Stage 2 (retrospective registration) — Does reducing intrinsic task rank K at fixed student capacity C improve held-out generalization?

Retrospective registration (frozen locally before execution via git commit 2991c8c and config SHA-256, registered on OSF after completion as an archival record). Tests whether reducing intrinsic task rank K at fixed student capacity C improves held-out generalization, within the region where every task is expressible (C >= K). Full design, endpoints, and integrity gates are in the Summary field and in the source repository (see the linked Zenodo record for code and raw results).