nx_reader_span_gate.nx

nx_reader_span_gate.nx -- RUNG 2 of the NEURAL PASSAGE READER arc (deep-research R2): the READER-SHAPED MODEL trains END-TO-END on the verified sovereign no-float stack. Composes nx_nofloat_autograd's PROVEN transformer block (the exact blk_fwd shape from the green blockfloat gates: embed -> rmsnorm -> Q/K/V -> RoPE -> QK^T -> causal softmax -> A.V -> Wo -> residual -> rmsnorm -> SwiGLU FFN -> residual) + a SPAN HEAD (matmul_nt of a trained probe against the contextual states -> [1,L] position logits) + fused softmax-CE. TASK (question-conditioned span pointing = the reader mechanism): passage = 3 (key,value) pairs in random order; question = one key token; GOLD SPAN = the position where the queried key occurs in the passage. To point there the model must ALIGN the question to the context (attention); a QUESTION-BLIND ablation (same params/budget, attention removed -> h_i sees only its own token, exactly the isolated-candidate reader that plateaued at 251/148/101) can at best fire on one fixed key = ~333 permille -> liar-kill. T1 reader (attention block) span accuracy > 700 permille held-out T2 question-blind ablation < 450 permille AND reader beats it by > 250 (liar-kill) T3 descent: one nfa_backward+nfa_sgd step on fresh weights DECREASES that example's loss (real gradients) T4 deterministic re-eval (integer stack -> bit-identical) expect_exit: 0 license_tier: ORIGINAL Sovereign: nx_nofloat_autograd + nx_syscalls.