGAP-025

Category

Verification gap

Subsystem

Testing & evidence

Status

Narrowed (partially fixed; describes only what survives)

Verified against

CNA c1c316b9 (c1c316b9c7a846ce8002809c151fcd1af14942c9)

Evidence basis

Source-verified: read at TARGET, not executed

Tests touching this area

Yes: see Current tests

Affected contract

WEBGPU and SDL_GPU renderer identities (and SDL_GPU's Direct3D 12 driver route)

Expected behaviour

Published renderer identities should be exercised by at least one automated lane on a representative host.

Actual behaviour at TARGET

A search of .github/workflows at c1c316b9 finds no WEBGPU or SDL_GPU configuration: WEBGPU appears only in emscripten-multi-renderer-ci.yml, which links a WEBGL2+WEBGPU bundle and does not run it (the workflows configure VULKAN, OPENGLES3, SDL_RENDERER, SOFTWARE, HEADLESS, STUB and, on macOS, METAL). CNA's own records place WEBGPU native runs on Linux x86_64 and on a Mac mini M4 (wgpu-native's Metal backend), and its browser runs in scripted headless Chrome; SDL_GPU's Metal-driver runs are recorded on the same Mac (apple-platforms.md, “Mac mini M4 qualification”), while its cross-compiled Windows executables run through run-wine-vkd3d-headless.sh (Wine evidence, not native presentation).

The same search finds no workflow that selects FNA3D or DIRECTX9 either, so the parity fixtures that the WEBGPU and SDL_GPU example trees register through cna_register_parity_fixtures are local evidence; the browser renderers (CNA-VGAP-023), the Windows identities (CNA-VGAP-026) and the recorded glTF golden sets (CNA-VGAP-030) have entries of their own. The classic ^SdlGpu CTests have no Linux/Vulkan record after the Apple campaign's SDL_GPU fixes (AM4-064 to AM4-070, AM4-124, AM4-174, AM4-183, AM4-184): the last one, SMG-0043 (plan_sdlgpu_modern_graphics.md), recorded 26 of 200 failing before those changes, while CNA's Mac mini M4 run of the whole SDL_GPU tree on SDL's Metal driver passes.

Source locations

• scripts/run-wine-vkd3d-headless.sh — SDL_GPU Direct3D 12 route under Wine/vkd3d-proton
• modules/renderers/sdl-gpu/examples/CMakeLists.txt — SDL_GPU registrations
• modules/renderers/webgpu/examples/CMakeLists.txt — WEBGPU registrations
• .github/workflows/emscripten-multi-renderer-ci.yml — WEBGPU linked into the WEBGL2+WEBGPU bundle, never run
• .github/workflows/gltf-renderer-stride-ci.yml — conformance matrix and the OPENGLES3-only l7-corpus job
• docs/apple-platforms.md — Mac mini M4 qualification: SDL_GPU and WEBGPU recorded runs on Metal

Evidence

Checked by reading the named sources at c1c316b9; nothing was built or executed for this entry.

Narrowed at c1c316b9: the Apple campaign registered SDL_GPU's headless tests for SDL's Metal driver (9eb060aa1, AM4-069), made the WebGPU GPU tests run on macOS instead of skipping (d89b9043e, AM4-088) and recorded full-tree runs of both identities on CNA's Mac mini M4 (e761c5f64 and 173be6d9d, AM4-093 and AM4-208, docs/apple-platforms.md). These are CNA's recorded local runs, not CI, so the entry no longer says the Metal route has no evidence; read at c1c316b9, not repeated here.

Independently reported as CNA-VGAP-048 (folded into this entry by the adversarial audit): Only the OPENGLES3 whole-corpus L7 rung runs in CI; WEBGPU, SDL_GPU, FNA3D and DIRECTX9 are configured by no workflow (WEBGPU is only linked into a browser bundle that is never run), so their registered parity fixtures are local evidence.

Focused reproduction

No focused reproduction is known. Nothing has been invented here; the evidence above is what exists.

Current tests

Large local suites exist for both renderers (SdlGpu_*, WebGPU_*); no CI job runs them.

Regression test

A Linux CI job per identity (lavapipe/llvmpipe-class drivers suffice for SDL_GPU on Vulkan and wgpu-native), plus a recorded native Windows run for SDL_GPU's Direct3D 12 driver.

Blast radius

All WEBGPU and SDL_GPU rendering; regressions surface only in local runs.

Workaround

No workaround is known.