Build system

• Replaced gyp + bundled Python with a small CMakeLists.txt (header-only INTERFACE library, mapbox::pixelmatch alias, install rules for downstream find_package).
• Removed deps/gyp/, deps/run_gyp, pixelmatch.gyp, the wrapper Makefile, and .travis.yml.
• New GitHub Actions workflow (Ubuntu-only — header-only stdlib code, no platform variance to test).
• Added .clang-format enforcement in CI via clang-format --dry-run --Werror.

Tests

• Tests now decode PNG fixtures directly via a vendored picoPNG (~530-line single header), eliminating the JS/pngjs convert_tests.js step and the committed *.rgba blobs.
• Imported new fixtures (1, 3, 4, 6 from upstream; 2, 5, 7 generated against an upstream branch with checkerboard: false to match our white-blend semantics).

Algorithm port (from JS pixelmatch)

Brought the implementation up to date with ~10 years of upstream improvements, while deliberately keeping:

• White-background blending (skipping the new checkerboard variant).
• The original options surface (no alpha / aaColor / diffColor / diffColorAlt / diffMask).

Performance changes

1. Identical-image fast path — per-row memcmp short-circuits the entire loop when both inputs are byte-equal.
2. Per-pixel 4-byte equality check before invoking colorDelta, so identical pixels skip all YIQ math.
3. colorDelta rewrite — operates on raw channel diffs and only blends with white when α<255. The opaque case (the common one) now skips the blend math entirely. Mathematically equivalent to the old white-blend formula.
4. antialiased rewrite:
   • Center pixel cached outside the 8-neighbor loop (one read instead of nine).
   • Edge-pixel zeroes initialization fix (treats off-image neighbors as implicit zeroes, matching upstream behavior).
   • The recursive call to antialiased(...) for the sibling-confirmation step replaced with a much simpler hasManySiblings doing raw RGBA equality via memcmp. This was the biggest single win.
5. Stride-aware antialiased and hasManySiblings (closes a latent bug where the AA path assumed tightly-packed input even when called via the stride overload).

Correctness changes

• drawGrayPixel now reads from img1 with the stride-aware offset (was using the tightly-packed output offset, broken for stride ≠ width × 4).
• New AA edge-pixel zeroes=1 initialization fixes a small over-flagging on image borders. This is what changes the 4a vs 4b count from 36089 → 36049 (40 pixels at edges previously misclassified as differences).
• All four reused upstream diff fixtures (1, 3, 4, 6) now match byte-for-byte.

Benchmark results

Added a bench_pixelmatch executable that runs the full fixture set with auto-calibrated iteration counts.