[yxshee]

Changes proposed in this pull request:

• When convert() is given a matrix, the transparency color in info is now calculated the same way as the pixels, m[0] * r + m[1] * g + m[2] * b + m[3] + 0.5. The Python helper had m[3] * 0.5, so the offset was halved and the rounding term was lost, and the transparent color no longer matched the pixel it was meant to mark.

from PIL import Image

im = Image.new("RGB", (1, 1), (255, 0, 0))
im.info["transparency"] = (255, 0, 0)
out = im.convert("L", (1, 0, 0, -10))
print(out.getpixel((0, 0)), out.info["transparency"])  # 245 250, expected 245 245

With (0.5, 0.5, 0, 0) the pixel is 128 and the transparency was 127.

• Added test_matrix_transparency, which checks that the transparency equals the converted pixel. On main it fails with assert 127 == 128 and assert (123, 61, 100) == (123, 62, 101).
• test_matrix_xyz expected (105, 54, 4), but the C conversion turns (255, 0, 0) into (105, 54, 5) with that matrix, so the expected value is updated to match.

The full test suite passes locally, apart from test_grab, which also fails on main here because of macOS screen capture permissions. pre-commit and tox -e mypy pass.

[radarhere]

You're pointing out that

Pillow/src/libImaging/Matrix.c

Line 52 in a26b63c

float v = m[0] * in[0] + m[1] * in[1] + m[2] * in[2] + m[3] + 0.5;

and

Pillow/src/PIL/Image.py

Line 1098 in a26b63c

value = m[0] * v[0] + m[1] * v[1] + m[2] * v[2] + m[3] * 0.5

disagree.

The C code has been there since PIL. The Python code was added in #3205 by me. I am inclined to believe that I made a mistake.