LR-PROF-CAMERACOLOR: accept ARW camera-look fits up to 0.10 held-out RMS

The 0.055 cap rejected half of the public Sony samples tried (ILCE-6400,
ILCE-6700, ILCE-7M4) although their fits were 1.5-3x closer to the
camera JPEG than the fallback and visibly better (mean dE vs the embedded
preview 19-26 -> 5-10). Unrelated previews stay rejected (synthetic test
lands at 0.115).

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

crates/engine/src/camera_preview.rs

@@ -43,6 +43,14 @@ pub(crate) fn fit_preview(raw: &RawImage, bytes: &[u8], transform: &CameraTransf
     Some(look)
 }
 
+/// A fit must cut the held-out squared error to below this share of the fallback's.
+const MIN_IMPROVEMENT: f64 = 0.7;
+/// ...and stay within this per-channel RMS of the camera JPEG (linear display values). On public
+/// raw.pixls.us samples (eight Sony bodies) good fits that still beat the fallback 1.5–3× landed
+/// at 0.065–0.093 (camera local tone, vignetting and lens processing that a global matrix + curve
+/// cannot follow): visibly better renders that 0.055 rejected.
+const MAX_HOLDOUT_RMS: f64 = 0.10;
+
 fn luma(p: [f64; 3]) -> f64 {
     p[0] * 0.2627 + p[1] * 0.6780 + p[2] * 0.0593
 }
@@ -131,7 +139,7 @@ fn fit_pairs(sensor: &Rgb32f, reference: &Rgb32f) -> Option<CameraLook> {
     if lightcraft_pipeline::profiling() {
         eprintln!("[profile] ARW holdout RMS {:.5} -> {:.5} ({samples} channels)", (before / samples as f64).sqrt(), (after / samples as f64).sqrt());
     }
-    if !after.is_finite() || samples == 0 || after >= before * 0.7 || after / samples as f64 > 0.055f64.powi(2) {
+    if !after.is_finite() || samples == 0 || after >= before * MIN_IMPROVEMENT || after / samples as f64 > MAX_HOLDOUT_RMS.powi(2) {
         return None;
     }
     Some(CameraLook { matrix, tone: curve })
@@ -216,6 +224,30 @@ mod tests {
         assert!(luma(p) > 1.0);
         assert!(tone.apply(0.2) < tone.apply(0.4));
     }
+    #[test]
+    fn accepts_a_much_better_fit_despite_local_camera_processing() {
+        // The camera JPEG departs from any global matrix + curve (local tone, vignetting): ±0.12
+        // per-pixel deviations, ~0.07 RMS. The fit is still far closer than the fallback.
+        let known = Mat3([[1.8, -0.4, -0.1], [-0.2, 1.5, -0.1], [-0.05, -0.3, 1.7]]);
+        let mut sensor = Rgb32f::new(64, 64);
+        let mut reference = sensor.clone();
+        let mut seed = 0x2545_f491_u32;
+        for (i, (src, dst)) in sensor.data.iter_mut().zip(&mut reference.data).enumerate() {
+            let ev = 0.05 + (i % 31) as f32 * 0.017;
+            *src = [ev * (0.8 + (i % 11) as f32 * 0.025), ev, ev * (0.8 + (i % 17) as f32 * 0.014)];
+            let p = known.apply_f32(*src);
+            let y = luminance_2020(p);
+            *dst = p.map(|v| {
+                seed ^= seed << 13;
+                seed ^= seed >> 17;
+                seed ^= seed << 5;
+                let noise = (seed % 2001) as f32 / 1000.0 - 1.0;
+                (v * (1.0 - (-2.5 * y).exp()) / y + 0.12 * noise).clamp(0.005, 0.97)
+            });
+        }
+        assert!(fit_pairs(&sensor, &reference).is_some());
+    }
+
     #[test]
     fn rejects_monochrome_invalid_and_unrelated_previews() {
         let mut sensor = Rgb32f::new(32, 32);

docs/camera-preview-colour.md

@@ -10,7 +10,7 @@ Two thirds of the usable pixels train a ridge-regularised 3×3 **chromaticity**
 
 The matrix is applied to scene-linear sensor pixels. The curve is carried in `SourceInfo` and applied in the pipeline finish stage, after exposure/WB/local edits. It is never baked into decoded luminance or clipped to JPEG precision. CPU and GPU use the same tone LUT. Contrast, whites and blacks adjust the starting look through the display-domain controls. JPEG neutral rendering and the existing DNG colour model are unchanged.
 
-One third of the pixels are held out from both fits. The combined model must reduce linear-display squared error by at least 30% and have per-channel RMS error ≤0.055. These are provisional acceptance gates, not a colour-accuracy certification. Monochrome, singular, unrelated, aspect-mismatched and poor-fitting references retain the documented fallback. Both fits use the same fixed proxy at every decode resolution.
+One third of the pixels are held out from both fits. The combined model must reduce linear-display squared error by at least 30% and have per-channel RMS error ≤0.10 (0.055 rejected half of the public raw.pixls.us samples tried — ILCE-6400, -6700, -7M4 — whose fits were still 1.5–3× closer to the camera JPEG than the fallback and visibly better). These are provisional acceptance gates, not a colour-accuracy certification. Monochrome, singular, unrelated, aspect-mismatched and poor-fitting references retain the documented fallback. Both fits use the same fixed proxy at every decode resolution.
 
 ## Source metadata and white balance