nx_nofloat_propose_chat_gate.nx

nx_nofloat_propose_chat_gate.nx -- PROPOSE->VERIFY rung 3 (2026-07-15): a BETTER PROPOSER on the SAME RULER. The raw 3-shot completion proposer measured solve-rate 1/10 on the 10-question inversion ruler (nx_nofloat_propose_solve_gate -- the model pattern-copies instead of inverting). This gate asks the SAME 10 questions through the serve's CHATML path (gp[12]=1: <|im_start|>user ... <|im_start|>assistant -- the instruct-tuned surface, natural-language question + answer-format instruction) and measures whether the prompting MODE alone moves the solve-rate. Same kernel, same questions, same teeth => the delta is the prompting mode's, honestly. Teeth stay model-IQ-independent (the RATE is data, not a tooth): T1 mechanism live: >=8/10 proposals parse as integers; zero generation or kernel failures T2 soundness cross-check: kernel verdict on the instantiation == independent in-gate arithmetic (0 violations) T3 NEG-CONTROL: all 10 wrong-by-one instantiations (in-grammar: x+1, stepping to x-1 at the operand caps) REFUTED T4 NEG-CONTROL: the LIVE unsolvable chat question ('2 times what number equals 9?') can NEVER certify T5 determinism: re-running Q1 reproduces BYTE-IDENTICAL text (greedy no-float exceed, chat path included) INSTRUMENT v2 (2026-07-15, after the 1.5B first-light): max_new 8->24 (EOS-complete -- chattier models were TRUNCATED mid-sentence: 'The number is 7 - 3' cut before the result) + LAST-integer extraction (pvl_last_uint -- the answer position in sentence replies; first-int grabbed operands). Applied uniformly to every model on this instrument; v1 numbers (0.5B 3/10, first-int, max_new=8) stay banked as v1. argv[1] = optional gguf path. Requires the gguf. Heavy: ~2-8 min (model-size dependent). Return from main (pool reap), never sys_exit. No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL