pdf2md v5: gescannte und digitale PDFs, EPUB-Export, Tests

- Seitenklassifizierung (Text / Scan / OCR-Ebene / leer) mit OCR-Plan pro Seite:
  Scans per Tesseract, gerade OCR-Ebenen übernehmen, schiefe neu erkennen,
  Foto-/Coverseiten nur nach Probe-OCR; --ocr auto|always|never, --jobs
- Nachbearbeitung: Satzende-Erkennung mit « “ ‹, Silbentrennung, Überschriften-
  ebenen aus PDF-Lesezeichen, Rückstufung von Satzresten, Inhaltsverzeichnis
  trennen, Seitenzahlen/Kolumnentitel/Bildtext entfernen
- Patches für pymupdf4llm: Bindestrich-Komposita am Zeilenende, Absätze per
  Einzug in Scan-Blöcken
- --epub über pandoc (Metadaten, Cover, Kapitelaufteilung)
- pytest-Suite mit synthetischen Test-PDFs (79 Tests)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tests/test_classify.py

@@ -0,0 +1,144 @@
+"""Seitenklassifizierung und OCR-Planung."""
+
+import pymupdf
+import pytest
+
+import pdf2md_v5 as p
+from conftest import needs_tesseract
+
+
+def kinds_of(path):
+    with pymupdf.open(path) as doc:
+        return [p.classify_page(page) for page in doc]
+
+
+def test_digital_pages_are_text(pdfs):
+    assert kinds_of(pdfs["digital"]) == [p.PAGE_TEXT] * 4
+
+
+def test_scanned_pages_are_scans(pdfs):
+    assert kinds_of(pdfs["scanned"]) == [p.PAGE_SCAN] * 4
+
+
+@needs_tesseract
+def test_scans_with_ocr_layer_are_detected(pdfs):
+    assert kinds_of(pdfs["scanned_ocr"]) == [p.PAGE_OCR_LAYER] * 4
+
+
+def test_mixed_document(pdfs):
+    assert kinds_of(pdfs["mixed"]) == [p.PAGE_TEXT, p.PAGE_TEXT, p.PAGE_SCAN, p.PAGE_SCAN]
+
+
+def test_blank_page_is_empty(pdfs):
+    assert kinds_of(pdfs["blank"]) == [p.PAGE_EMPTY]
+
+
+def test_plan_never_uses_no_ocr(pdfs):
+    with pymupdf.open(pdfs["scanned"]) as doc:
+        _, plan = p.plan_pages(doc, "never", "deu", 300)
+    assert plan == [p.RUN_PLAIN] * 4
+
+
+@needs_tesseract
+def test_plan_auto_scanned(pdfs):
+    with pymupdf.open(pdfs["scanned"]) as doc:
+        _, plan = p.plan_pages(doc, "auto", "deu", 300)
+    assert plan == [p.RUN_OCR] * 4
+
+
+@needs_tesseract
+def test_plan_auto_reuses_existing_ocr_layer(pdfs):
+    with pymupdf.open(pdfs["scanned_ocr"]) as doc:
+        _, plan = p.plan_pages(doc, "auto", "deu", 300)
+    assert plan == [p.RUN_REUSE] * 4
+
+
+def _scan_with_text_layer(path, angle):
+    """Scanseite mit unsichtbarer, um `angle` Grad gedrehter Textebene."""
+    doc = pymupdf.open()
+    page = doc.new_page(width=420, height=595)
+    pix = pymupdf.Pixmap(pymupdf.csGRAY, pymupdf.IRect(0, 0, 200, 280), 0)
+    pix.set_rect(pix.irect, (230,))
+    page.insert_image(page.rect, pixmap=pix)
+    for i in range(20):
+        point = pymupdf.Point(50, 80 + 20 * i)
+        page.insert_text(
+            point,
+            "Eine Zeile der vorhandenen OCR-Ebene",
+            render_mode=3,
+            morph=(point, pymupdf.Matrix(1, 1).prerotate(-angle)),
+        )
+    doc.save(path)
+    return path
+
+
+@needs_tesseract
+@pytest.mark.parametrize("angle, expected", [(0, p.RUN_REUSE), (0.5, p.RUN_OCR)])
+def test_plan_auto_redoes_skewed_ocr_layer(tmp_path, angle, expected):
+    path = _scan_with_text_layer(tmp_path / f"layer_{angle}.pdf", angle)
+    with pymupdf.open(path) as doc:
+        kinds, plan = p.plan_pages(doc, "auto", "deu", 300)
+    assert kinds == [p.PAGE_OCR_LAYER]
+    assert plan == [expected]
+
+
+@needs_tesseract
+def test_plan_auto_mixed_probes_scan_pages(pdfs):
+    # 2 Text- zu 2 Scanseiten gilt nicht als Scan-Dokument; die Scanseiten
+    # werden per Probe-OCR geprüft und bestehen sie.
+    with pymupdf.open(pdfs["mixed"]) as doc:
+        _, plan = p.plan_pages(doc, "auto", "deu", 300)
+    assert plan == [p.RUN_PLAIN, p.RUN_PLAIN, p.RUN_OCR, p.RUN_OCR]
+
+
+@needs_tesseract
+def test_plan_auto_skips_photo_page(pdfs):
+    with pymupdf.open(pdfs["photo"]) as doc:
+        kinds, plan = p.plan_pages(doc, "auto", "deu", 300)
+    assert kinds[-1] == p.PAGE_SCAN
+    assert plan == [p.RUN_PLAIN] * 5
+
+
+@needs_tesseract
+def test_plan_always(pdfs):
+    with pymupdf.open(pdfs["digital"]) as doc:
+        _, plan = p.plan_pages(doc, "always", "deu", 300)
+    assert plan == [p.RUN_OCR] * 4
+
+
+def test_missing_language_gives_clear_error():
+    if not pymupdf.get_tessdata():
+        pytest.skip("Tesseract nicht installiert")
+    with pytest.raises(RuntimeError, match="tesseract-ocr-xyz"):
+        p.check_ocr_language("deu+xyz")
+
+
+@pytest.mark.parametrize(
+    "text, language, expected",
+    [
+        ("Der Brief lag auf dem Tisch. " * 10, "deu", True),
+        ("Am Abend saß er lange in der Küche und trank einen Tee. Die Gedanken "
+         "an den Brief ließen ihn nicht los.", "deu", True),
+        ("The letter was lying on the table in the kitchen, and he read it "
+         "twice before he put it away.", "deu+eng", True),
+        # Cover mit Titel und Autor: zu wenig Wörter
+        ("Reinhard Otto Kranz OPERATION LJUTSCH Band I Schlüssel zur Deutschen Einheit", "deu", False),
+        ("fi@ |l1 ;; ¦¦ ~~ ‚, ., " * 20, "deu", False),
+        # echte Tesseract-Ausgabe einer Rauschseite
+        ("©Se te\nansa |\naa\n.\nerNee =\nae\neea\nBa:Ben\nBena\needaaf\neeee\naar.\nEl eeBR\neen\n", "deu", False),
+        ("Zu kurz.", "deu", False),
+    ],
+)
+def test_looks_like_text(text, language, expected):
+    assert p.looks_like_text(text, language) is expected
+
+
+def test_make_tasks_groups_and_chunks():
+    plan = [p.RUN_PLAIN] * 3 + [p.RUN_OCR] * 12 + [p.RUN_PLAIN]
+    tasks = p.make_tasks(plan)
+    assert [(run, pages[0], pages[-1]) for run, pages in tasks] == [
+        (p.RUN_PLAIN, 0, 2),
+        (p.RUN_OCR, 3, 12),
+        (p.RUN_OCR, 13, 14),
+        (p.RUN_PLAIN, 15, 15),
+    ]