feat(stage5): implement claim extraction — atomic verifiable claims from sources

- Claim model with provenance, evidence_span, attribution, claim_type
- Stage5Extractor: LLM-based atomic claim extraction from source content
  - Never summarizes — always extracts atomic, verifiable claims
  - Claims require evidence span (exact quote from source)
  - Attribution per claim (who says what)
  - Claim types: fact, opinion, prediction, recommendation, claim
  - Confidence score 0.0–1.0 per claim
  - Bounded concurrency, SSRF-safe, max content truncation
- REST API: GET/POST /research/{run_id}/claims
- 36 tests: parsing, edge cases, integration, validation

src/nsct/models/claim.py

@@ -0,0 +1,107 @@
+"""Pydantic v2 schema — Claim für Stage 5: Claim Extraction.
+
+Jeder Claim ist eine atomare, überprüfbare Behauptung mit Provenance.
+Keine Zusammenfassungen — immer einzelne, isolierte Claims.
+"""
+
+from __future__ import annotations
+
+import json
+from datetime import datetime
+from enum import Enum
+from typing import Any
+from uuid import UUID, uuid4
+
+from pydantic import BaseModel, Field, field_validator
+
+
+class ClaimType(str, Enum):
+    """Klassifikation eines Claims nach seinem Epistemischen Status."""
+
+    FACT = "fact"
+    OPINION = "opinion"
+    PREDICTION = "prediction"
+    RECOMMENDATION = "recommendation"
+    CLAIM = "claim"
+
+
+class Claim(BaseModel):
+    """Eine atomare, überprüfbare Behauptung aus einer Quelle.
+
+    Felder:
+        id: UUID — Primärschlüssel
+        research_run_id: UUID — Zuordnung zum Research-Run
+        source_id: UUID — Quelle, aus der der Claim extrahiert wurde
+        claim_text: str — Der atomare Claim-Text (NOT NULL)
+        evidence_span: str — Das exakte Zitat aus dem Original
+        claim_type: ClaimType — Typisierung
+        source_url: str — URL der Quelle
+        confidence: float 0.0-1.0 — Wie sicher ist der Claim?
+        metadata: dict — Zusätzliche Kontextdaten
+        created_at: datetime — Erstellungszeitpunkt
+    """
+
+    id: UUID = Field(default_factory=uuid4)
+    research_run_id: UUID = Field(
+        ..., description="Research-Run-UUID für Gruppierung"
+    )
+    source_id: UUID = Field(
+        ..., description="Source-UUID, aus der dieser Claim stammt"
+    )
+    claim_text: str = Field(
+        ..., min_length=1, description="Der atomare Claim-Text"
+    )
+    evidence_span: str = Field(
+        ..., description="Exakter Textabschnitt im Original als Evidenz"
+    )
+    claim_type: ClaimType = Field(default=ClaimType.CLAIM)
+    source_url: str = Field(
+        ..., description="URL der Quelle"
+    )
+    confidence: float = Field(
+        default=1.0, ge=0.0, le=1.0, description="Confidence 0-1"
+    )
+    metadata: dict[str, Any] = Field(
+        default_factory=dict, description="Zusätzliche Metadaten"
+    )
+    created_at: datetime = Field(
+        default_factory=datetime.utcnow
+    )
+
+    @field_validator("claim_text")
+    @classmethod
+    def claim_text_not_empty(cls, v: str) -> str:
+        if not v.strip():
+            raise ValueError("claim_text darf nicht leer sein")
+        return v
+
+    @field_validator("evidence_span")
+    @classmethod
+    def evidence_span_not_empty(cls, v: str) -> str:
+        if not v.strip():
+            raise ValueError("evidence_span darf nicht leer sein")
+        return v
+
+    @field_validator("source_url")
+    @classmethod
+    def source_url_not_empty(cls, v: str) -> str:
+        if not v.strip():
+            raise ValueError("source_url darf nicht leer sein")
+        return v
+
+    model_config = {"frozen": True}
+
+
+class ClaimExtractionResult(BaseModel):
+    """Ergebnis einer Claim-Extraktion pro Dokument."""
+
+    source_id: UUID
+    source_url: str
+    research_run_id: UUID
+    claims: list[Claim] = Field(default_factory=list)
+    total_tokens: int = 0
+    extraction_tool: str = "llm"
+    metadata: dict[str, Any] = Field(
+        default_factory=dict,
+        description="Extraktions-Metadaten (z.B. Token-Anzahl, Dauer)"
+    )