nx_gate_anchor_edit.nx

nx_gate_anchor_edit.nx -- THE MECHANICAL HALF OF THE D001 ANCHOR RUNG (seq585; D001-B lane 2026-07-31). nx_gate_migrate is deliberately a VERIFIER and NEVER an auto-editor, so that it can never mangle 1918 bespoke gates. That is the right call -- but it leaves the ~953 unjudgeable gates with no mechanical EDIT step, and an edit-per-gate routed through a session context does not scale to a 1057-gate lane. This organ is that missing step. The division of labour is PRESERVED, not weakened: THE EDITOR PROPOSES A CANDIDATE FILE, THE VERIFIER (nx_gate_migrate anchor) REMAINS THE SOLE AUTHORITY ON WHETHER IT MAY BE COMMITTED. This organ never writes a gate in place -- only ever a separate candidate path -- so its blast radius is bounded BY CONSTRUCTION, not by promise (rule 26 posture). THE TRANSFORM (narrow by construction; the local awk v4.x failures are encoded here as REFUSALS, not hopes). Insert verdict= immediately before a GREEN or RED token, but ONLY where ALL of these hold: (a) the line is NOT a comment (first non-blank bytes are not //). A mangled comment is invisible to a behaviour-preserving oracle -- which is exactly how MEASURED -> MEASUverdict=RED shipped locally. (b) the line ALSO contains sys_exit( or return -- the verdict line is the one that decides the exit. (c) the token is STANDALONE: neither neighbouring byte is [A-Za-z0-9_], so MEASURED / RECOVERED / TEXTURED can never be glued mid-word (MEASURED literally contains RED at index 5). (d) the token is not ALREADY preceded by verdict= -- idempotent (rule 10), so re-runs are no-ops. Every other byte is copied VERBATIM. Exit codes are distinct so a batch driver can branch without parsing prose: 0 = candidate written 5 = NOEDIT (nothing matched -> hand-treat; never a silent success) 2 = usage 4 = I/O error license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0