fix(bpc): estimate black through the tag family the xform actually uses

[maxderhak]

Problem

CIccXform::Create() chooses between the AToBx/BToAx colorimetric tags and the DToBx/BToDx MPE tags, and a caller forces the colorimetric ones by passing bUseD2BxB2DxTags=false to AddXform().

CIccApplyBPC rebuilds its own cmm objects to find the black point, and hardcoded that choice as version >= V5 ? colorimetric : MPE instead of asking the xform it is adjusting:

pICC->m_Header.version >= icVersionNumberV5 ? false : true

For a v4 profile carrying both families this is backwards. The transform applies BToAx while black is measured through BToDx, so the scale and offset BPC installs come from a pipeline that is never applied. These are exactly the profiles where the two pipelines are known to differ — that is why a caller opts out — so the error is not marginal, and it is silent.

Only the B-side is affected. Create()'s input branch gates its DToBx lookup on spectralPCS || version >= V5, so a plain v4 profile resolves device→PCS through AToBx whatever the flag says. The output branch has no such gate and follows the flag alone.

Fix

m_bUseD2BTags already existed on CIccXform but was dead — set false in the constructor and never written since.

• Record it at the SetParams() sites in CIccXform::Create() and CIccXformMpe::Create(), after the icXformLutSpectral and icXformLutColorimetric adjustments have rewritten it, so it names the family the selection code actually walked rather than what the caller asked for.
• Expose UseD2BTags() / SetUseD2BTags() as inline accessors — no layout or vtable change, so ABI is unaffected.
• CalcFactors() already receives the xform, so only pixelXfm() and getBlackXfm() gain a parameter.

The Create() overload that is handed a tag outright keeps the constructor default, with a comment saying why: there is no selection to record, and no way for CIccApplyBPC to reproduce the caller's choice by rebuilding from the profile.

Blast radius

The V5 arm of the ternary is preserved, so only the explicit opt-out changes behaviour and the default path is byte-identical. The CLI tools cannot reach the changed configuration — iccApplyToLink decodes intent 10 + (no D2Bx) and 40 + (BPC) as mutually exclusive switch cases, so BPC there always implies bUseD2BxB2DxTags=true. No existing end-to-end baseline can shift; this is a library-API-only change.

Regression coverage

iccdev.bpc-d2b-tag-family builds a v4.3 CMYK printer profile in code whose BToA0 and BToD0 return different inks (cyan 0.60 vs 0.95) and whose AToB0 turns those into L* 40 and L* 5 — both clearing calcSrcBlackPoint()'s L*≤50 clip as distinct values, since a pair that both clipped would leave the test toothless.

Applying one CMYK pixel with BPC enabled:

	flag true	flag false	BToD0 absent (reference)
before	0.499921	0.499921	0.349680
after	0.499921	0.349680	0.349680

Before the change the flag was inert and black came from BToD0 either way. After, false matches the BToA0-only reference exactly. The test exits 1 before and 0 after.

Verification performed

Configured and built with the repo's own vs2022-x64 preset (VS2022 Professional, MSVC x64 Release, vcpkg toolchain) and exercised through CTest:

• iccdev.bpc-d2b-tag-family registers as test Update Xcode Project Files, Add Unit Test for CVE-2023-44062, Build Scripts for libtiff + libxml2, Fix FPE #67 with labels bpc cmm iccdev regression xform, builds, and passes.
• Reverting only the three ternaries and rebuilding makes it fail under the same toolchain, so the test is a genuine pin on MSVC and not an artefact of one compiler.
• ctest -L "cmm|xform" — 16/16 pass.
• ctest -L regression — 95/95 pass.
• Full non-slow suite — 101/101 pass.

Re-verified after the branch was rebased onto current master (53 newer commits, including the apply-path guard retirement #2307 and the CLUT channel-count bounds #2306): full reconfigure and rebuild, test #90 passes with byte-identical output, and the full non-slow suite is 133/133.

Also cross-checked earlier with g++ against a hand-built IccProfLib, producing byte-identical numbers to the MSVC run.

Left open

Two pre-existing gaps outside this fix's scope:

• pixelXfm() still hardcodes icXformLutColorimetric, which closes the DToBx path on the input side for any profile. Moot for v4 non-spectral, but it means a v5 MPE-only profile cannot resolve a source black point at all.
• The tag-passing Create() overload records nothing, so BPC on an xform built that way still assumes the colorimetric family.

🤖 Generated with Claude Code

[xsscx]

2026-08-25 20:41:22 UTC

[Copilot]

Pull request overview

Updates BPC black-point estimation to use the tag-family preference recorded by the adjusted transform.

Changes:

• Records and exposes the selected DToB/BToD tag-family preference.
• Propagates that preference through BPC helper transforms.
• Adds a focused CTest regression.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
IccProfLib/IccCmm.h	Adds tag-family accessors.
IccProfLib/IccCmm.cpp	Records tag-family selection during transform creation.
IccProfLib/IccApplyBPC.h	Extends BPC helper signatures.
IccProfLib/IccApplyBPC.cpp	Reuses the transform’s tag-family preference.
Build/Cmake/Testing/CMakeLists.txt	Registers the regression test.
.github/ci/regression/bpc-d2b-tag-family.cpp	Adds the synthetic-profile regression fixture.