Ajout d'un outil de benchmark des modèles d'analyse + support des modèles à raisonnement

- Nouvelle commande `inkflow benchmark` : compare la sortie d'analyse aux fichiers
  de référence (data/<slug>/reference/), met plusieurs modèles en concurrence,
  table rich + rapport JSON. Métriques : attribution de locuteur, incises, type/glued.
  Flags --models, --temperature, --reasoning, --stream, --use-cached + suivi par chapitre.
- analysis/benchmark.py : scoring pur (testable) + runner multi-modèles (un MLX à la fois).
- gemma.py : support des modèles à raisonnement (retrait de la pensée, désactivation
  via enable_thinking hors --reasoning, arrêt anticipé sur JSON complet, plafond +
  température dédiés anti-boucle), récupération du chat_template manquant (fix Mistral),
  streaming des tokens (set_token_sink).
- settings.py : gemma_reasoning, gemma_reasoning_max_tokens, gemma_reasoning_temperature.
- Tests : test_benchmark.py (scoring pur), test_gemma_reasoning.py.

Conclusion benchmark : Qwen3.6-27B-8bit non-raisonnant = meilleur modèle d'analyse.

backend/tests/test_benchmark.py

@@ -0,0 +1,159 @@
+"""Tests purs du scoring de benchmark (sans Gemma ni disque).
+
+Monte des `ChapterAnalysis` synthetiques et verifie les metriques :
+alignement, attribution du locuteur (avec normalisation d'alias), incises
+(exact vs chevauchement), type/glued, et micro-moyenne sur plusieurs chapitres.
+"""
+from __future__ import annotations
+
+from inkflow.analysis.benchmark import (
+    _score_counts,
+    align,
+    aggregate,
+    score_chapter,
+)
+from inkflow.models import (
+    Cast,
+    Character,
+    ChapterAnalysis,
+    Incise,
+    Segment,
+    SegmentType,
+)
+
+
+def _seg(text, *, type="narration", speaker="narrateur", glued=False, incises=None):
+    return Segment(
+        type=SegmentType(type), text=text, speaker=speaker,
+        glued_to_prev=glued, incises=incises or [])
+
+
+def _chap(index, segments):
+    return ChapterAnalysis(index=index, title=f"ch{index}", segments=segments)
+
+
+# --- Alignement --------------------------------------------------------------
+
+def test_alignement_1_1_textes_identiques():
+    ref = _chap(5, [_seg("Bonjour."), _seg("— Salut.", type="dialogue")])
+    hyp = _chap(5, [_seg("bonjour.  "), _seg("— salut.", type="dialogue")])  # espaces/casse
+    pairs = align(ref, hyp)
+    assert len(pairs) == 2
+    assert all(r is not None and h is not None for r, h in pairs)
+
+
+def test_alignement_segment_hypothese_en_trop():
+    ref = _chap(5, [_seg("A"), _seg("B")])
+    hyp = _chap(5, [_seg("A"), _seg("X"), _seg("B")])
+    cnt = _score_counts(ref, hyp, None)
+    # le segment "X" non couvert par la reference -> warning
+    assert any("sans correspondance" in w for w in cnt.warnings)
+
+
+# --- Attribution du locuteur -------------------------------------------------
+
+def test_speaker_parfait():
+    segs = [
+        _seg("narration"),
+        _seg("— Bonjour.", type="dialogue", speaker="Holden"),
+        _seg("— Salut.", type="dialogue", speaker="Kajri"),
+    ]
+    ref = _chap(5, [s.model_copy(deep=True) for s in segs])
+    hyp = _chap(5, [s.model_copy(deep=True) for s in segs])
+    score = score_chapter(ref, hyp)
+    assert score.speaker_acc_dialogue == 1.0
+    assert score.speaker_acc_all == 1.0
+    assert score.errors == []
+
+
+def test_speaker_avec_erreurs():
+    ref = _chap(5, [
+        _seg("— A.", type="dialogue", speaker="Holden"),
+        _seg("— B.", type="dialogue", speaker="Kajri"),
+    ])
+    hyp = _chap(5, [
+        _seg("— A.", type="dialogue", speaker="Holden"),
+        _seg("— B.", type="dialogue", speaker="Drummer"),  # faux
+    ])
+    score = score_chapter(ref, hyp)
+    assert score.speaker_acc_dialogue == 0.5
+    assert len(score.errors) == 1
+    assert score.errors[0].expected == "Kajri"
+    assert score.errors[0].got == "Drummer"
+    assert score.confusion["Kajri"]["Drummer"] == 1
+
+
+def test_speaker_normalisation_alias():
+    ref = _chap(5, [_seg("— A.", type="dialogue", speaker="Camina Drummer")])
+    hyp = _chap(5, [_seg("— A.", type="dialogue", speaker="Drummer")])  # alias
+    cast = Cast(characters=[Character(name="Camina Drummer", aliases=["Drummer"])])
+    # sans cast : compte comme une erreur (noms differents)
+    assert score_chapter(ref, hyp, None).speaker_acc_dialogue == 0.0
+    # avec cast : l'alias est resolu -> correct
+    assert score_chapter(ref, hyp, cast).speaker_acc_dialogue == 1.0
+
+
+# --- Incises -----------------------------------------------------------------
+
+def test_incises_exact_vs_overlap():
+    ref = _chap(5, [_seg("— A dit-il.", type="dialogue", speaker="X",
+                          incises=[Incise(start=4, end=11)])])
+    # span decale mais largement chevauchant -> overlap ok, exact non
+    hyp = _chap(5, [_seg("— A dit-il.", type="dialogue", speaker="X",
+                         incises=[Incise(start=4, end=10)])])
+    score = score_chapter(ref, hyp)
+    assert score.incise_exact_f1 < 1.0
+    assert score.incise_overlap_f1 == 1.0
+
+
+def test_incises_faux_positif_baisse_precision():
+    ref = _chap(5, [_seg("— A.", type="dialogue", speaker="X", incises=[])])
+    hyp = _chap(5, [_seg("— A.", type="dialogue", speaker="X",
+                         incises=[Incise(start=0, end=3)])])  # invente une incise
+    score = score_chapter(ref, hyp)
+    assert score.incise_overlap_p < 1.0
+    assert score.incise_overlap_r == 1.0  # rien a rappeler
+
+
+def test_incises_manque_baisse_rappel():
+    ref = _chap(5, [_seg("— A dit-il.", type="dialogue", speaker="X",
+                          incises=[Incise(start=4, end=11)])])
+    hyp = _chap(5, [_seg("— A dit-il.", type="dialogue", speaker="X", incises=[])])
+    score = score_chapter(ref, hyp)
+    assert score.incise_overlap_r < 1.0
+    assert score.incise_overlap_p == 1.0
+
+
+# --- Type / glued ------------------------------------------------------------
+
+def test_type_et_glued():
+    ref = _chap(5, [_seg("A", type="narration"), _seg("— B", type="dialogue", glued=True)])
+    hyp = _chap(5, [_seg("A", type="dialogue"), _seg("— B", type="dialogue", glued=False)])
+    score = score_chapter(ref, hyp)
+    assert score.type_acc == 0.5
+    assert score.glued_acc == 0.5
+
+
+# --- Agregat (micro-moyenne) -------------------------------------------------
+
+def test_aggregate_micro_moyenne():
+    # ch1 : 1 dialogue correct ; ch2 : 3 dialogues dont 1 faux
+    ref1 = _chap(1, [_seg("— A.", type="dialogue", speaker="X")])
+    hyp1 = _chap(1, [_seg("— A.", type="dialogue", speaker="X")])
+    ref2 = _chap(2, [
+        _seg("— B.", type="dialogue", speaker="X"),
+        _seg("— C.", type="dialogue", speaker="Y"),
+        _seg("— D.", type="dialogue", speaker="Z"),
+    ])
+    hyp2 = _chap(2, [
+        _seg("— B.", type="dialogue", speaker="X"),
+        _seg("— C.", type="dialogue", speaker="Y"),
+        _seg("— D.", type="dialogue", speaker="WRONG"),
+    ])
+    c1, c2 = _score_counts(ref1, hyp1, None), _score_counts(ref2, hyp2, None)
+    s1, s2 = score_chapter(ref1, hyp1), score_chapter(ref2, hyp2)
+    agg = aggregate([s1, s2], [c1, c2])
+    # micro : 3 corrects / 4 dialogues = 0.75 (et non moyenne de 1.0 et 0.666)
+    assert agg.n_dialogue == 4
+    assert abs(agg.speaker_acc_dialogue - 0.75) < 1e-9
+    assert agg.index == -1