nx_nofloat_sample_gate.nx

nx_nofloat_sample_gate.nx -- TEMPERATURE SAMPLING decode in pure integer Q16 (CAP-NF-SAMPLE). Real LLMs GENERATE by sampling from the softmax distribution (with a temperature), not just greedy argmax. The no-float stack had only argmax; this adds a CORRECT fixed-point sampler: a deterministic LCG PRNG + softmax-with- temperature (scale logits by 1/T, max-subtract, fxexp, normalize) + inverse-CDF pick. All integer -> the sampling is REPRODUCIBLE given a seed (the no-float determinism exceed: same seed+weights -> same text, which float/GPU sampling cannot guarantee bit-for-bit). T1 LOW temperature concentrates on the argmax (>= 90% of draws) -- approaches greedy as T->0. T2 at temperature=1 the EMPIRICAL sample frequencies match softmax(logits) (a CORRECT sampler, measured). T3 REPRODUCIBLE: two runs with the same seed produce byte-identical samples (no-float determinism). T4 HIGH temperature is measurably MORE UNIFORM than low (argmax-frequency drops) -- temperature has teeth. Evidence -> knowledge/status/nofloat_sample.log. Sovereign: nx_nofloat_autograd (nfa_fxexp/nfa_qmul) + nx_syscalls. expect_exit: 0 license_tier: ORIGINAL