[msaroufim]

The sort checker accepted invented increasing values because input magnitude depended on the seed and correctness used relative tolerance. At seed 10^12, the old 4096-element generator collapses to one float32 value, while an unrelated sequence spanning ±4,000,000 passes its checker.

Require exact sorted values and float16 addition; generate sort inputs independently of seed magnitude with varied distributions. Reject malformed shapes, dtypes, devices and non-finite results across all eight PMPP v2 tasks. Fix histogram scalar broadcasting and replace prefix-sum tolerance that grew to ~16% with fixed relative tolerance and an input-energy roundoff allowance. Calibrate the scan allowance against legitimate PyTorch scans through the largest benchmark. Bound mismatch diagnostics for large incorrect outputs and preserve reference determinism settings.

Validation: 118 CPU regression tests, Ruff, diff checks and GitHub CI pass. All 32 GPU positive controls pass: every PMPP v2 task on A100, B200, H100 and L4, including every benchmark shape. Fabricated sort, histogram and scan negative controls are rejected. Private historical revalidation also applies the exact sort checker to the original generator, catching input-detecting approximations that fall back to real sorting on the repaired generator. Contestant source is excluded from this PR; leaderboard remediation is tracked separately. All eight production task definitions were synced and their embedded files, correctness cases and benchmarks verified against the merged suite.