Add an LLM fact-check pass over the generated briefing

After summarisation the report is verified against the collected
payload (numbers, directions, unsupported claims) by a second LLM call
with a Verification JSON schema. Major issues trigger one revision and
a re-check; minor-only verdicts are normalized to ok. Failures never
block delivery. The verdict is stored in the result JSON and shown on
the dashboard (badge + 검증 메모). Disable with US2KR_VERIFY_ENABLED.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

us2kr/models.py

@@ -91,6 +91,25 @@ class Report(BaseModel):
     data_gaps: list[str] = Field(default_factory=list, description="수집 실패로 반영하지 못한 항목")
 
 
+class VerificationIssue(BaseModel):
+    claim: str = Field(description="문제가 된 브리핑 속 서술 (원문 그대로 인용)")
+    problem: str = Field(description="제공된 데이터와 어떻게 다른지 한 문장")
+    severity: Literal["minor", "major"] = Field(
+        default="minor", description="major: 수치·방향이 데이터와 모순 / minor: 근거 불충분·표현 문제"
+    )
+
+
+class Verification(BaseModel):
+    """LLM fact-check verdict for a generated Report, judged only against the payload."""
+
+    ok: bool = Field(description="major 이슈가 하나도 없으면 true")
+    issues: list[VerificationIssue] = Field(default_factory=list, description="발견된 문제 목록 (없으면 빈 배열)")
+
+    @property
+    def has_major(self) -> bool:
+        return any(i.severity == "major" for i in self.issues)
+
+
 class BriefingResult(BaseModel):
     payload: MarketPayload
     report: Report | None
@@ -98,3 +117,4 @@ class BriefingResult(BaseModel):
     model_used: str | None = None
     degraded: bool = False  # True when LLM failed and only the numeric table was produced
     error: str | None = None
+    verification: Verification | None = None  # None: not run (disabled / degraded / verifier failed)