nx_ta_transformer_gradcheck_gate.nx

nx_ta_transformer_gradcheck_gate.nx -- LIAR-KILL for the 9 f32 transformer ops added to nx_autograd_tensor (2026-07-09, for the R3e neural reader). Each backward identity was HAND-PORTED from its Q16 nfa_* twin; a transposed index or flipped sign would train the reader to garbage and waste the slow f32 run. This gate verifies every one by FINITE DIFFERENCE: build op -> scalar loss (mse-vs-0, or the CE itself for softce) -> analytic grad (ta_backward) vs numeric grad ((L(+h)-L(-h))/2h). A wrong identity => order-of-magnitude/sign mismatch. Requires each op to show a NON-TRIVIAL gradient (|g|>=20 milli) so it cannot pass at 0==0. PASS <op> for each of: matmul, matmul_nt, cmul, softmax_rows(causal), rope_tab, hadamard, silu, rmsnorm_rows, softce_rows. expect_exit: 0 license_tier: ORIGINAL Sovereign: nx_autograd_tensor + nx_syscalls.