nx_gate_dry_apply.nx

nx_gate_dry_apply.nx -- THE D001 CANDIDATE GENERATOR (2026-07-31, ws=gate-dry-d001). WHAT IT IS: a MECHANICAL REWRITER that emits a MINIMAL-FORM migration candidate for one gate. It is explicitly NOT a correct rewriter and does not need to be -- nx_gate_migrate verify is the safety net, and that net is independently gate-proven to BITE on real files (a check-dropping edit was refused on clause (c) while clauses (a) and (b) both waved it through, and the original was restored byte-intact). So the division of labour is: this organ GUESSES, the verifier DECIDES, and a wrong guess costs nothing because the verifier restores the original itself. THE MINIMAL FORM, and why it is the mechanizable one: leave EVERY check row byte-untouched and replace ONLY the hand-rolled verdict tail. Clause (c) -- the PASS/FAIL vector, which is the hardest clause to satisfy -- then cannot diverge BY CONSTRUCTION rather than by the author having hand-checked it. And the verdict tail is precisely the debt D001 names ("2601 gates roll their own verdict emission"), so the minimal form eats the debt in full. Full gv_check adoption is a separate, optional, later rung. SAFE BY REFUSAL (the nx_map_delegate discipline): every structural landmark must be found EXACTLY, and the counter idiom must match one of the three known shapes. Anything else -> SKIP, emit nothing, exit 1. A SKIP is a fine outcome and must never be dressed up as a migration. SHAPES RECOGNISED (measured over the local buildroot: 950 of 1390 candidates carry a counter idiom): A tot[0] / tot[1] -- an mmap'd pair, pass in [0] and fail in [1] (nx_brand_snap_gate) B var pass / var fail -- two scalars, pass and fail counted separately C var pass / var tot -- pass counted, total counted usage: nx_gate_dry_apply <gate> <out-candidate-path> then: nx_gate_migrate verify <gate> <out-candidate-path> exit 0 = candidate written · 1 = SKIP (shape not recognised) · 2 = usage · 4 = harness/root failure license_tier: ORIGINAL No hw writes (Rule 26).