Try to fix badly encoded json tex commands from the flash-lite calls

copienator/commands/gemini_for_enonce.py

@@ -22,6 +22,7 @@ from copienator import (
     workspace_from_args,
 )
 from copienator.platform import validate_windows_labels
+from copienator.latex_json import repair_latex_json
 from copienator.filesystem import staged_directory
 from copienator.utils import compile_to_pdf
 
@@ -69,7 +70,7 @@ class ExamContext(BaseModel):
 # --- Modèles pour la Requête 4 (Barèmes) ---
 class RubricItem(BaseModel):
     label: str = Field(description="Label exact de la question, à conserver sans traduction.")
-    rubric_content: str = Field(description="Barème détaillé sur 4 points : toutes les consignes, explications et justifications doivent être rédigées en français.")
+    rubric_content: str = Field(description="Barème détaillé sur 4 points, en français, avec formules LaTeX. Échapper chaque antislash LaTeX conformément à la syntaxe JSON.")
 
 class GroupRubrics(BaseModel):
     rubrics: list[RubricItem]
@@ -101,6 +102,27 @@ question fournie, cette liste contient un objet avec son `label` exact et
 son barème en français dans `rubric_content`.
 """
 
+LATEX_JSON_INSTRUCTION = r"""
+IMPORTANT — Échappement JSON du LaTeX :
+Dans les chaînes JSON, chaque antislash LaTeX doit être doublé.
+Exemple d’échappement dans un barème (adapte les clés au schéma demandé) :
+{"rubrics":[{"label":"1)","rubric_content":"1 point : $\\frac{1}{2}$ et $f \\circ g$.\n1 point : $x \\rightarrow 0$ et $x \\not=0$."}]}
+Ne confonds pas les commandes LaTeX avec les échappements JSON :
+\\frac, \\rightarrow, \\not, \\theta et \\begin sont des commandes LaTeX ;
+\n est réservé à un véritable saut de ligne du texte.
+"""
+
+
+def validate_rubric_content(content: str) -> str:
+    """Catch decoded control escapes; ordinary JSON newlines remain valid.
+
+    A misescaped LaTeX command starting with n is indistinguishable from
+    an intentional newline. Do not guess or rewrite those after decoding.
+    """
+    if any((ord(char) < 32 and char != "\n") or ord(char) == 127 for char in content):
+        raise ValueError("Rubric contains control characters, likely from unescaped LaTeX")
+    return content
+
 # --- Modèle fusionné (pour le reste du script) ---
 class QuestionItem(BaseModel):
     label: str
@@ -178,6 +200,12 @@ traduire, et conserve le LaTeX ainsi que les labels exacts.
 Renvoie le résultat dans la liste `contexts` de l'objet JSON attendu.
 """
 
+PROMPT_1 += LATEX_JSON_INSTRUCTION
+PROMPT_2 += LATEX_JSON_INSTRUCTION
+PROMPT_3 += LATEX_JSON_INSTRUCTION
+PROMPT_4 += LATEX_JSON_INSTRUCTION
+
+
 def find_file(folder: Path, base_name: str) -> Path | None:
     for ext in [".org", ".tex"]:
         path = folder / f"{base_name}{ext}"
@@ -202,10 +230,10 @@ def generate_rubrics(client, group_context_text: str) -> dict[str, str]:
             response_json_schema=GroupRubrics.model_json_schema(),
         ),
     )
-    rubrics = GroupRubrics.model_validate_json(response.text).rubrics
+    rubrics = GroupRubrics.model_validate_json(repair_latex_json(response.text)).rubrics
     if len({item.label for item in rubrics}) != len(rubrics):
         raise ValueError("Gemini returned duplicate rubric labels")
-    return {item.label: item.rubric_content for item in rubrics}
+    return {item.label: validate_rubric_content(item.rubric_content) for item in rubrics}
 
 
 def validate_groups(groups: list[list[str]], labels: list[str]) -> None:
@@ -364,7 +392,7 @@ def process_exam(
         print("Saving questions to cache...")
         atomic_write_text(cache_q_file, response_q_text)
 
-    questions_data = ExamQuestions.model_validate_json(response_q_text)
+    questions_data = ExamQuestions.model_validate_json(repair_latex_json(response_q_text))
 
     # ==========================================
     # REQUÊTE 2 : Extraction des Corrections
@@ -403,7 +431,7 @@ def process_exam(
         print("Saving solutions to cache...")
         atomic_write_text(cache_s_file, response_s_text)
 
-    solutions_data = ExamSolutions.model_validate_json(response_s_text)
+    solutions_data = ExamSolutions.model_validate_json(repair_latex_json(response_s_text))
 
     # ==========================================
     # REQUÊTE 3 : Extraction du Contexte (Notations, etc.)
@@ -440,7 +468,7 @@ def process_exam(
         print("Saving context to cache...")
         atomic_write_text(cache_c_file, response_c_text)
 
-    context_data = ExamContext.model_validate_json(response_c_text)
+    context_data = ExamContext.model_validate_json(repair_latex_json(response_c_text))
 
     # ==========================================
     # FUSION des trois résultats

copienator_gui/app.py

@@ -1551,8 +1551,7 @@ class CopienatorApp(tk.Tk):
         if self.current_step and self.current_step.optional:
             step_id = self.current_step.id
             self._mark_step("skipped")
-            if step_id.startswith("refaire_"):
-                self._move_selection_from(step_id, 1)
+            self._move_selection_from(step_id, 1)
 
     def _poll_runner(self) -> None:
         while True:

tests/test_gui_convenience.py

@@ -190,7 +190,9 @@ class GuiConvenienceTests(unittest.TestCase):
         command = self.app._make_command()
         self.assertEqual(command[-4:], ["crop-margins", str(source), "--workers", "5"])
         self.app._skip_step()
+        self.app.update()
         self.assertEqual(self.app.state_store.step("crop_blank_margins")["status"], "skipped")
+        self.assertNotEqual(self.app.current_step.id, "crop_blank_margins")
 
     def test_optional_answer_bottom_crop_uses_parallel_workers_and_can_be_skipped(self):
         answers = self.evaluation / "Copies" / "Copie01"
@@ -207,9 +209,11 @@ class GuiConvenienceTests(unittest.TestCase):
             ["crop-answer-bottoms", self.app._evaluation_arg(), "--workers", "5"],
         )
         self.app._skip_step()
+        self.app.update()
         self.assertEqual(
             self.app.state_store.step("crop_exercise_bottoms")["status"], "skipped"
         )
+        self.assertNotEqual(self.app.current_step.id, "crop_exercise_bottoms")
 
     def test_redo_targets_selected_copy_and_copies_runnable_command(self):
         for folder in ("Copies", "Copies Originales"):

tests/test_statement_personal_flow.py

@@ -134,6 +134,42 @@ class SelectiveGeminiTests(unittest.TestCase):
                 gemini.refine_existing(self.workspace, "persp", api_client=self.client)
             self.assertEqual(self.snapshot(), before)
 
+    def test_rubric_latex_and_line_breaks_survive_json_and_file_writes(self):
+        expected = (
+            r'Barème : $f \circ g \rightarrow \frac{1}{2} \not= 0$.'
+            '\n' + r'$\begin{matrix}a & b \\ c & d\end{matrix}$, $\theta$, $\beta$, $\nu$.'
+            '\nUne "justification" vaut 1 point.'
+        )
+        self.client.models.generate_content.side_effect = [
+            self.response({'rubrics': [{'label': label, 'rubric_content': expected} for label in group]})
+            for group in (self.labels[:2], self.labels[2:])
+        ]
+        gemini.refine_existing(self.workspace, 'persp', api_client=self.client)
+        for label in self.labels:
+            self.assertEqual((self.root / 'Persp' / label).read_bytes(),
+                             (label + '\n' + expected).encode('utf-8'))
+
+    def test_invalid_latex_escapes_are_repaired_without_retry(self):
+        for command in (r'\circ', r'\rightarrow', r'\frac', r'\not', r'\theta', r'\beta'):
+            with self.subTest(command=command):
+                # Deliberately malformed wire JSON: do not json.dumps the command.
+                broken = SimpleNamespace(text='{"rubrics":[{"label":"Ex 2","rubric_content":"$'
+                                         + command + '$"}]}')
+                self.client.models.generate_content.reset_mock()
+                self.client.models.generate_content.side_effect = [broken]
+                self.assertEqual(gemini.generate_rubrics(self.client, 'context'),
+                                 {'Ex 2': '$' + command + '$'})
+                self.assertEqual(self.client.models.generate_content.call_count, 1)
+
+    def test_unrepairable_control_character_preserves_all_existing_rubrics(self):
+        before = self.snapshot()
+        broken = self.response({'rubrics': [{'label': 'Ex 2', 'rubric_content': 'invalid \x01 control'}]})
+        self.client.models.generate_content.side_effect = [
+            self.rubric_response(self.labels[:2]), broken]
+        with self.assertRaisesRegex(ValueError, 'control characters'):
+            gemini.refine_existing(self.workspace, 'persp', api_client=self.client)
+        self.assertEqual(self.snapshot(), before)
+
     def test_cli_dispatches_selective_modes_without_full_extraction(self):
         with patch.object(gemini, "refine_existing", return_value=ExitCode.SUCCESS) as refine, patch.object(
             gemini, "process_exam"