Stage 3 (pilot, descriptive, retrospective) — Does lower coefficient-spectrum participation ratio at fixed teacher/student width correlate with faster feature learning?

Retrospective, descriptive/pilot registration (not confirmatory) — frozen before execution of the 96-run pilot grid on 2026-07-11, commit f906f26317eb727f12a7833451fbeddcf97b2eca. An engineering pilot with only 4 outer units, reporting paired signs and effect ranges only (no p-values, no confirmatory claim). Its purpose was to remove a confound found in Stage 2's exploratory analysis by holding exact teacher width constant. Full design and provenance in the Summary field and in the linked Zenodo record.