fix(verifier): fail closed and isolate paper splits

[enzo-adami]

This is a focused extension of your NousResearch#150 branch, prompted by two falsifiable admission failures reproduced on the current head.

Changes

• force the composite admission score to 0.0 when objective correctness is zero; a wrong but well-formed answer can no longer retain roughly half credit
• allocate papers, not individual tasks, to train/validation/holdout before task sampling so the same paper cannot leak across splits
• preserve deterministic sampling and requested dataset sizes
• bring the three touched verifier files up to the current Ruff contract

Validation

• full repository: 187 passed
• Ruff on all touched files: PASS
• git diff --check: PASS

This intentionally does not add another statistical-power implementation. NousResearch#162 already supplies the stronger paired significance/power foundation and should remain the single owner of that layer.

If you prefer, the functional commit is 2041d57; the second commit is lint-only.