feat: Inhaltsinventar in SQLite — KI-Inhaltsangabe je Objekt (URL + Hash)

Vereinheitlicht den bisherigen JSON-Ledger zu EINEM abfragbaren SQLite-Store
pro Site (data/content_inventory.db): jedes geprüfte Objekt mit Inhalts-Hash,
URL(s), Sicherheits-Verdikt UND neutraler KI-Inhaltsangabe.

- baseline.py: load/save_ai_ledger jetzt SQLite-gestützt (Dict-Schnittstelle
  bleibt → Analyzer unverändert). Tabellen objects + object_urls. Einmalige
  Migration eines vorhandenen ai_ledger.json → SQLite (.migrated). UPSERT mit
  first_seen-Erhalt; object_urls transaktional ersetzt. Neu: query_inventory,
  export_inventory_csv (beide migrieren failsafe).
- ai_analyzer.py: Antwort-Schema + Prompt um neutrales Feld 'description'
  erweitert (im selben Call, keine Extrakosten); _make_entry speichert es.
- __main__.py: neues Kommando 'inventory' (Übersicht, --search, --kind, --csv).
- Doku: README + Bedienungsanleitung.

Audio/Video sind im Schema (kind) vorbereitet (Phase 2: Extractor + Modalitäten).
bredelar.info real migriert: 245 Objekte (200 Bilder, 45 Texte).

Tests: 266 grün (+6: Round-Trip mit description, Migration, first_seen-Erhalt,
invalidate, Suche/CSV-Export, description landet im Ledger).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tests/test_ai_analyzer.py

@@ -506,6 +506,17 @@ class TestParallel:
         assert len(res["findings"]) == 2                  # aber Fund für beide URLs
         assert {f["url"] for f in res["findings"]} == {"https://x.de/a", "https://x.de/b"}
 
+    def test_description_stored_in_ledger(self, tmp_path, monkeypatch):
+        """Die neutrale Inhaltsangabe aus dem Verdikt landet im Inventar."""
+        monkeypatch.setenv("OPENROUTER_API_KEY", "test")
+        bm = BaselineManager(tmp_path)
+        v = {"description": "Eine Seite über die Dorfgeschichte.", "category": "clean",
+             "severity": "none", "confidence": 0.9, "explanation": "ok"}
+        with patch("scanner.ai_analyzer._openrouter_chat", return_value=v):
+            run_ai_analysis(_cfg(), bm, _snap(_LONG), {})
+        entry = next(iter(bm.load_ai_ledger()["entries"].values()))
+        assert entry["description"] == "Eine Seite über die Dorfgeschichte."
+
     def test_changed_page_always_analyzed_despite_backlog_limit(self, tmp_path, monkeypatch):
         """Geänderte Seiten werden IMMER geprüft — die Bestands-Drossel gilt nicht für sie."""
         monkeypatch.setenv("OPENROUTER_API_KEY", "test")

tests/test_baseline.py

@@ -159,3 +159,67 @@ class TestRebuildBaseline:
 
         baseline = bm.load_baseline()
         assert baseline["pages"]["https://b.info/"]["text"] == "Neu"
+
+
+# ---------------------------------------------------------------------------
+# AI content inventory (SQLite store: verdict + description + url→hash)
+# ---------------------------------------------------------------------------
+
+def _entry(kind="image", url="https://b.info/x.jpg", category="clean", desc="Ein Logo"):
+    return {"kind": kind, "url": url, "description": desc, "category": category,
+            "severity": "none", "confidence": 0.9, "explanation": "ok",
+            "model": "m", "dismissed": False, "checked_at": "2026-01-01T00:00:00"}
+
+
+class TestAiInventory:
+    def test_roundtrip_preserves_entries_and_url_hashes(self, tmp_path):
+        bm = BaselineManager(tmp_path)
+        bm.save_ai_ledger({"entries": {"H1": _entry(desc="Schützenfest-Plakat")},
+                           "url_hashes": {"https://b.info/x.jpg": "H1"}})
+        led = bm.load_ai_ledger()
+        assert led["entries"]["H1"]["description"] == "Schützenfest-Plakat"
+        assert led["entries"]["H1"]["dismissed"] is False
+        assert led["url_hashes"] == {"https://b.info/x.jpg": "H1"}
+
+    def test_migrates_json_ledger_once(self, tmp_path):
+        (tmp_path).mkdir(parents=True, exist_ok=True)
+        old = tmp_path / "ai_ledger.json"
+        old.write_text(json.dumps({"entries": {"H1": _entry()}, "url_hashes": {"u": "H1"}}),
+                       encoding="utf-8")
+        bm = BaselineManager(tmp_path)
+        led = bm.load_ai_ledger()
+        assert "H1" in led["entries"]
+        assert not old.exists()                                  # umbenannt
+        assert (tmp_path / "ai_ledger.json.migrated").exists()
+        assert (tmp_path / "content_inventory.db").exists()
+
+    def test_first_seen_preserved_last_seen_updated(self, tmp_path):
+        bm = BaselineManager(tmp_path)
+        bm.save_ai_ledger({"entries": {"H1": _entry()}, "url_hashes": {}})
+        rows1 = bm.query_inventory()
+        first1, last1 = rows1[0]["first_seen"], rows1[0]["last_seen"]
+        import time; time.sleep(0.01)
+        bm.save_ai_ledger({"entries": {"H1": _entry(category="propaganda")}, "url_hashes": {}})
+        rows2 = bm.query_inventory()
+        assert rows2[0]["first_seen"] == first1          # bleibt
+        assert rows2[0]["last_seen"] >= last1            # aktualisiert
+        assert rows2[0]["category"] == "propaganda"      # Feld aktualisiert
+
+    def test_invalidate_removes_url(self, tmp_path):
+        bm = BaselineManager(tmp_path)
+        bm.save_ai_ledger({"entries": {}, "url_hashes": {"a": "H1", "b": "H2"}})
+        assert bm.invalidate_ai_url_hashes(["a"]) == 1
+        assert bm.load_ai_ledger()["url_hashes"] == {"b": "H2"}
+
+    def test_query_search_and_csv_export(self, tmp_path):
+        bm = BaselineManager(tmp_path)
+        bm.save_ai_ledger({"entries": {
+            "H1": _entry(url="https://b.info/kloster.jpg", desc="Foto des Klosters Bredelar"),
+            "H2": _entry(url="https://b.info/auto.jpg", desc="Ein Auto"),
+        }, "url_hashes": {}})
+        hits = bm.query_inventory(search="kloster")
+        assert len(hits) == 1 and "Kloster" in hits[0]["description"]
+        out = tmp_path / "inv.csv"
+        assert bm.export_inventory_csv(out) == 2
+        content = out.read_text(encoding="utf-8")
+        assert "description" in content and "Kloster" in content