nx_nofloat_arch_config_gate.nx

nx_nofloat_arch_config_gate.nx -- proves ARCH-CONFIGURABLE inference: read the model's dims from GGUF METADATA (nac_read_config) instead of hardcoding, then run the REAL no-float forward with those read dims and confirm it STILL generates ' Paris' (id 12095). This is the keystone that lets our inference run ANY Qwen2/Llama-schema gguf -- the real 0.5B AND a small from-scratch model -> closes the sovereign train->serve loop (our f32 trainer -> gguf writer -> THIS inference). expect_exit: 0 license_tier: ORIGINAL D001 MIGRATION 2026-08-25. This gate hand-rolled its verdict: a pass/total pair, inline PASS/FAIL prints and a prose "verdict=" line, with main returning a bare 0 or 1. A gate whose EXIT CODE does not carry its verdict can print FAIL and still exit 0, silently blessing every failure it finds -- so it now inherits nx_gate_verdict. gv_ctr makes declared and executed tooth counts ONE number (a tooth that stops running can no longer read GREEN), gv_check names every tooth, and gv_verdict owns the exit code: 0 GREEN / 1 RED / 3 SKIP. THE MODEL FILE IS A PRECONDITION, NOT A SUBJECT. If the fixture gguf is absent this gate proves NOTHING about arch-configurable inference; it used to return 1, which reports its own environment in the same word it reports a regression. It now SKIPs. SKIP is not a pass -- it blocks any claim that this works. The old single compound T1 (seven dims nested in one if) is decomposed into one tooth per dimension. The predicate set is unchanged; what changes is that a failure now NAMES the dim that drifted instead of saying "config mismatch" and leaving the reader to guess which of the seven conjuncts fired.