[refactor] avoid Kelvin/mired round-trip in DNG solver (#273)

[thc1006]

TL;DR

XYZ_to_color_temperature interpolates in mired then converts to Kelvin and
clamps to [2000, 50000]. Its only in-tree caller, find_camera_to_XYZ_matrix,
divides that Kelvin back into mired immediately, producing a redundant
mired -> Kelvin -> mired round-trip flagged by #273.

This PR adds a private helper XYZ_to_mired returning the unclamped mired
value, refactors XYZ_to_color_temperature into a wrapper that preserves the
existing Kelvin signature and clamp, and switches the caller to use the mired
helper directly. Public contract and existing test golden numbers unchanged.

Resolves #273.

What changes

• src/rawtoaces_core/rawtoaces_core_priv.h: add XYZ_to_mired declaration.
• src/rawtoaces_core/rawtoaces_core.cpp: split the existing function into
  XYZ_to_mired (the Robertson interpolation, unclamped) plus a two-line
  XYZ_to_color_temperature wrapper that does mired_to_kelvin and the
  [2000, 50000] K clamp. Switch find_camera_to_XYZ_matrix to call the
  mired helper directly.
• tests/testDNGIdt.cpp: add testIDT_XYZToMired pinning the mired return
  for the same XYZ input as the existing Kelvin test, with a cross-check that
  the wrapper agrees with clamp(mired_to_kelvin(mired), 2000, 50000).

3 files, +48/-12.

Why this shape (and not a rename)

The wrapper is preserved on purpose:

• The Kelvin clamp [2000, 50000] corresponds exactly to mired anchors
  robertson_mired_table[2] = 20 and robertson_mired_table[26] = 500, and
  1e6/2000 = 500.0 and 1e6/50000 = 20.0 are exact double-precision
  reciprocals. So the wrapper preserves observable behaviour bit-for-bit.
• All existing testIDT_XYZToColorTemperature* golden numbers continue to
  verify with the same 1e-5 tolerance; no test re-derivation was needed.
• The Kelvin spelling stays available for any consumer that wants
  user-reportable CCT (EXIF / DNG metadata / light_source_to_color_temp).

If a hard rename to a mired-only API is preferred, it is a one-step
deprecation of the wrapper on top of this PR.

Verification

Built and tested inside aswf/ci-rawtoaces:2026.2 (the same image the VFX
2026 CI lane uses), C++20, conan toolchain, RTA_ENABLE_LENSFUN=ON,
ENABLE_SHARED=ON.

• Release / gcc 14.2: full ctest 17/17 pass (baseline 17/17, post-change 17/17).
• Release / clang 19.1: full ctest 17/17 pass.
• cmake --install + tests/config_tests ctest 2/2 pass.
• RTA_SANITISER_MODE=address Debug build: 15/15 pass for tests not
  pre-affected by an unrelated lensfun ASan issue
  (Test_LensCorrection, Test_Python_ImageConverter::test_conversion).
  That issue reproduces on main without these changes; trace is
  lfModifier::AddColorCallback allocating 32 B vs ~lfModifier deleting
  24 B inside liblensfun.so.1 (RHEL/Rocky lensfun-0.3.2-15.el8),
  with no rawtoaces frame other than the lens-correction call site.
• clang-format --dry-run --Werror clean over the entire C++ tree using
  clang-format 16.0.6 (the version pinned by clang-format-check.yml).

Out of scope

• The inverse color_temperature_to_XYZ has the same Kelvin/mired
  asymmetry. Happy to file a follow-up to symmetrise it.
• The unrelated CAT target concern raised in [refactor] DNG solver #267 review is untouched.

DCO

Signed-off-by: thc1006 <84045975+thc1006@users.noreply.github.com>

[linux-foundation-easycla]

The committers listed above are authorized under a signed CLA.

• ✅ login: thc1006 / name: thc1006 (d2b457f)

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.