GPU tests: drop the edge-pixel tolerance; the coverage mask makes edges match

#207 let up to two edge pixels differ because the GPU's f32 framing could
fall on the other side of the lens-warp boundary. 'GPU: decide warped-geometry
edges with a CPU coverage mask' (this PR) makes the GPU use the CPU's f64
decision, so every case is back within the plain MAX_LSB check (embedded
lens + perspective: max 1 LSB, was 91 before the mask).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

crates/gpu/tests/equivalence.rs

@@ -73,23 +73,7 @@ fn check(name: &str, src: &Arc<Rgb32f>, info: &SourceInfo, s: &DevelopSettings,
     let gpu = lightcraft_gpu::render(src, info, s, req, None).expect("gpu render");
     let (mean, max, over) = diff(&cpu, &gpu.image);
     eprintln!("{name:<28} {}x{}  mean {mean:.4}  max {max}  >1: {:.4}%", cpu.width, cpu.height, over * 100.0);
-    // A pixel on the very edge of the picture can fall either side of the sampling boundary (the
-    // lens warp with a shifted red plane lands exactly outside it): the CPU and a GPU's rounding
-    // may then disagree on one edge pixel. Tolerate a couple of those, nothing in the interior.
-    let (w, h) = (cpu.width, cpu.height);
-    let strays: Vec<(usize, usize)> = cpu
-        .data
-        .iter()
-        .zip(&gpu.image.data)
-        .enumerate()
-        .filter(|(_, (p, q))| (0..3).any(|c| p[c].abs_diff(q[c]) > MAX_LSB))
-        .map(|(i, _)| (i % w, i / w))
-        .collect();
-    let on_the_edge = |&(x, y): &(usize, usize)| x == 0 || y == 0 || x + 1 == w || y + 1 == h;
-    assert!(
-        mean < MEAN_LSB && (max <= MAX_LSB || (strays.len() <= 2 && strays.iter().all(on_the_edge))),
-        "{name}: mean {mean:.4} LSB, max {max} LSB, pixels over {MAX_LSB}: {strays:?}"
-    );
+    assert!(mean < MEAN_LSB && max <= MAX_LSB, "{name}: mean {mean:.4} LSB, max {max} LSB");
     (mean, max)
 }