raw (910 rows): every sample, including weak/unsure labels.
binary (459 rows): subset where binary_strict ∈ {positive, negative} AND
the task was successfully migrated to a runnable task directory. This is
the canonical 5-fold CV evaluation set used by quality-gate experiments.
Excludes 51 binary-labeled rows whose v2 migration failed.
