Replaces the one-off audit with a re-runnable instrument: per-module AST function extraction, name-preserving matching against Go/Rust sources, FULL/PARTIAL/NOT-TOUCHED verdicts with missing-function naming. Current honest measure: 13% function-level (290/2225 fns; 7 FULL / 50 PARTIAL / 121 NOT TOUCHED) — the true queue, regenerated with --write after every wave.