[22nsuk]

writetrainingdata sets sgfGameWinner to Black for the valid unscored White-win result RE[W+]. This selects the Black-win branch when constructing forced result targets. Recognize White correctly while preserving the existing incomplete-record flags and target weighting.

Factor only the two bare B+/W+ branches into a file-local helper used by the converter and a native regression. The regression reads actual SGF RE properties, checks both colors plus lowercase/whitespace normalization, verifies the forced win/loss/no-result/score targets, and checks that other result forms are left to their existing handlers. It runs in katago runtests, so the existing upstream CI exercises it without a model or workflow changes.

Validation:

• Negative control retaining the old W+ -> Black mapping: the new regression fails at winner == testCase.winner.
• Fixed Eigen Release build: full katago runtests passes, including six accepted result variants and six non-bare-result controls.
• Additional runoutputtests encountered the existing snapshot != NULL assertion at cpp/tests/testsearchnonn.cpp:2657. An unmodified export of base d91ea855 reproduced the same failure (first run passed, second failed). This is the separate terminal-leaf invariant issue covered by Fix visit-cap test invariants for terminal leaves #1258; this PR does not change those tests or claim an overall output-test pass.

Based directly on current upstream master (d91ea855110dae533f0aada947b2b7d78cc8a4e1), with one commit touching only writetrainingdata.cpp, runtests.cpp, and tests.h. This native regression checks result parsing and target construction, not a full SGF-to-NPZ conversion. No fork-specific worker settings, Python harnesses, or CI changes are included.