Descent of the simplified residual under finite-precision initialization
Determine whether the simplified residual direction used by hardware, obtained by subtracting the finite-precision ZF residual error from the true residual, is a descending direction for the conditioned quadratic recovery surrogate; specifically, establish when Re{(r_true^(0))^H M^{-1} r_impl^(0)} > 0 holds.
References
If hardware instead uses the simplified residual \mathbf{r}{\rm impl}{(0)}=\mathbf{r}{\rm true}{(0)}-\mathbf{e}_0, then, with \mathbf{M}=\operatorname{diag}(\mathbf{A}), its direction is descending only when \operatorname{Re}{(\mathbf{r}{\rm true}{(0)})H\mathbf{M}{-1}\mathbf{r}{\rm impl}{(0)}}>0; this condition is not established by the present synthesis data.
Polar correction removes cancellation in the main term but is not a guarantee: in some parameter states the residual dominates, and how often training visits them is open.