Neutral Cross-Framework Benchmarking Reveals Rz Sign Convention Divergence in Quantum Simulation Backends

We present CleitonForge, a neutral benchmarking layer for quantum circuit simulation implemented in Rust. By routing identical circuits through a canonical intermediate representation to four independent simulation backends (native statevector, quantrs2, roqoqo, q1tsim), we identify a sign convention disagreement in the Rz rotation gate. The divergence produces cross-backend fidelity F=0.0 on QAOA and VQE circuits but is invisible to Clifford benchmarks and the IBM Quantum Volume test. We prove algebraically (Theorem 1) that for circuits built from Rz, Ry, and CX, the sign inversion produces U_wrong = U*_ref, leaving all measurement probabilities exactly identical — making the bug undetectable by any probability-based metric including Quantum Volume. We introduce the Quantum Gate Convention Standard (QGCS) v0.1, a 12-check phase-sensitive conformance suite. Results: native/roqoqo/q1tsim score 12/12 (OpenQASM 3 compliant); quantrs2 scores 10/12. The cforge certify CLI and cf.certify() Python API automate these checks. v2 (2026-07-14): adds QGCS section, algebraic blindness proof, corrected HOG data, Python API. Source: https://github.com/cleitonaugusto/CleitonForge