audit: composite AUDIT_PASSED verdict; missing shards are unmeasured, not clean

build/audit_contamination.py

@@ -129,10 +129,14 @@ def main():
     hits = {k: 0 for k in ref_sets}
     hits_by_source = {}
     docs = grams_seen = 0
+    missing = []
     import build_mix as BM
     for rec in man["shards"]:
         p = os.path.join(a.root, rec["file"])
         if not os.path.exists(p):
+            # A shard that is not on disk is not "clean", it is unmeasured. Silence here would let an
+            # audit over an empty directory report zero contamination (E-017's shape).
+            missing.append(rec["file"])
             continue
         for d in BM.iter_shard_docs(p):
             docs += 1
@@ -150,22 +154,35 @@ def main():
         if docs % 200000 < 5000:
             print(f"  audited {docs:,} docs, {grams_seen:,} mix grams", flush=True)
 
+    total_overlap = sum(hits.values())
     report = {
         "k": K,
         "reference": {k: {"distinct_grams": len(v)} for k, v in ref_sets.items()},
         "reference_unreadable": errors,
+        "shards_expected": len(man["shards"]),
+        "shards_missing": missing,
         "mix_documents_audited": docs,
         "mix_grams_audited": grams_seen,
         "overlap_grams_total_by_ref": hits,
         "overlap_rate_by_ref": {k: round(v / max(1, grams_seen), 8) for k, v in hits.items()},
         "overlap_by_source": hits_by_source,
+        # Composite, so Gate 2 cannot be claimed on an audit that measured nothing. Every clause here
+        # has been observed to fail independently: unreadable refs, absent shards, zero docs.
+        "COVERED_ALL_SHARDS": not missing and len(man["shards"]) > 0,
+        "MEASURED": docs > 0 and grams_seen > 0 and len(ref_sets) >= 6 and not errors,
+        "CLEAN": total_overlap == 0,
+        "AUDIT_PASSED": bool(not missing and man["shards"] and docs > 0 and grams_seen > 0
+                             and len(ref_sets) >= 6 and not errors and total_overlap == 0),
         "note": ("counts only; no item text was read or printed. Reference splits are train/validation/dev "
                  "per task -- never test, per §3.3. Any source with a non-zero overlap here must be dropped "
                  "or cleaned before Gate 2 is claimed."),
     }
+    print(f"VERDICT AUDIT_PASSED={report['AUDIT_PASSED']} docs={docs:,} overlap={total_overlap} "
+          f"missing_shards={len(missing)} refs_ok={len(ref_sets)} refs_failed={len(errors)}")
     print("AUDIT_JSON_BEGIN")
     print(json.dumps(report, indent=1, default=str))
     print("AUDIT_JSON_END")
+    raise SystemExit(0 if report["AUDIT_PASSED"] else (6 if report["MEASURED"] else 7))
 
 
 if __name__ == "__main__":