CNA-VGAP-031: The real-world Khronos comparison matrix passes on three renderers; the planned fourth (GLTF ROBUST row 12) is open

CNA snapshot 009d40f5  ·  Known Issues › Verification gaps  ·  source links pinned to 009d40f5

Gate C's retake matrix (Fox, DamagedHelmet, Sponza and the sparse cases against the Khronos sample renderer) passes on OPENGLES3, VULKAN and DIRECTX11; the fourth renderer the plan requires is not qualified.

Identifier

CNA-VGAP-031

Category

Verification gap

Subsystem

Testing & evidence

Status

Open

Verified against

CNA 009d40f5 (009d40f5dd085c4e674d3479675fac84b12b3e0a)

Evidence basis

Recorded by CNA: CNA's own recorded run, not repeated here

Tests touching this area

Yes: see Current tests

Affected contract

glTF import and rendering compared with the Khronos glTF-Sample-Renderer on real assets

Expected behaviour

plan_gltf.md §27.2 row 12 asks for the retake matrix to pass on at least four renderers, including one Direct3D path.

Actual behaviour at TARGET

The plan records row 12 as "3 of 4": Gate C is recorded as passing all 14 rows (15 cases) on OPENGLES3, VULKAN and DIRECTX11 (under Xvfb, Wine and DXVK) against each renderer's own goldens. The fourth renderer is not qualified. The candidate the plan tried was not byte-deterministic across a full-corpus sequence and has since been retired (removed-renderers.md, 2026-09-17). The plan's own later re-assessment names SOFTWARE as the cheapest remaining candidate, with whole-corpus tolerance-0 image evidence at 148 dispositions, blocked by the harness's two external inputs (a built Khronos glTF-Sample-Renderer and a sparse glTF-Sample-Assets checkout) that were absent on the recording host; the row's own sentence that SOFTWARE remains unusable is stale against that. gltf-viewer-retake-report.json records only the OPENGLES3 run (2026-08-14, CNA commit d3c5bca6); the VULKAN and DIRECTX11 passes exist only as prose in plan_gltf.md.

Source locations

• plans/plan_gltf.md — section 27.2 row 12 and GLTF-459
• docs/gltf-viewer-retake-report.json — the recorded Gate C run and its pins

Evidence

Checked by reading at 009d40f5; not executed.

Independent re-verification: Checked by reading at 009d40f5; not executed. The three-renderer pass is CNA's own recorded result. The committed report was written once, on 2026-08-14, and never updated for the later VULKAN and DIRECTX11 runs.

Focused reproduction

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

Current tests

The Gate C record above; no CI job reruns it.

Regression test

A fourth renderer's Gate C run with byte-identical two-process captures, recorded in the report.

Blast radius

Evidence only: real-world asset parity is established for three renderers.

Workaround

No workaround is known.