macOS native parity + GPU tone curve shader fix

CLAUDE.md

@@ -60,20 +60,20 @@ gamma-encoded. If you touch either function, keep them in sync.
 
 ## Native RAW processor build paths
 
-The canonical C source is `lib/ffi/raw/raw_processor_common.{c,h}`. Three
-consumers include/compile it:
+The canonical C source is `lib/ffi/raw/raw_processor_common.{c,h}`. All four
+consumers compile a tiny wrapper that `#include`s it:
 
-- Production Linux build (CMake) — `linux/raw_processor/raw_processor_wrapper.c`
-  which `#include`s `../../lib/ffi/raw/raw_processor_common.c`.
+- Production Linux build (CMake) — `linux/raw_processor/raw_processor_wrapper.c`.
 - Test libraries (`scripts/build_test_libs.sh`) — same wrapper.
 - Flatpak (`dev.myyc.aks.yaml`) — same wrapper.
-- macOS (`macos/raw_processor/raw_processor.c`) — **standalone copy**, not
-  wrapped. If you change the common source, mirror it here too or convert
-  macOS to use a wrapper the same way.
+- macOS (`macos/raw_processor/raw_processor_wrapper.c`) — same wrapper.
+  `macos/build.sh` and `macos/Makefile` pass `-I../lib/ffi/raw` so the
+  canonical header is picked up; Xcode's "Build Native Libraries" phase
+  lists the wrapper + canonical `.c`/`.h` as input dependencies.
 
-All three Linux paths were consolidated onto the wrapper in the
-`refactor/cleanup-and-imagestate` branch; prior to that, stale duplicates
-at `linux/raw_processor/raw_processor.c` kept drifting.
+Consolidation history: Linux paths were unified on the
+`refactor/cleanup-and-imagestate` branch; macOS was migrated on the
+follow-up `followup/macos-parity-and-tone-curve-test` branch.
 
 ## Histogram
 

linux/vulkan_processor/shaders/image_process.comp

@@ -81,21 +81,22 @@ vec3 applyToneCurves(vec3 color) {
     if (params.toneCurveEnabled == 0.0) {
         return color;
     }
-
-    // Convert to 0-255 range
-    ivec3 indices = ivec3(clamp(color * 255.0, 0.0, 255.0));
-
-    // Apply RGB master curve first
-    indices.r = int(getLutValue(rgbLut.data, uint(indices.r)));
-    indices.g = int(getLutValue(rgbLut.data, uint(indices.g)));
-    indices.b = int(getLutValue(rgbLut.data, uint(indices.b)));
-
-    // Then apply individual channel curves
+
+    // Round (not truncate) when sampling the 256-entry byte LUTs so
+    // colour 0.50196 (which is byte 128) hits index 128, not 127.
+    ivec3 indices = ivec3(clamp(color * 255.0 + 0.5, 0.0, 255.0));
+
+    // Must match the CPU order (cpu_processor.dart:_applyToneCurve):
+    // per-channel first, then master RGB. The two orders are not
+    // commutative for non-identity curves.
     indices.r = int(getLutValue(redLut.data, uint(indices.r)));
     indices.g = int(getLutValue(greenLut.data, uint(indices.g)));
     indices.b = int(getLutValue(blueLut.data, uint(indices.b)));
-
-    // Convert back to 0-1 range
+
+    indices.r = int(getLutValue(rgbLut.data, uint(indices.r)));
+    indices.g = int(getLutValue(rgbLut.data, uint(indices.g)));
+    indices.b = int(getLutValue(rgbLut.data, uint(indices.b)));
+
     return vec3(indices) / 255.0;
 }
 

macos/Makefile

@@ -10,9 +10,10 @@ BUILD_DIR = .
 # Targets
 all: libraw_processor.dylib
 
-libraw_processor.dylib: raw_processor/raw_processor.c
+libraw_processor.dylib: raw_processor/raw_processor_wrapper.c ../lib/ffi/raw/raw_processor_common.c
 	$(CC) $(CFLAGS) -o $(BUILD_DIR)/libraw_processor.dylib \
-		raw_processor/raw_processor.c \
+		raw_processor/raw_processor_wrapper.c \
+		-I../lib/ffi/raw \
 		$(LIBRAW_FLAGS) -lm
 
 clean:

macos/Runner.xcodeproj/project.pbxproj

@@ -301,8 +301,9 @@
 			inputFileListPaths = (
 			);
 			inputPaths = (
-				"$(SRCROOT)/raw_processor/raw_processor.c",
-				"$(SRCROOT)/raw_processor/raw_processor.h",
+				"$(SRCROOT)/raw_processor/raw_processor_wrapper.c",
+				"$(SRCROOT)/../lib/ffi/raw/raw_processor_common.c",
+				"$(SRCROOT)/../lib/ffi/raw/raw_processor_common.h",
 			);
 			name = "Build Native Libraries";
 			outputFileListPaths = (

macos/build.sh

@@ -84,9 +84,13 @@ if [ -n "$LIBRAW_STATIC" ]; then
 
     # Set minimum macOS version to match system libraries
     # Use 12.0 as a reasonable minimum that supports both Intel and Apple Silicon
+    # Source is the wrapper which includes the canonical C file at
+    # ../lib/ffi/raw/raw_processor_common.c — shared across Linux / macOS /
+    # Flatpak so the FFI surface never drifts across platforms.
     clang -shared -fPIC -o libraw_processor.dylib \
         -mmacosx-version-min=12.0 \
-        raw_processor/raw_processor.c \
+        raw_processor/raw_processor_wrapper.c \
+        -I"../lib/ffi/raw" \
         -I"$LIBRAW_INCLUDE" \
         "$LIBRAW_STATIC" \
         $JPEG_LIB \
@@ -99,7 +103,8 @@ else
     echo -e "${YELLOW}Using dynamic linking for libraw (may have dependency issues)${NC}"
     clang -shared -fPIC -o libraw_processor.dylib \
         -mmacosx-version-min=12.0 \
-        raw_processor/raw_processor.c \
+        raw_processor/raw_processor_wrapper.c \
+        -I"../lib/ffi/raw" \
         -I"$LIBRAW_INCLUDE" -L"$LIBRAW_LIB" -lraw -lm \
         -Wl,-rpath,@loader_path
 fi