fix(#2520): read the spectral Absolute flag, and set all three emission observer scales

[colourbill-ctrl]

Fixes #2520.

Three defects in IccProfLib/IccMpeSpectral.cpp, all from the 2015 import (1f0a9dd), all feeding the multiply in CIccMpeSpectralObserver::Apply(). The first two are the ones #2520 reports. The third turned up in review of the fix: without it, the fix would make things worse for copied profiles.

Change

site	before	after
CIccMpeReflectanceCLUT::Begin(), CIccMpeSpectralObserver::Apply(), CIccMpeReflectanceObserver::Begin()	m_flags & icRelativeSpectralData	m_flags & icAbsoluteSpectralData
CIccMpeEmissionObserver::Begin()	m_xyzscale[0] assigned three times	[0], [1], [2]
CIccMpeSpectralObserver default ctor, copy ctor, copyData()	m_flags never set or copied	set to 0, copied
CIccMpeReflectanceCLUT::Begin() local xyzscale[3]	uninitialized	{1, 1, 1}

1. Absolute flag. icRelativeSpectralData is 0 (icProfileHeader.h:1189), so all three tests were false. ICC.2:2023 §11.2.12 and §11.2.13 apply the media white adaptation (Formulae (43) to (46), (51) to (54)) only "when the Absolute flag is zero". icAbsoluteSpectralData (bit 0) was referenced nowhere in the tree.
2. Emission scale. No constructor initializes m_xyzscale, so Apply() scaled Y and Z by leftover memory on every pixel.
3. Observer flags. An observer built in code, and every copy of one, carried leftover memory as m_flags. CIccCmm::AddXform(CIccProfile&) and CIccXform::Create(CIccProfile&) copy the profile, and the MPE tag copies each element with NewCopy(). On master only the Lab bit was exposed. Once (1) makes bit 0 live, a copied Flags=0 profile could turn absolute. The spectral CLUT constructors already set and copy the field (:825, :868, :911).
4. Initializer. Once the flag test is no longer constant, GCC 15.2 cannot connect the !bUseAbsolute that fills xyzscale to the one that reads it. Under the ci-pr-gcc15 configuration it fails the LTO link with -Werror=maybe-uninitialized, measured in CI's image. 1.0 means "no adaptation", so neither path changes.

Test: iccdev.spectral-absolute-flag

A custom 81-step PCC with a flat illuminant and a sloped observer (x rises, y flat, z falls), and a sloped white, so the media white is a different colour from the illuminant white. With the flat observer of the #1671 harness every spectrum has the same chromaticity, and relative and absolute output could differ only by a scale. Input spectrum = white.

check	asserts
rclt, robs, Flags=0	media white → illuminant white, full XYZ (0.5 1.0 0.5)
rclt, robs, Flags=1	media white keeps its own chromaticity, X/Y and Z/Y only (0.585417, 0.414583)
eobs, Flags=0	white emission → Y = 1 with the observer's chromaticity (Formula (26)); m_xyzscale is poisoned to NaN before Begin()
robs built in code, flags never set	relative, as Flags=0
robs copied with NewCopy() / operator= from Flags=1	the copy is absolute
fixture	media white and illuminant white differ by > 0.05 in both ratios, so the absolute checks cannot pass vacuously

Red/green: each check fails against its own defect

ASan+UBSan Debug (clang 18), detect_leaks=1. The test returns normally, so LeakSanitizer runs.

library	ok	FAIL	failing
master 74579c4c	12	6	rclt abs, robs abs (X/Y 0.5 Z/Y 0.5); eobs (Y nan); flags never set (100 0 0, a Lab result); NewCopy(); operator=
flag tests reverted	14	4	rclt abs, robs abs, both copies
m_xyzscale reverted	17	1	eobs
m_flags init/copy reverted	15	3	flags never set, NewCopy(), operator=
this branch	18	0	—

With the m_flags fix reverted in a Release build (no ASan malloc fill), 2 fail: NewCopy() and operator=. "Flags never set" passed there because the reused memory happened to be zero. Only the operator= case is deterministic on every build. The other two read leftover memory, which ASan's fill makes wrong.

What this does not change

• Relative (Flags=0) output. Master passes every Flags=0 check. k appears in both kCr and m_w = kCw, so it cancels out of Formulae (43) to (46).
• Absolute magnitude. iccDEV's absolute branch normalizes the observer against the illuminant (getEmissiveObserver()), so a perfect reflector has Y = 1. Formulae (41) and (50) print k = 1 / Σ c_y,i·w_i. This PR re-enables the branch and does not rule on which k is intended. The test checks only ratios, so it will not need changing either way.
• Emission observer Absolute. getEmissiveObserver() always applies the relative k of Formula (26), so the flag still has no effect on eobs.
• Corpus. No tracked profile applies a reflectance or observer element. The only tracked eobs/robs are two parse-rejection fixtures in .github/ci/test-data/.

Pre-flight

lane	result
ASan+UBSan Debug, clang 18, detect_leaks=1	red/green above; 0 sanitizer reports, 0 leaks
Release, clang 18, full ctest -j4	265 / 266. The one failure is iccdev.spectral-tiff-preview: ModuleNotFoundError: No module named 'imagecodecs' on this machine, raised before any iccDEV code runs
Strict clang 21.1.3, -Wall -Wextra -Wpedantic -Werror	strict warnings ENABLED, 0 warnings
GCC 15.2.0 strict + LTO, ghcr.io/internationalcolorconsortium/iccdev:latest, ci-pr-gcc15 configure	0 warnings, 0 errors; new test and iccdev.reflectance-observer-illum-range pass
ci-pr-action, ci_scope=source	run 34552413648 at c9e66037: success, 14 jobs green. iccdev.spectral-absolute-flag passed in the job logs of Windows MSVC, Windows ClangCL, MinGW UCRT64 and GCC Strict ASAN+UBSAN Debug (262 / 262)

[xsscx]

2026-09-11 02:06:58 UTC