Loss scale is 512, true gradient norm is 4, and clip threshold is 1. Explain correct processing and the clip-before-unscale error. Use idealized norm arithmetic and ignore numerical epsilon in clipping.

Unscale to norm 4, then clip to norm 1 before the optimizer update. Clipping scaled norm 2,048 to 1 first yields norm 1/512 after unscaling, far smaller than intended. Check scaler and skipped-update behavior under the framework contract.