metabric: the label is partly derived from the features #151

Problem

The metabric label is the PAM50 + claudin-low subtype, which was assigned from the tumours' expression profiles. Four of its 11 features (BCL2, CCNB1, EGFR, ERBB2) are among the 50 genes of the PAM50 classifier, so the label partly derives from the features rather than being independent of them (docs/reference/datasets/metabric.md, Feature subset). The feature subset follows L2M v2, which gives no rationale.

Scope

• Dataset generation for metabric; hard-budget setting (no method runs it in the soft-budget setting). Not in the kdd26 dataset list.

Options to decide

• Keep the L2M feature set for comparability with L2M and keep documenting the dependence.
• Or drop the PAM50 genes, or choose a label not derived from expression, which departs from L2M.

Acceptance criteria

• The decision is recorded on the dataset page.
• just qa passes.