Golden-reference regression suite tool

A dependency-free (pure standard-library) Python tool that runs the golden-reference regression SUITE this article describes — distinct from 5.5's single-build gate. It loads a FROZEN baseline (the golden master: per-clip, per-rendition VMAF mean and 5th percentile blessed when the suite was created) and a candidate build, and for each reference clip it (1) compares the candidate to the frozen baseline within a NOISE-SIZED tolerance band computed per clip — margin = z * sqrt(model_sigma^2 + run_