Read CalGray and CalRGB, and let a JPEG reach its colour space

[tannevaled]

Two defects, and the second is what made the first unreachable.

The spaces were read as their device namesakes

CalGray and CalRGB were mapped straight onto DeviceGray and DeviceRGB. They are not those spaces: each says what CIE XYZ its numbers stand for, under a white point of its own choosing, so reaching sRGB is a gamma, a matrix and a chromatic adaptation.

With the gamma the format defaults to, a CalGray component is linear luminance — a mid grey encodes to 188, not the 128 a device grey gives.

calibrated.go reads the two dictionaries and carries the format defaults; gfx/color v0.22.0 does the arithmetic (go-gfx/gfx#39, measured against poppler at 1 level in 256).

And a JPEG could not reach a colour space anyway

jpegThroughSpace asked the colour space only for a one-component picture in one of four named families:

switch sp.name {
case "Separation", "DeviceN", "Indexed", "Lab":
default:
    return nil
}
if sp.components != 1 { return nil }

So a three-component JPEG in a calibrated space was drawn from the decoder’s own colours and the space was never consulted — the new library would have had no caller. The rule is now what it should have been: a space that is not one of the four device singletons has something to say about the samples, whatever their count. The device spaces are package-level singletons, so identity is the whole test.

What it is worth

openpdf’s PDF 2.0 test file draws the same 1466-byte JPEG stream twice, once through a CalRGB (Adobe RGB matrix, D50, gamma 2.2) and once through DeviceRGB, and its own page text says the two should differ. Against pdfimages:

	before	after
peak	110	20
share of pixels differing	0.5764	0.0098
mse	909.34	0.30

The 20 that remains is not colour

Our conversion applied to poppler’s own decoded samples disagrees with poppler by one level. But a sample error of ±3 — exactly what the direct bucket measures for this same stream — moves the answer by up to 35 levels once a gamma of 2.2 and an Adobe RGB matrix have amplified it.

A calibrated space magnifies a decoder disagreement. That is worth knowing before reading any per-channel bound: the conformance gate is derived for a picture read directly.

Lab is deliberately left alone

Correction, after merge. This section said poppler's GfxLabColorSpace::getXYZ does not multiply by the white point where ISO 32000-2 8.6.5.4 says to. That was wrong, and it was wrong because I read one function instead of running the program: getXYZ indeed does not, but ::getRGB multiplies immediately after calling it (X *= whiteX;).

Settled by experiment: a hand-built four-pixel Lab document, extracted with pdfimages, matches the specification's formula on 4 of 4 pixels within one level and departs from the no-white-point formula by 13 levels on a neutral mid tone. The judge follows the specification, so there is no conflict and labSpace()'s grey stand-in is simply a defect of ours. It is fixed in a follow-up rather than here.

Lab is unchanged in this PR. For scale: 12 of 2268 form documents mention /Lab, and none of the 1013 scans.

Coverage stays at 100%.

🤖 Generated with Claude Code