Shared noisy evaluator and saved-results audit

All 128 checks pass after moving the short experiment into the shared evaluator used by the next study. The independent saved-results auditor also rejects altered prediction targets, operation counts and covariance normalization.

The same twelve short cases preserve their original interpretation: accurate point forecasts, numerical nonconvergence and unidentified calibration. Repeating them for software qualification creates no new physical evidence.

All 162 artifact hashes, including 153 source snapshots, are verified. No original or protected measurements are used.

Next: the separately frozen two-minute study, with more training observations and the full evaluation ceiling.