CNA-BUG-234: CLAUDE.md's 'WebGPU Is Active' section still gives the current WEBGPU baseline as clear/present, Texture2D, buffer uploads and SpriteBatch, contradicting its own renderer paragraph, which describes a real 3D route with render targets

CNA snapshot 009d40f5  ·  Known Issues › Current bugs  ·  source links pinned to 009d40f5

The renderer paragraph and the 'WebGPU Is Active' section of CNA's CLAUDE.md disagree about what WEBGPU implements. The section's 'current baseline' list and its 'remaining shader, state, effect, render-target ... tasks' clause date from the 2026-07-12 activation; the paragraph was refreshed on 2026-08-26, AGENTS.md's copy of the section already says WebGPU is well past the initial 2D slice, and WebGPURenderer::SupportsCapability reports multiple render targets, occlusion queries, wireframe and multi-stream vertex input. The section's caution against calling WebGPU Vulkan-level or full XNA 3D parity is not itself contradicted.

Identifier

CNA-BUG-234

Category

Bug

Subsystem

Documentation & release tooling

Status

Open

Verified against

CNA 009d40f5 (009d40f5dd085c4e674d3479675fac84b12b3e0a)

Severity

Low (a triage suggestion, not a project priority)

Evidence basis

Source-verified: read at TARGET, not executed

Tests touching this area

Yes: see Current tests

Affected contract

WEBGPU renderer status as described in CLAUDE.md

Expected behaviour

One description of the WEBGPU renderer's scope, consistent with the renderer and with docs/webgpu-renderer.md, which CLAUDE.md names as the capability boundary.

Actual behaviour at TARGET

CLAUDE.md's renderer paragraph: WEBGPU is 'experimental but past a 2D baseline, with a real 3D route (stock effects, instancing, render targets, MSAA, fog parity, GPU-native compressed textures, stencil ops) and Emscripten browser support'. Its 'WebGPU Is Active' section: 'The current baseline implements native surface/device setup, clear/present, Texture2D, buffer uploads and WGSL SpriteBatch. Do not describe it as Vulkan-level or full XNA 3D parity until the remaining shader, state, effect, render-target, readback and test tasks are actually complete.' AGENTS.md's version of that section was rewritten to say the renderer is well past the initial 2D slice; CLAUDE.md's was not. WebGPURenderer::SupportsCapability (WebGPURenderer.cpp) reports multiple render targets, occlusion queries and wireframe through the shared default and multi-stream vertex input explicitly, with comments naming the tasks that made them real, and WebGPU_MRT and WebGPU_OcclusionQuery are registered tests.

Source locations

• CLAUDE.md — renderer selection paragraph and the 'WebGPU Is Active (Experimental)' section
• modules/renderers/webgpu/src/WebGPURenderer.cpp — WebGPURenderer::SupportsCapability
• docs/webgpu-renderer.md — the capability boundary CLAUDE.md defers to

Evidence

Checked by reading at 009d40f5; not executed. The instruction to not describe more than the baseline predates the 3D work the other paragraph describes.

Independent re-verification: Checked by reading at 009d40f5; not executed. The baseline bullet was written at the 2026-07-12 activation and the renderer paragraph was corrected on 2026-08-26 without touching it. The bullet's instruction is a caution against claiming Vulkan-level or full XNA 3D parity, not an instruction to describe nothing beyond the baseline, so what is stale is the baseline enumeration and the 'remaining ... tasks' clause. plan_webgpu.md counts 200 of 201 rows done, the open one being external-platform verification.

Focused reproduction

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

Current tests

WebGPU's own CTests cover the renderer; nothing reads CLAUDE.md.

Regression test

None applies; remove or date the stale baseline bullet and keep the pointer to plans/plan_webgpu.md's status summary.

Blast radius

Contributors and agents following the instruction file; the renderer is unaffected.

Workaround

Use docs/webgpu-renderer.md and plans/plan_webgpu.md's status summary.