## macOS parity
Convert the macOS raw processor to the wrapper pattern (same as Linux),
which was the last standalone drift point after the Linux cleanup branch.
Before: `macos/raw_processor/raw_processor.c` was a 250-line standalone
copy ~60 lines out of sync with canonical. Most importantly it lacked
`raw_processor_set_highlight_mode`, so any macOS build after the Linux
refactor would crash at image load when the Dart FFI looked up the
missing symbol.

- Add `macos/raw_processor/raw_processor_wrapper.c` (one-liner that
  `#include`s `../../lib/ffi/raw/raw_processor_common.c`).
- Delete `macos/raw_processor/raw_processor.{c,h}`.
- Update `macos/build.sh` and `macos/Makefile` to compile the wrapper
  and pass `-I../lib/ffi/raw` for the canonical header.
- Update Xcode project input-paths so the "Build Native Libraries" phase
  re-runs on wrapper/canonical changes.
- Update `CLAUDE.md` — macOS no longer needs special mirroring.

## GPU tone curve fixes
Audit of ROADMAP's "GPU tone curve pending" claim: the shader code is
all there, but there were two real bugs that a CPU↔GPU comparison test
would have caught:

1. **Operation order was reversed** from CPU. The shader applied
   master RGB curve first then per-channel curves; CPU does per-channel
   first. Non-commutative for non-identity curves — outputs differed
   by several LSBs.
2. **Float→int truncation** when indexing the byte LUT: sRGB 0.50196
   (byte 128) got truncated to index 127 instead of rounded to 128. Off
   by one at every quantisation boundary.

Both fixed in `linux/vulkan_processor/shaders/image_process.comp`.
Added a CPU↔GPU comparison test with an S-curve + distinct per-channel
curves; CPU and GPU now agree to within 1 LSB (measured: maxDiff 0,
avgDiff 0.00 on the full test fixture).

## Tests
31 native tests pass (was 30), 66 unit tests pass, `dart analyze` clean.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>