feat: 0.2.0 – Konsens Qwen = Gemma = GLM als Standard, Gehrke komplett verarbeitet

- Dritte Lesart (GLM-4.6V über OpenRouter, ohne Nachdenken) muss zustimmen;
  keine 2:1-Annahme; fehlender Schlusspunkt der dritten Lesart zählt nicht
- run in Paketen zu 50 mit Statuszeile, beide lokalen Modelle parallel,
  GLM nur für Einige; nur die übernommene Lesart wird gegen das Bild geprüft
- Bild-Veto: Ähnlichkeit >= 0,65, Schriftgrößen-Prüfung abgeschaltet
- Antworten auf 800 Tokens begrenzt; Prompt-Version und Nachdenk-Einstellung
  im Cache-Schlüssel; Verfahrensversion für Prüfergebnisse
- Behoben: Gleichungsnummern aus alten Lesarten ("(3.7)"), gleichzeitiges
  Speichern zweier Engines, Absturz bei Formeln ohne aktuelle Lesart
- apply: deutsches Dezimalkomma 3{,}6; Review-Galerie zeigt aktuelle Lesarten
- tools/vergleich.py auf Konsens-Ablauf umgestellt
- Doku: README, BEDIENUNGSANLEITUNG, neues CHANGELOG

Gehrke-Brückenkurs: 942 von 1505 Formelbildern automatisch ersetzt (63 %),
62 von 62 von Hand geprüften korrekt; Pandoc/MathML 945 von 945.

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

tests/test_pipeline.py

@@ -131,22 +131,33 @@ def test_consensus_mode(epub2md_dir: Path) -> None:
     ws.save(display)
     pipeline.verify_all(source, ws, preset="cm", workers=2, quiet=True)
     st = {f.asset.rsplit("/", 1)[1]: f for f in ws.all()}
-    assert st["0001-inline.jpg"].status == "auto_accepted", st["0001-inline.jpg"].reason
+    assert pipeline.needs_third(st["0001-inline.jpg"]), st["0001-inline.jpg"].reason    # A = B, C missing
     assert st["0002-display.jpg"].status == "needs_review"
     assert "uneinig" in st["0002-display.jpg"].reason
 
-    # a third reading breaks the tie
-    d = st["0002-display.jpg"]
+    # the third reading confirms A = B -> accepted; for A != B a third reading does not decide 2:1
+    i, d = st["0001-inline.jpg"], st["0002-display.jpg"]
+    i.candidates.append(_reading("cloud", "C", r"L=\left\{-\sqrt{2};\sqrt{2}\right\}"))
     d.candidates.append(_reading("cloud", "C", r"x_{2,k} = \arcsin(-\frac35) + 2k\pi"))
+    ws.save(i)
     ws.save(d)
     pipeline.verify_all(source, ws, preset="cm", workers=2, quiet=True)
-    d = ws.load(d.id)
-    assert d.status == "auto_accepted" and "Stichwahl" in d.reason, d.reason
+    i, d = ws.load(i.id), ws.load(d.id)
+    assert i.status == "auto_accepted" and "dritte Lesart bestätigt" in i.reason, i.reason
+    assert d.status == "needs_review"
+
+    # a disagreeing third reading blocks the acceptance
+    i.candidates[-1] = _reading("cloud", "C", r"L=\{-\sqrt{2};\sqrt{3}\}")
+    i.status = "pending"
+    ws.save(i)
+    pipeline.verify_all(source, ws, preset="cm", workers=2, quiet=True)
+    i = ws.load(i.id)
+    assert i.status == "needs_review" and "weicht ab" in i.reason, i.reason
 
 
 def test_consensus_rules() -> None:
     from formula_img2latex.decide import consensus_thresholds, decide_consensus
-    th = consensus_thresholds()
+    th = consensus_thresholds({"min_p_mean": 0.9, "min_p_min": 0.3, "require_third": 0})
     a, b = _reading("local", "A", "x^2"), _reading("local2", "B", "x^{2}")
     veto = {"ok": True, "metrics": {"ncc": 0.9, "px_per_pt": 1.4}}
     assert decide_consensus([a, b], [(0, 1, "normalform")], veto, True, 1.4, th)[0] == "auto_accepted"
@@ -156,11 +167,18 @@ def test_consensus_rules() -> None:
     bad = {"ok": True, "metrics": {"ncc": 0.5, "px_per_pt": 1.4}}
     assert decide_consensus([a, b], [(0, 1, "normalform")], bad, True, 1.4, th)[0] == "needs_review"     # veto
     big = {"ok": True, "metrics": {"ncc": 0.9, "px_per_pt": 1.8}}
-    assert decide_consensus([a, b], [(0, 1, "normalform")], big, True, 1.4, th)[0] == "needs_review"     # size
+    sized = {**th, "veto_scale_dev": 0.15}
+    assert decide_consensus([a, b], [(0, 1, "normalform")], big, True, 1.4, sized)[0] == "needs_review"  # size check on
     assert decide_consensus([a, b], [(0, 1, "normalform")], veto, False, 1.4, th)[0] == "needs_review"   # pandoc
     nf = [_reading("local", "A", None, is_formula=False), _reading("local2", "B", None, is_formula=False)]
     assert decide_consensus(nf, [], None, True, 1.4, th)[0] == "not_formula"
     assert decide_consensus([a], [], veto, True, 1.4, th)[0] == "needs_review"                          # one model
+    strict = consensus_thresholds()                                                     # default: third must agree
+    c = _reading("cloud", "C", "x^2")
+    assert decide_consensus([a, b], [(0, 1, "normalform")], veto, True, 1.4, strict)[0] == "needs_review"
+    assert decide_consensus([a, b, c], [(0, 1, "normalform"), (0, 2, "normalform"), (1, 2, "normalform")],
+                            veto, True, 1.4, strict)[0] == "auto_accepted"
+    assert decide_consensus([a, b, c], [(0, 2, "normalform")], veto, True, 1.4, strict)[0] == "needs_review"  # no 2:1
 
 
 def test_canonical_form() -> None:
@@ -185,3 +203,34 @@ def test_logprob_confidence() -> None:
     p_min, p_mean = latex_token_probs(content, items)
     assert p_min == 0.5 and 0.8 < p_mean < 0.9          # only the four LaTeX tokens count
     assert parse_confidence(content) == 97
+
+
+def test_equation_number_only_from_chosen_reading() -> None:
+    from formula_img2latex.store import Formula
+    f = Formula(id="f-x-d", sha256="x", asset="a.png", display=True)
+    f.candidates = [{"engine": "local", "model": "old", "tex": r"a=b \tag{3.7}"}]    # an old, hallucinated number
+    assert pipeline._with_tag(f, "a=b") == "a=b"
+    assert pipeline._with_tag(f, r"a=b \tag{2.1}") == r"a=b\qquad\text{(2.1)}"
+    f.display = False
+    assert pipeline._with_tag(f, r"a=b \tag{2.1}") == "a=b"
+
+
+def test_german_decimal_comma(epub2md_dir: Path) -> None:
+    import json
+    source, _ws = pipeline.open_all(epub2md_dir, None, create=True)
+    assert source.format_math("3,6 + (1,2)", False) == "$3,6 + (1,2)$"          # no language: unchanged
+    rep = json.loads((epub2md_dir / "conversion-report.json").read_text())
+    rep["metadata"] = {"language": "de"}
+    (epub2md_dir / "conversion-report.json").write_text(json.dumps(rep))
+    source, _ws = pipeline.open_all(epub2md_dir, None)
+    assert source.format_math(r"\frac{72}{3,6} = 20, x", True) == r"$$\frac{72}{3{,}6} = 20, x$$"
+
+
+def test_verify_skips_formulas_without_current_reading(epub2md_dir: Path) -> None:
+    source, ws = pipeline.open_all(epub2md_dir, None, create=True)
+    pipeline.scan(source, ws, quiet=True)
+    f = next(f for f in ws.all() if f.asset.endswith("0001-inline.jpg"))
+    f.candidates = [{"engine": "local", "model": "A", "tex": "x", "is_formula": True, "prompt": 1},   # old reading
+                    {"engine": "local", "model": "A", "tex": None, "is_formula": False, "prompt": 2}]  # current: none
+    ws.save(f)
+    pipeline.verify_all(source, ws, preset="cm", workers=1, quiet=True)    # must not crash