Fix: edge visibility bug

[danielzhong]

Description

Problem:

1. GLB files produced by TilesetPublisher contain degenerate triangles — triangles where two or more vertices share the same position, giving them zero area. In buildTriangleAdjacency, the cross product of the two edge vectors for such a triangle is the zero vector. Calling Cartesian3.normalize on it throws:
   Error: normalized result is not a number

These have always existed in the data, but previously went unnoticed because face normal computation was done in the GPU shader, where normalize(vec3(0)) silently returns vec3(0).

#13110 moved face normal computation to the CPU, introducing the first JS call to Cartesian3.normalize on per-triangle cross products. Unlike the GPU, Cartesian3.normalize throws a DeveloperError when given the zero vector, causing the tile to fail to load entirely.

This fix makes Cesium defensively skip degenerate triangles, restoring the previous tolerance for imperfect mesh data regardless of its origin.

2. Cleanup and simplify silhouette normal decoding

The previous implementation decoded GLB signed-byte normals, then immediately re-encoded them to 16-bit oct format, packed two into a Uint32, and later unpacked and decoded them back to floats before writing into the Float32Array vertex buffer. This round-trip (signed byte → oct16 → Uint32 → oct16 → float) introduced quantization loss and unnecessary complexity.

The normals are now decoded once — signed byte → normalized Cartesian3 — and stored directly in a plain JS array. Each edge lookup reads directly from this array. The final vertex buffer (Float32Array) is unchanged. The unused Cartesian2 and AttributeCompression imports are removed.

Issue number and link

https://github.com/CesiumGS/tilers/issues/2156

Testing plan

Author checklist

• I have submitted a Contributor License Agreement
• I have added my name to CONTRIBUTORS.md
• I have updated CHANGES.md with a short summary of my change
• I have added or updated unit tests to ensure consistent code coverage
• I have updated the inline documentation, and included code examples where relevant
• I have performed a self-review of my code

AI acknowledgment

• I used AI to generate content in this PR
• If yes, I have reviewed the AI-generated content before submitting

If yes, I used the following Tools(s) and/or Service(s):
GitHub Copilot

If yes, I used the following Model(s):
Claude 4.6 Sonnect

[github-actions]

Thank you for the pull request, @danielzhong!

✅ We can confirm we have a CLA on file for you.

[javagl]

I don't know how often this is called, and how performance-critical it is. So the following might not be important, but I'll mention it anyhow: Performance purists could claim that the (squared) length is computed twice here now - once for the check, and once for the normalize call. And these purists could suggest to change the normalize call to

multiplyByScalar(... 1.0 / Math.sqrt(crossMagnitudeSquared) ...);

But my gut feeling is that this will not immediately affect actual performance in any way.

[danielzhong]

Great point, modified!

[Copilot]

Pull request overview

This PR hardens EdgeVisibilityPipelineStage against malformed mesh data and simplifies the silhouette-normal decoding path used by EXT_mesh_primitive_edge_visibility, preventing tiles from failing to load when degenerate triangles are present.

Changes:

• Skip zero-area (degenerate) triangles during triangle adjacency/face-normal computation to avoid Cartesian3.normalize throwing.
• Replace the previous silhouette-normal decode/re-encode/pack/unpack pipeline with a single decode-to-float representation.
• Remove unused imports related to the removed encoding/packing path.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[markschlosseratbentley]

LGTM - make sure you mark the checklist items in the PR desc.