nx_kda_gate.nx

nx_kda_gate.nx -- two things nx_nofloat_kda_gate does NOT cover: (A) the ALGEBRAIC IDENTITY that licenses the shipped kernel's factored update, and (B) the KDA:MLA interleave schedule, a separate census feature that no organ implemented. (A) WHY IT MATTERS. Kimi Linear (arXiv 2510.26692) states the rule as S_t = (I - beta k k^T) Diag(alpha) S_{t-1} + beta k v^T which literally builds a d_k x d_k matrix: O(d_k^2 * d_v). The shipped kernel (_hdl_build/nx_nofloat_kda.nx) does NOT do that -- it computes pred = k^T S, then S += k (x) beta(v-pred), which is O(d_k * d_v). That is a real optimisation and it is CORRECT only because (I - beta k k^T) is identity-plus-rank-1. nx_nofloat_kda_gate proves the kernel's BEHAVIOUR (overwrite, gating, determinism); nothing proved the factorisation itself. T1 does, over two steps from a NON-ZERO state -- one step from S=0 would agree even for a wrong factorisation. T3 guards against proving the identity on a DEGENERATE case: if k were a basis vector, (I - beta k k^T) would be diagonal and both forms would agree trivially. T3 asserts the off-diagonal term is non-zero, so T1 exercised the genuinely non-diagonal path. Deliberately NOT retested here (owned by nx_nofloat_kda_gate 5/5): delta-overwrite idempotence, per-channel vs scalar gating, bit-exact determinism. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0