feat(B-SENS): Sensitive Data Boundary + AI/Data Exposure Policy + AIProvider Compliance

B-SENS: app/core/sensitive_data.py (NEU) — zentrale Sensitive-Field-Verwaltung
- SENSITIVE_FIELDS dict für contact/user/mail_account/system_settings
- is_sensitive(), sanitize_dict(), register_sensitive_fields()
- Integration: errors.py (Log-Redaction), audit.py (Audit-Masking), export_service.py (Export-Filter), embedding.py (Index-Filter)

B-DATA-POL: AI/Data Exposure Policy
- DATA_EXPOSURE_POLICY: pro Entity+Field welche Systeme erlaubt (llm_context/search/embeddings/rag/agent_memory/export)
- filter_for_llm_context/search/embeddings/export/rag/agent_memory()

B-AIPROV-COMP: AIProvider Compliance Metadata
- Migration 0119: 7 neue Spalten an ai_providers (region, hosting_type, dpa_status, retention_policy, training_on_customer_data, transfer_notice, allowed_data_classes)
- llm_client.py: get_provider_compliance() + check_data_class_allowed()

B-PRIV-TEST: 76 Tests in test_sensitive_data.py — alle grün
- Sensitive Fields, Exposure Policy, Provider Compliance, Secrets-always-blocked
- Keine Regression: 39 LLM-Client Tests grün

app/plugins/builtins/unified_search/embedding.py

@@ -147,6 +147,29 @@ async def index_entity(
             logger.debug("Empty embedding text for %s/%s", entity_type, entity_id)
             return False
 
+        # Apply sensitive-data filter: ensure no sensitive fields leak into
+        # embedding text.  The provider builds text from DB columns, so we
+        # rely on the provider selecting only non-sensitive columns.  This
+        # is a secondary safety net — providers should use filter_for_embeddings
+        # when constructing embedding text from dict-like data.
+        from app.core.sensitive_data import get_sensitive_fields
+
+        sensitive = get_sensitive_fields(entity_type)
+        if sensitive:
+            # If any sensitive field name appears as a substring in the text,
+            # it's likely a key=value pair — redact it.  This is a best-effort
+            # guard; providers are expected to exclude sensitive columns at
+            # the SQL level.
+            for field_name in sensitive:
+                # Only redact if the field name appears as a key-like pattern
+                import re
+                text = re.sub(
+                    rf"\b{re.escape(field_name)}\s*[=:]\s*\S+",
+                    f"{field_name}=***REDACTED***",
+                    text,
+                    flags=re.IGNORECASE,
+                )
+
         embedding = await generate_embedding(text, db=db, tenant_id=tenant_id)
         if not embedding:
             return False