fix: restore scoring signals lost in the evidence DB round-trip

Production scores come from _rescore_submission, which re-reads evidence via
_domain_evidence_from_ui and re-scores. Serialization dropped the promoted /
entity fields and the precise EvidenceSource, so on every rescore Naver
promotion (+40) and Google web-entity (+40) silently scored 0 and skip
evidence was mis-restored as FAILURE (+30 phantom).

- store_serialization: persist domainSource / promoted / promotionReason /
  entity / category in _evidence_payload and restore them in
  _domain_evidence_from_ui; resolve the source from the persisted enum
  (keyword heuristic kept only as a legacy fallback).
- search_result_promoter: structural no-result detection (covers image/blog/
  web variants), scan only result fields not the query, add Korean rights
  markers (excluding over-generic substrings).
- risk_scoring: dedup promoted Naver hits by query_signature; drop zero-point
  skip reasons from the operator-facing reasons list.

Tests: round-trip regression (promoted naver / web entity / two distinct
entities score > 0; skip adds no phantom points) + skip-reason exclusion.

src/rights_filter/analysis/risk_scoring.py

@@ -8,6 +8,7 @@ class RiskScorer:
         score = 0
         reasons: list[str] = []
         counted_web_buckets: set[str] = set()
+        counted_naver_buckets: set[str] = set()
 
         for item in evidence:
             if item.data.get("operator_status") in {"irrelevant", "false_positive"}:
@@ -37,9 +38,18 @@ class RiskScorer:
                         counted_web_buckets.add(bucket)
             elif item.source == EvidenceSource.NAVER_SEARCH:
                 points = _naver_search_points(item)
+                if not points:
+                    continue
+                # Count one promoted hit per executed query. Multiple promoted
+                # results from the same search are one signal, not N stacked +40s
+                # (which would push a single query straight to the high band).
+                bucket = _naver_search_bucket(item)
+                if bucket and bucket in counted_naver_buckets:
+                    continue
                 score += points
-                if points:
                 reasons.append(item.reason)
+                if bucket:
+                    counted_naver_buckets.add(bucket)
             elif item.source == EvidenceSource.LLM_SUMMARY:
                 continue
             elif item.source in {
@@ -54,13 +64,12 @@ class RiskScorer:
                 # e.g. "External API failed") still contribute the +30 signal.
                 if item.reason.startswith("LLM assistance failed"):
                     continue
-                score += (
+                if item.source in {EvidenceSource.FAILURE, EvidenceSource.ENRICHMENT_FAILURE}:
-                    30
+                    score += 30
-                    if item.source
-                    in {EvidenceSource.FAILURE, EvidenceSource.ENRICHMENT_FAILURE}
-                    else 0
-                )
                     reasons.append(item.reason)
+                # SEARCH_SKIPPED / EXTERNAL_SKIPPED contribute zero points and are
+                # operational status ("quota reached", "provider disabled"), not
+                # risk signals — keep them out of the operator-facing reasons list.
 
         score = min(100, score)
         if score >= 70:
@@ -113,6 +122,16 @@ def _naver_search_points(item: Evidence) -> int:
     return 0
 
 
+def _naver_search_bucket(item: Evidence) -> str:
+    # Dedup promoted Naver hits by the query that produced them. Distinct queries
+    # remain distinct signals; repeated hits from one query collapse to one.
+    signature = str(item.data.get("query_signature", "")).strip()
+    if signature:
+        return f"naver:{signature}"
+    query = str(item.data.get("query", "")).strip().casefold()
+    return f"naver-query:{query}" if query else ""
+
+
 def _unique(values: list[str]) -> list[str]:
     result: list[str] = []
     for value in values:

src/rights_filter/analysis/search_result_promoter.py

@@ -14,12 +14,24 @@ class SearchResultPromoter:
         if item.source != EvidenceSource.NAVER_SEARCH:
             return item
 
-        if item.reason == "Naver search returned no results":
+        # Structural no-result detection covers the image/blog/web variants
-            return replace(item, confidence=0.0, data={**item.data, "promoted": False})
+        # (each emits a different "... returned no results" reason). A real hit
+        # always carries at least one result field; a no-result item carries
+        # none, so it must never be promoted even if its query text contained a
+        # rights marker.
+        if not _has_result_content(item.data):
+            return replace(
+                item,
+                confidence=0.0,
+                data={**item.data, "promoted": False, "promotion_reason": "no search results"},
+            )
 
+        # Scan only the actual result fields — never the query. The query is our
+        # own generated input; including it self-promotes any result whose query
+        # happened to contain a marker word.
         text = " ".join(
             str(item.data.get(key, ""))
-            for key in ("title", "description", "query", "result_url")
+            for key in ("title", "description", "result_url")
         ).lower()
         promoted = _has_named_rights_signal(text)
         data = {
@@ -32,9 +44,15 @@ class SearchResultPromoter:
         return replace(item, confidence=0.8 if promoted else 0.2, data=data)
 
 
-def _has_named_rights_signal(text: str) -> bool:
+def _has_result_content(data: dict) -> bool:
-    markers = {
+    return any(data.get(key) for key in ("title", "description", "result_url", "image_url"))
-        "iu",
+
+
+# Markers that indicate a named person / work / rights context. English markers
+# plus the Korean equivalents actually emitted by Naver results and by this
+# project's own query templates, since Naver is a Korean-language provider and an
+# English-only list produces systematic false negatives.
+_RIGHTS_MARKERS = {
     "official",
     "album",
     "cover",
@@ -45,8 +63,33 @@ def _has_named_rights_signal(text: str) -> bool:
     "drama",
     "copyright",
     "celebrity",
-    }
+    "goods",
-    # Match whole words only. Plain `in` matching let short markers like "iu"
+    "공식",
-    # or "cover" hit inside unrelated words/URLs ("premium", "discover"),
+    "앨범",
-    # falsely promoting context-only results.
+    "커버",
-    return any(re.search(rf"\b{re.escape(marker)}\b", text) for marker in markers)
+    "캐릭터",
+    "웹툰",
+    "화보",
+    "저작권",
+    "연예인",
+    "아이돌",
+    "굿즈",
+    "포스터",
+    # NOTE: intentionally NOT including bare '게임'/'방송' — as CJK substrings they
+    # over-match benign Korean text ('방송통신위원회', casual game blogs) and
+    # over-promote. Their English counterparts (game/broadcast) stay, matched on
+    # word boundaries.
+}
+
+
+def _has_named_rights_signal(text: str) -> bool:
+    # ASCII markers match whole words only (so short ones like "cover" don't hit
+    # inside "discover"); CJK markers have no word boundaries in regex, so match
+    # them as substrings (Korean has no spaces between a marker and its context).
+    for marker in _RIGHTS_MARKERS:
+        if marker.isascii():
+            if re.search(rf"\b{re.escape(marker)}\b", text):
+                return True
+        elif marker in text:
+            return True
+    return False

src/rights_filter/server/store_serialization.py

@@ -261,6 +261,16 @@ def _evidence_payload(submission_id: str, evidence: Evidence) -> dict[str, Any]:
         "id": _evidence_id(submission_id, evidence),
         "group": "watchlist" if knowledge_entry_status == "watchlist" else "face_web" if face_crop_search else _ui_group(evidence.source),
         "source": source,
+        # Persist the precise domain source and the scoring-relevant data fields
+        # so the production round-trip (_domain_evidence_from_ui -> RiskScorer in
+        # _rescore_submission) reproduces the original score. Without these the
+        # coarse "source" string collapses SEARCH_SKIPPED/FAILURE together and the
+        # promoted/entity/category signals vanish, silently zeroing risk points.
+        "domainSource": evidence.source.value,
+        "promoted": bool(evidence.data.get("promoted", False)),
+        "promotionReason": str(evidence.data.get("promotion_reason", "")),
+        "entity": str(evidence.data.get("entity", "")),
+        "category": str(evidence.data.get("category", "")),
         "title": evidence.reason,
         "confidence": evidence.confidence,
         "query": str(evidence.data.get("query", "")),
@@ -327,6 +337,10 @@ def _domain_evidence_from_ui(payload: dict[str, Any]) -> Evidence:
             "match": payload.get("matchType", ""),
             "rank": payload.get("rank", ""),
             "score": payload.get("providerScore", ""),
+            "promoted": bool(payload.get("promoted", False)),
+            "promotion_reason": payload.get("promotionReason", ""),
+            "entity": payload.get("entity", ""),
+            "category": payload.get("category", ""),
             "contributed": payload.get("contributed", True),
             "status": payload.get("status", ""),
             "weak_hint": bool(payload.get("faceCropSearch", False)) or _is_google_weak_label_payload(payload),
@@ -344,6 +358,17 @@ def _domain_evidence_from_ui(payload: dict[str, Any]) -> Evidence:
 
 
 def _domain_source_from_ui_payload(payload: dict[str, Any]) -> EvidenceSource:
+    # Prefer the persisted precise source. This distinguishes SEARCH_SKIPPED
+    # (0 points) from FAILURE (+30) instead of guessing from title keywords,
+    # which previously misclassified skip evidence as failure on every rescore.
+    domain_source = str(payload.get("domainSource", ""))
+    if domain_source:
+        try:
+            return EvidenceSource(domain_source)
+        except ValueError:
+            pass
+
+    # Legacy fallback for rows written before domainSource was persisted.
     source = str(payload.get("source", ""))
     title = str(payload.get("title", "")).lower()
     if source == "failure" and (

tests/rights_filter/analysis/test_risk_scoring.py

@@ -280,3 +280,21 @@ def test_same_web_entity_repeated_does_not_double_count():
 
     assert result.score == 45
     assert result.band == "medium"
+
+
+def test_search_skipped_reason_is_not_listed_as_a_risk_reason():
+    # 2026-07-11 #16: a zero-point skip ("quota reached") is operational status,
+    # not a risk signal, and must not pollute the operator-facing reasons list.
+    result = RiskScorer().score(
+        [
+            Evidence(
+                source=EvidenceSource.SEARCH_SKIPPED,
+                reason="search API usage limit reached",
+                confidence=1.0,
+                data={},
+            ),
+        ]
+    )
+
+    assert result.score == 0
+    assert result.reasons == []