Add Vision-LLM mode for direct image-to-JSON extraction

Tesseract OCR fails on rotated/low-contrast CD back covers.
New vision_llm module sends images directly to qwen3-vl via
Ollama chat API, bypassing OCR entirely. Robust JSON extraction
handles thinking tags, markdown blocks, and empty responses.
CLI scan/process commands gain --vision flag.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

tests/test_vision_llm.py

@@ -0,0 +1,37 @@
+"""Tests für die Vision-LLM JSON-Extraktion."""
+
+import pytest
+
+from musiksammlung.vision_llm import _extract_json
+
+
+def test_extract_pure_json():
+    text = '{"artist": "Test", "album": "Album"}'
+    assert '"Test"' in _extract_json(text)
+
+
+def test_extract_json_from_markdown_block():
+    text = 'Hier ist das Ergebnis:\n```json\n{"artist": "Test"}\n```\nFertig.'
+    assert '"Test"' in _extract_json(text)
+
+
+def test_extract_json_with_thinking_tags():
+    text = '<think>Ich denke nach...</think>\n{"artist": "Test", "album": "X"}'
+    result = _extract_json(text)
+    assert '"Test"' in result
+
+
+def test_extract_json_with_surrounding_text():
+    text = 'Das JSON:\n{"artist": "A", "album": "B"}\nEnde.'
+    result = _extract_json(text)
+    assert '"A"' in result
+
+
+def test_extract_json_empty_raises():
+    with pytest.raises(ValueError, match="Leere Antwort"):
+        _extract_json("")
+
+
+def test_extract_json_no_json_raises():
+    with pytest.raises(ValueError, match="Kein JSON"):
+        _extract_json("Hier ist kein JSON, nur Text.")