Render band test: the comments say what row_stats divides by and what the artifact case measures (#2080)

row_stats sums the differences of the pixels above noise and divides by the whole row's channel count; the doc now says so. The artifact case changes only the colour channels, so its mean is 2.25, which the earlier allowance let through, and its comment now says it fails on the mean while its count stays within a twentieth of the row.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Model-Effort: Opus 5.5 drove and wrote at high; no brief

crates/render/src/band_tests.rs

@@ -61,9 +61,10 @@ fn view(scale: f64) -> View {
 /// it runs level, as at the top of an ellipse), not a band artifact.
 const NOISE: u8 = 24;
 
-/// One row of two renders: the mean channel difference over the pixels that are not noise (see
-/// `NOISE`), and how many pixels differ by more than 2 at all. A band artifact differs along a
-/// whole row, by more than noise.
+/// One row of two renders: the channel differences of the pixels above noise (see `NOISE`) summed
+/// and divided by the whole row's channel count, so noise pixels add nothing but still count in
+/// the denominator; and how many pixels differ by more than 2 at all. A band artifact differs along
+/// a whole row, by more than noise.
 fn row_stats(x: &[u8], y: &[u8]) -> (f64, usize) {
     let mut sum = 0.0;
     let mut off = 0;
@@ -84,8 +85,8 @@ fn row_diffs(a: &Pixmap, b: &Pixmap) -> Vec<(f64, usize)> {
     a.data().chunks(row).zip(b.data().chunks(row)).map(|(x, y)| row_stats(x, y)).collect()
 }
 
-/// Whether a row passes: the original mean limit over non-noise pixels, and at most a twentieth
-/// of the row differing at all.
+/// Whether a row passes: the original 0.25 limit on the mean from `row_stats`, and at most a
+/// twentieth of the row differing at all.
 fn row_ok(mean: f64, off: usize, width: usize) -> bool {
     mean < 0.25 && off * 20 <= width
 }
@@ -94,10 +95,13 @@ fn row_ok(mean: f64, off: usize, width: usize) -> bool {
 fn a_real_band_artifact_fails_but_edge_noise_passes() {
     let width = 880; // scale 2
     let flat = vec![100u8; width * 4];
-    // 44 pixels differing by 60 per channel: under the old mean allowance, but a real artifact.
+    // 44 pixels whose colour channels differ by 60 (alpha unchanged): the mean is
+    // 44 * 3 * 60 / (880 * 4) = 2.25, which the earlier `0.25 + 0.05 * off` allowance (2.45 here)
+    // let through; the count, 44, is within a twentieth of the row. A real artifact, so it fails
+    // on the mean now.
     let mut artifact = flat.clone();
     for px in artifact.chunks_mut(4).take(44) {
-        px.iter_mut().for_each(|c| *c += 60);
+        px.iter_mut().take(3).for_each(|c| *c += 60);
     }
     let (mean, off) = row_stats(&flat, &artifact);
     assert!(!row_ok(mean, off, width), "{mean} {off}");