paper-reprise

Audit a paper-reprise run's numerical fidelity beyond the end2end metric: Part A (paper-internal closed-form properties) applies to from-scratch implementations (impl/, no official repo to diff against); Part B (kernel-composition correctness) applies to ANY run — from-scratch or official-repo — whose eval path composes with an imported/modified kernel (e.g. Quark's MXFP4 quantize/dequantize, or an official repo's own kernel swapped/patched during setup). Triggers on "does this impl actually match the paper", "verify this from-scratch implementation", "why is this claim PARTIAL/FAIL", "is the MXFP4 kernel integration correct", "before I trust this MATCH", "the setup loop patched/replaced a kernel — is that patch correct". Do NOT use for grading a claim's numeric tolerance (that's `paper-reprise grade`, pure code).