vello_gpu: fix Extend::Pad with bilinear/bicubic sampling

[nicoburns]

Bug noticed when smoke testing AnyRender upgrade to vello_gpu 0.3.

Fix entirely generated by LLM (but it's very small).

Summary

If an image paint uses Extend::Pad with ImageQuality::Medium, vello_gpu fills everything past the image's right and bottom edges with one flat colour: an exact 50/50 blend of the last two texels. vello_cpu repeats the edge texels, which is the expected output.

vello_cpu (expected)	vello_gpu before	vello_gpu after

(8×8 checkerboard, scaled 8×, filled into a rect 4×3 times the image size.)

Cause: extend_mode clamped the Pad position to [0, size - 1] before the bilinear sampler subtracted the half-texel offset. For an 8-wide image, every point past the edge became 7.0 - 0.5 = 6.5, so the sampler blended texels 6 and 7 at fract = 0.5. Bicubic had the same problem.

Fix: external_bilinear_sample and external_bicubic_sample already clamp each tap to the image bounds, which matches what vello_cpu does with per-tap extend_mode. So Pad no longer clamps the position at all, and the only path that relied on that clamp, the nearest-neighbour textureLoad, now clamps to the image bounds itself:

 // helpers/extend.wesl
 case EXTEND_PAD: {
-    return clamp(t, 0.0, max - 1.0);
+    return t;
 }
 // helpers/external_texture.wesl, sample_external_image
-return textureLoad(external_texture, vec2<u32>(coords), 0);
+return textureLoad(external_texture, vec2<u32>(clamp(coords, image_offset, image_max)), 0);

Tests:

• New image_bilinear_pad and image_bicubic_pad snapshot tests: a 4×4 checkerboard on a 100×100 canvas, with snapshots of 368 B and 1 KB. Both fail on GPU before this change and pass after it.
• The GPU-only snapshots external_texture_skewed and external_texture_with_scene_transform were regenerated. Only their right and bottom edge pixels change, by at most 4/255, because of the half-texel fix at those edges.

[LaurenzV]

Is this not fixed by #1963?

[LaurenzV]

I think it should be. It will still take a bit but I hope I can land it next week! 😄

[nicoburns]

Is this not fixed by #1963?

Might well be. Bug is present in 0.3. I didn't check for open PRs.

[LaurenzV]

@nicoburns Could you check whether this works fine on newest main?