Put the -1 on the greens along the chroma axis, not across it

The Malvar "R at green in R row" kernel weights the two greens two
sites away along the row at -1 and the pair up and down the column at
+1/2. The shader had the two swapped, in the comment as well as the
code, so the transcription checked against itself. Both sum to zero
and reconstruct a flat patch exactly, which is all the tests fed it.

On an edge the correction at green sites is half strength and the
false colour doubles: 0.375 against 0.19 on a grey step, and a
blue/yellow zipper around every clipped highlight at 1:1. The other
three kernels and the CFA tables were right.

A grey vertical step now runs through the pass; the transposed kernel
fails it at 0.375.

core/dr-gpu/src/demosaic.rs

@@ -1347,6 +1347,53 @@ mod tests {
         }
     }
 
+    #[test]
+    fn a_grey_step_edge_stays_grey() {
+        // A flat patch cannot tell the Malvar kernels from any other set of
+        // weights that sum to zero. An edge can. A grey vertical step, so
+        // every photosite records the same profile, must come back with the
+        // three channels close together on both sides; any spread is false
+        // colour from interpolating across the edge.
+        //
+        // The bound is set by the paper's kernels, which peak at 0.19 here.
+        // With the ±2 terms of the green-site kernels transposed — the bug
+        // this test was written against — the peak is 0.375.
+        let Some(ctx) = ctx() else { return };
+        let d = Demosaicer::new(&ctx).expect("demosaicer");
+
+        let size = 32u32;
+        let white = 16383u16;
+        let mut raw = flat_cfa(CfaPattern::Rggb, size, [0, 0, 0], 0, white);
+        for y in 0..size {
+            for x in size / 2..size {
+                raw.data[(y * size + x) as usize] = white;
+            }
+        }
+
+        let img = d.run(&raw).expect("demosaic");
+        let px = read_rgba(&ctx, &img);
+        let (w, _) = img.size();
+
+        let mut worst = (0.0f32, 0u32, 0u32);
+        for y in 2..size - 2 {
+            for x in 2..size - 2 {
+                let p = px[(y * w + x) as usize];
+                let spread = (p[0] - p[1]).abs().max((p[2] - p[1]).abs());
+                if spread > worst.0 {
+                    worst = (spread, x, y);
+                }
+            }
+        }
+        assert!(
+            worst.0 < 0.25,
+            "false colour of {} at ({}, {}) on a grey edge — the green-site \
+             kernels are interpolating across the edge",
+            worst.0,
+            worst.1,
+            worst.2
+        );
+    }
+
     #[test]
     fn output_is_free_of_nan_and_negatives() {
         // f16 NaN propagates silently through every later stage; a negative

core/dr-gpu/src/shaders/demosaic.wgsl

@@ -160,12 +160,18 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
         // Green is measured. Red and blue are interpolated from their own
         // axis, with a correction from the green Laplacian.
         //
-        // Malvar "G at R/B locations" kernels, transposed per axis:
+        // Malvar "R at green in R row" kernel, and its transpose:
-        //   chroma along the row:  (5c + 4(w1+e1) - (nw+ne+sw+se) - (n2+s2) + 0.5(w2+e2)) / 8
+        //   chroma along the row:  (5c + 4(w1+e1) - (nw+ne+sw+se) - (w2+e2) + 0.5(n2+s2)) / 8
+        //
+        // The -1 goes on the two greens *along* the chroma axis and the +0.5
+        // on the pair across it. Transposed, both kernels still sum to zero
+        // and reconstruct a flat patch exactly, but on an edge the correction
+        // at green sites is half strength and the false colour doubles: a
+        // blue/yellow zipper around every clipped highlight.
         let along_row =
-            (5.0 * c + 4.0 * (w1 + e1) - diag1 - vert2 + 0.5 * horiz2) * 0.125;
+            (5.0 * c + 4.0 * (w1 + e1) - diag1 - horiz2 + 0.5 * vert2) * 0.125;
         let along_col =
-            (5.0 * c + 4.0 * (n1 + s1) - diag1 - horiz2 + 0.5 * vert2) * 0.125;
+            (5.0 * c + 4.0 * (n1 + s1) - diag1 - vert2 + 0.5 * horiz2) * 0.125;
 
         let red_horizontal = red_is_horizontal(gid.x, gid.y);
         let r = select(along_col, along_row, red_horizontal);