Merge PR #583

# Conflicts:
#	crates/engine/src/media.rs

docs/camera-preview-colour.md

@@ -16,7 +16,9 @@ Matrix-only and matrix + table looks are each completed (tone curve, then chroma
 
 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 ≤0.10 (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). When the fit on all pixels fails these gates, it is tried once more, with the same gates, on the pixels away from edges (where the 3×3 neighbourhood in either proxy spans a luminance ratio above 3). Many camera JPEGs are lens-corrected while the raw is not (Sony "Distortion Comp.: Auto", compacts and kit zooms), and on a detailed scene the misaligned edges alone fail the gate although the colours agree everywhere else (issue #232). The public ILCE-7RM2 samples (raw.pixls.us, one façade shot in seven compression/correction variants, 24–70 mm at 51 mm) are up to 15 px of 1440 off their JPEG; the six with distortion correction on missed the gate at held-out RMS 0.108–0.115 and opened in the grey neutral fallback, 13 L* too dark (ΔE76 16.6–16.9 against the JPEG after registering its geometry). Away from edges they pass at 0.049–0.052 and land at ΔE 3.7–3.8 with lightness within 0.6 L*, like the variant shot with distortion correction off (0.059 on all pixels, ΔE 3.9). The corpus GX80 kit-zoom sample (up to 63 px of 1440 off) passes too: 0.107 → 0.081, ΔE 19.1 → 7.9. Fits accepted on all pixels are unchanged. 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). When the fit on all pixels fails these gates, it is tried once more, with the same gates, on the pixels away from edges (where the 3×3 neighbourhood in either proxy spans a luminance ratio above 3). Many camera JPEGs are lens-corrected while the raw is not (Sony "Distortion Comp.: Auto", compacts and kit zooms), and on a detailed scene the misaligned edges alone fail the gate although the colours agree everywhere else (issue #232). The public ILCE-7RM2 samples (raw.pixls.us, one façade shot in seven compression/correction variants, 24–70 mm at 51 mm) are up to 15 px of 1440 off their JPEG; the six with distortion correction on missed the gate at held-out RMS 0.108–0.115 and opened in the grey neutral fallback, 13 L* too dark (ΔE76 16.6–16.9 against the JPEG after registering its geometry). Away from edges they pass at 0.049–0.052 and land at ΔE 3.7–3.8 with lightness within 0.6 L*, like the variant shot with distortion correction off (0.059 on all pixels, ΔE 3.9). The corpus GX80 kit-zoom sample (up to 63 px of 1440 off) passes too: 0.107 → 0.081, ΔE 19.1 → 7.9. Fits accepted on all pixels are unchanged. These are provisional acceptance gates, not a colour-accuracy certification. Monochrome, singular, unrelated and aspect-mismatched references retain the documented fallback; poor-fitting ones get the partial look below when it is clearly closer to the camera JPEG. Both fits use the same fixed proxy at every decode resolution.
+
+**A partial look when the full look misses the gates.** The neutral fallback (camera RGB taken as sRGB, the default tone map) is far from the camera JPEG. Rendered for every photo on this path in the CC0 archive (1,150 files with a camera JPEG), it is 10.7 L\* too dark (median; about one stop), keeps 0.36 of the camera's colourfulness and has hues 7° off on the coloured pixels (median CIEDE2000 13.6). Its white balance is right: on the pixels the camera renders near-neutral the cast is under 1 a\*/b\* (median, every maker). The "blue" fallback on newer Sony bodies (Δb\* −22 over the whole frame on the ILCE-7RM5 sample) is a warm scene losing its colour, not a cast: its near-neutral pixels are within −1.6 b\*. So before falling back, a partial look is fitted (`fit_partial`): the quantile tone curve (registration-free, the largest part of the error), the chroma curve, and, when the scene has enough colour to learn a matrix (the full look's condition), the photo's own colour matrix at half weight, the other half the identity (`PARTIAL_MATRIX_WEIGHT`). No hue/saturation table. Half weight because, on the held-out proxy pixels of the 23 colourful photos without a look, the full matrix left 2 of them more than 2° further from the camera's hues than the neutral fallback and half weight none (median hue error 9.6° with the generic colour, 5.1° at half weight, 4.2° at full weight; on the 1,080 photos with a look 3.9° and 2.7°). Without a matrix (tone and chroma curve alone, or with one global saturation scale) about a third of all photos ended up more than 2° further from the camera's hues than the fallback, as the added colour makes the generic matrix's hue errors visible. The partial look is tried on all pixels, then away from edges, and used only when it cuts the held-out squared error below 0.7 of the neutral fallback's in linear and in gamma-encoded display values, and when the JPEG's luminance ranks the pixels like the scene's (Spearman ≥ 0.5, `PARTIAL_MIN_RANK_CORRELATION`): a curve matched to the brightness of an unrelated picture also beats the dark fallback pixel for pixel. The accepted full looks have 0.55 or more; three old Coolpix files whose raws decode wrongly have 0.07–0.27 and keep the fallback. There is no limit on the partial look's own error. On the archive (1,783 renders): all 1,087 full looks are byte-identical; of the 55 photos that reach the partial fit, 49 get it (40 on all pixels, 9 away from edges; 13 with the damped matrix, 41 with a chroma curve) and 6 keep the fallback (lens-corrected compacts: RX10, TZ80 1:1, D-LUX 4, G10 16:9, GF2 ×2). On the 49, median CIEDE2000 to the camera JPEG 14.0 → 2.1, ΔL\* −11.5 → +0.6, colourfulness 0.36 → 0.87, hue error 10.1° → 7.1°; every one is closer, by at least 1.3. Most of the others without a look never reach a fit: on 48 files (Nikon D4, Df, D300/D300S, D7000, D5100 and others) the camera JPEG proxy comes out a pixel narrower than the sensor proxy and no pairs are collected (the resize in PR #531 fixes this; with it 46 of them pass the full look's gates, CIEDE2000 17.9 → 2.9).
 
 ## Source metadata and white balance