32k pixel as hard group limit, better manual resolution support

copienator/commands/annotating_by_label.py

@@ -27,7 +27,7 @@ from copienator.commands import annotating, annotating_with_checks
 from copienator.filesystem import staged_directory
 from copienator.utils import natural_key
 
-MAX_HEIGHT_PX = 25000
+MAX_HEIGHT_PX = 32000
 
 
 def render_item(item):
@@ -63,6 +63,11 @@ def save_batch(batch, prefix, group_id, output_root: Path) -> None:
     print(f"Generating group PDF: {prefix} G{group_id} ({len(batch)} elements)")
     max_width = max(item[2].width for item in batch)
     total_height = sum(item[2].height for item in batch)
+    if total_height > MAX_HEIGHT_PX:
+        raise CliError(
+            f"{prefix} G{group_id} would be {total_height} px high; "
+            f"the maximum is {MAX_HEIGHT_PX} px."
+        )
     concatenated = Image.new("RGB", (max_width, total_height), "white")
     draw = ImageDraw.Draw(concatenated)
     checkbox_map: list[dict[str, Any]] = []
@@ -268,32 +273,44 @@ def _existing_group_state(
 
 
 def split_batches(rendered):
-    def split(maximum_height: float):
     batches = []
     current = []
     current_height = 0
-        previous_student = None
-        for item in rendered:
-            student_id = item[0]
-            image_height = item[2].height
-            if (
-                current
-                and current_height + image_height > maximum_height
-                and student_id != previous_student
-            ):
+
+    def flush() -> None:
+        nonlocal current, current_height
+        if current:
             batches.append(current)
             current = []
             current_height = 0
-            current.append(item)
-            current_height += image_height
-            previous_student = student_id
-        if current:
-            batches.append(current)
-        return batches
 
-    strict = split(MAX_HEIGHT_PX)
-    relaxed = split(1.1 * MAX_HEIGHT_PX)
-    return relaxed if len(relaxed) < len(strict) else strict
+    index = 0
+    while index < len(rendered):
+        student_id = rendered[index][0]
+        end = index + 1
+        while end < len(rendered) and rendered[end][0] == student_id:
+            end += 1
+        student_items = rendered[index:end]
+        student_height = sum(item[2].height for item in student_items)
+
+        if student_height > MAX_HEIGHT_PX:
+            labels = ", ".join(item[1] for item in student_items)
+            raise CliError(
+                f"Copie{student_id} ({labels}) is {student_height} px high; "
+                f"one student's labels cannot be split across groups and the "
+                f"maximum group height is {MAX_HEIGHT_PX} px."
+            )
+
+        # Consider the student's complete block before adding its first image.
+        # Otherwise later images from that same student could overflow the batch.
+        if current and current_height + student_height > MAX_HEIGHT_PX:
+            flush()
+        current.extend(student_items)
+        current_height += student_height
+        index = end
+
+    flush()
+    return batches
 
 
 def _generate_groups(

copienator/commands/correction.py

@@ -1026,7 +1026,9 @@ def run_configured(args: argparse.Namespace) -> ExitCode:
         manual_path = INPUT_DIR / "manual_resolutions.txt"
         atomic_write_text(
             manual_path,
-            "###  Use -> x>, -x, ss, sx, xx, xs, c{43}1>, c{43}2x\n"
+            "### Use -> x>, -x, ss, sx, xx, xs, c{43}1>, c{43}2x; "
+            "delete a line to dismiss its conflict and keep its PDFs unchanged, "
+            "then run resolve-manual (even if all lines were deleted).\n"
             + "\n".join(unresolved_delayed)
             + "\n",
         )

copienator/commands/cutleft.py

@@ -195,6 +195,7 @@ class ImageReviewer:
         self.root.bind("N", lambda _event: self.on_shift(2 * CROP_SHIFT_STEP))
         self.root.bind("t", lambda _event: self.on_shift(-CROP_SHIFT_STEP))
         self.root.bind("l", lambda _event: self.on_enlarge(CROP_WIDTH_STEP))
+        self.root.bind("L", lambda _event: self.on_enlarge(-CROP_WIDTH_STEP))
         self.root.bind("1", lambda _event: self.on_set_max_pages(1))
 
         Thread(target=self.prefetch_worker, daemon=True).start()
@@ -290,7 +291,7 @@ class ImageReviewer:
                 f"Files: {schema['number_of_files']} | "
                 f"Cols: {schema['columns_per_file']}\n"
                 "Enter: Save and next | s: flag error and skip | n: +50 | N: +100 | t: -50 | "
-                "l: widen by 50 | 1: use full pages"
+                "l: widen by 50 | Shift+l: undo widening by 50 | 1: use full pages"
             ),
             fg="black",
         )
@@ -305,7 +306,7 @@ class ImageReviewer:
     def on_enlarge(self, amount: int) -> None:
         if self.is_processing:
             return
-        self.current_width_offset += amount
+        self.current_width_offset = max(0, self.current_width_offset + amount)
         print(f"Applying extra width: {self.current_width_offset}")
         self.trigger_processing(self.files[self.index], self.current_shift)
 

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/commands/gemini_for_labels.py

@@ -4,9 +4,10 @@ import argparse
 import re
 import time
 import typing
+from threading import Lock
 from collections import defaultdict
 from collections.abc import Callable, Sequence
-from concurrent.futures import ThreadPoolExecutor
+from concurrent.futures import ThreadPoolExecutor, wait
 from pathlib import Path
 
 from google import genai
@@ -271,6 +272,10 @@ def generate_request(
     ]
 
     generate_content_config = types.GenerateContentConfig(
+        http_options=types.HttpOptions(
+            timeout=120_000,
+            retry_options=types.HttpRetryOptions(attempts=1),
+        ),
         automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
         temperature=1.0,
         top_p=0.95,
@@ -282,9 +287,17 @@ def generate_request(
     return (contents, generate_content_config)
 
 TARGET_INTERVAL = 3.5
+MAX_ATTEMPTS = 3
+PROGRESS_INTERVAL = 30
 Sleep = Callable[[float], None]
 
 
+class IncompleteCopyError(Exception):
+    def __init__(self, images: list[Path]):
+        self.images = images
+        super().__init__(f"Label detection incomplete for {images[0].name}")
+
+
 def selected_images(
     workspace: EvaluationWorkspace,
     targets: list[Path],
@@ -395,11 +408,12 @@ def process_copy_group(
     overwrite: bool,
     sleep: Sleep = time.sleep,
     target_interval: float = TARGET_INTERVAL,
+    progress: Callable[[str, str, float], None] | None = None,
 ) -> int:
     """Process one student's image parts sequentially to preserve context."""
     accumulated_labels: list[str] = []
     generated = 0
-    for image_file in files:
+    for image_index, image_file in enumerate(files):
         started = time.monotonic()
         output_json = workspace.copies_dir / f"{image_file.stem}.json"
         if output_json.exists() and not overwrite:
@@ -414,9 +428,13 @@ def process_copy_group(
         attempt = 0
         label_retry_count = 0
         wrong_labels: list[str] = []
-        while True:
+        while attempt < MAX_ATTEMPTS:
+            if progress:
+                progress(group_key, f"{image_file.name}, attempt {attempt + 1}/{MAX_ATTEMPTS}", started)
             if attempt > 0:
+                print(f"[{group_key}] Retrying {image_file.name} in {10 * attempt}s...")
                 sleep(10 * attempt)
+            print(f"[{group_key}] {image_file.name}: attempt {attempt + 1}/{MAX_ATTEMPTS} (120s timeout).")
             try:
                 text_context_file, text_context = closest_text_context(
                     workspace, wrong_labels
@@ -486,15 +504,23 @@ def process_copy_group(
                 atomic_write_json(output_json, annotation.model_dump())
                 accumulated_labels.extend(box.label for box in annotation.list)
                 generated += 1
+                print(f"[{group_key}] Saved {image_file.name} after {time.monotonic() - started:.0f}s.")
                 break
             except KeyboardInterrupt:
                 raise
             except Exception as exc:  # noqa: BLE001 - remote API retry boundary
                 print(
-                    f"Error processing {image_file.name}: {exc}\n"
-                    "\tIt will be retried."
+                    f"Error processing {image_file.name} "
+                    f"(attempt {attempt + 1}/{MAX_ATTEMPTS}): {type(exc).__name__}: {exc}"
                 )
                 attempt += 1
+        else:
+            # Later pages need the context of this page; defer them until resume.
+            incomplete = [path for path in files[image_index:]
+                          if overwrite or not (workspace.copies_dir / f"{path.stem}.json").exists()]
+            print(f"[{group_key}] Failed after {MAX_ATTEMPTS} attempts; "
+                  "deferring remaining pages of this copy until the next run.")
+            raise IncompleteCopyError(incomplete)
         sleep(max(0.0, target_interval - (time.monotonic() - started)))
     return generated
 
@@ -528,8 +554,17 @@ def run(
         client = genai.Client(api_key=api_key)
 
     groups = group_images(images)
+    active: dict[str, tuple[str, float]] = {}
+    progress_lock = Lock()
+    incomplete: list[Path] = []
+    run_started = time.monotonic()
+
+    def report_progress(group: str, description: str, started: float) -> None:
+        with progress_lock:
+            active[group] = (description, started)
+
     with ThreadPoolExecutor(max_workers=max_workers) as executor:
-        futures = [
+        futures = {
             executor.submit(
                 process_copy_group,
                 workspace,
@@ -542,12 +577,33 @@ def run(
                 valid_names=valid_names,
                 overwrite=overwrite,
                 sleep=sleep,
-            )
+                progress=report_progress,
+            ): group_key
             for group_key, files in groups.items()
-        ]
-        for future in futures:
+        }
+        pending = set(futures)
+        while pending:
+            done, pending = wait(pending, timeout=PROGRESS_INTERVAL)
+            for future in done:
+                with progress_lock:
+                    active.pop(futures[future], None)
+                try:
                     future.result()
-    return ExitCode.PARTIAL if warnings else ExitCode.SUCCESS
+                except IncompleteCopyError as exc:
+                    incomplete.extend(exc.images)
+            if pending:
+                now = time.monotonic()
+                with progress_lock:
+                    snapshot = list(active.items())
+                print(f"Label detection: {now - run_started:.0f}s elapsed; "
+                      f"{len(pending)} copies outstanding ({len(pending) - len(snapshot)} queued).")
+                for group, (description, started) in snapshot:
+                    print(f"  [{group}] {description}: {now - started:.0f}s elapsed for image.")
+    if incomplete:
+        print("Incomplete images (rerun label detection to resume):")
+        for path in sorted(incomplete):
+            print(f"  {path.name}")
+    return ExitCode.PARTIAL if warnings or incomplete else ExitCode.SUCCESS
 
 
 def build_parser() -> argparse.ArgumentParser:

copienator/commands/page_splitter.py

@@ -187,7 +187,6 @@ class PDFPreviewer:
         action_map = {
             "move_left": self.move_line_left,
             "move_right": self.move_line_right,
-            "confirm_next": self.confirm_and_next_page,
             "rotate_page": self.rotate_page,
             "rotate_all_pages": self.rotate_all_pages,
             "rotate_all_files": self.rotate_all_files,
@@ -209,7 +208,6 @@ class PDFPreviewer:
 
         # self.master.bind("<Left>", self.move_line_left)
         # self.master.bind("<Right>", self.move_line_right)
-        # self.master.bind("<Return>", self.confirm_and_next_page)
         # self.master.bind("c", self.rotate_page)
         # self.master.bind("C", self.rotate_all_pages)
         # self.master.bind(",", self.rotate_all_files)
@@ -260,7 +258,9 @@ class PDFPreviewer:
             return
 
         page = self.doc.load_page(self.current_page_index)
-        self.page_label.config(text=f"Page {self.current_page_index + 1} of {len(self.doc)}")
+        self.page_label.config(
+            text=f"{self.pdf_path.name} — Page {self.current_page_index + 1} of {len(self.doc)}"
+        )
 
         # --- Calculate Scaling ---
         canvas_width = self.canvas.winfo_width()

copienator/commands/resolve_manual.py

@@ -132,6 +132,18 @@ def set_suffix_and_clean_error(
                 result["error"] = error
 
 
+def pending_manual_conflicts(results: dict[str, Any]) -> set[tuple[str, str, str]]:
+    """Return the copy/source/target conflicts listed by correction's manual file."""
+    return {
+        (item["id"], label, target)
+        for label, batches in results.items()
+        for batch in batches
+        for item in batch
+        for kind, target in item.get("result", {}).get("delayed", [])
+        if kind in {"wrong-label", "add-label"}
+    }
+
+
 def get_actual_pdf(copies_dir: Path, copy_id: str, label: str) -> Path:
     base = copies_dir / f"Copie{copy_id}" / f"{label}.pdf"
     for candidate in (
@@ -157,13 +169,7 @@ def _validate_pdf_inputs(
     missing: set[Path] = set()
     for instruction in instructions:
         source = initial_paths[(instruction.copy_id, instruction.old_label)]
-        destination = initial_paths[(instruction.copy_id, instruction.new_label)]
-        if instruction.should_merge:
-            if not source.exists():
-                missing.add(source)
-            if not destination.exists():
-                missing.add(destination)
-        elif instruction.should_copy and not source.exists():
+        if (instruction.should_merge or instruction.should_copy) and not source.exists():
             missing.add(source)
     if missing:
         rendered = ", ".join(str(path) for path in sorted(missing))
@@ -178,6 +184,10 @@ def resolve_manual(workspace: EvaluationWorkspace) -> ExitCode:
         raise CliError("correction.json must contain a JSON object")
     results: dict[str, Any] = loaded
     instructions = parse_instructions(workspace.manual_resolutions_file)
+    retained_conflicts = {
+        (item.copy_id, item.old_label, item.new_label) for item in instructions
+    }
+    dismissed_conflicts = pending_manual_conflicts(results) - retained_conflicts
     cut_sources = [(item.copy_id, item.old_label) for item in instructions if item.cut]
     if len(set(cut_sources)) != len(cut_sources):
         raise CliError("Une seule coupe par label source est autorisée dans une résolution.")
@@ -226,8 +236,10 @@ def resolve_manual(workspace: EvaluationWorkspace) -> ExitCode:
                 try:
                     if instruction.pipe_first:
                         writer.append(source)
+                        if destination.exists():
                             writer.append(destination)
                     else:
+                        if destination.exists():
                             writer.append(destination)
                         writer.append(source)
                     writer.write(temp_output)
@@ -285,6 +297,9 @@ def resolve_manual(workspace: EvaluationWorkspace) -> ExitCode:
         if temporary not in used_temps:
             temporary.unlink(missing_ok=True)
 
+    for copy_id, label, target in dismissed_conflicts:
+        # A removed instruction acknowledges the conflict without touching PDFs.
+        set_suffix_and_clean_error(results, copy_id, label, None, target)
     atomic_write_json(workspace.correction_file, results)
     refaire_tasks = [[copy_name, labels] for copy_name, labels in refaire_by_copy.items()]
     if refaire_tasks:
@@ -292,6 +307,8 @@ def resolve_manual(workspace: EvaluationWorkspace) -> ExitCode:
     workspace.manual_resolutions_file.unlink()
 
     print("Manual resolutions successfully applied.")
+    if dismissed_conflicts:
+        print(f"{len(dismissed_conflicts)} removed conflict(s) dismissed; their PDFs were kept unchanged.")
     if refaire_tasks:
         print(
             f"File {workspace.refaire_file.name} generated. Run "

copienator/commands/submit_batches.py

@@ -19,11 +19,7 @@ from copienator import (
 )
 
 
-def run(workspace: EvaluationWorkspace, *, client=None) -> ExitCode:
-    if client is None:
-        if not config.API_KEY:
-            raise CliError("GEMINI_API_KEY is not configured")
-        client = genai.Client(api_key=config.API_KEY)
+def run(workspace: EvaluationWorkspace, *, client=None, overwrite: bool = False) -> ExitCode:
     batches = (
         (
             "flash",
@@ -45,16 +41,32 @@ def run(workspace: EvaluationWorkspace, *, client=None) -> ExitCode:
     }
     if workspace.batch_jobs_file.is_file():
         previous = read_json(workspace.batch_jobs_file)
-        if isinstance(previous, dict) and isinstance(previous.get("jobs"), dict):
+        if not isinstance(previous, dict) or not isinstance(previous.get("jobs"), dict):
+            raise CliError(f"Invalid batch manifest: {workspace.batch_jobs_file}")
         manifest["jobs"] = previous["jobs"]
     started = 0
+    skipped = 0
     for tier, file_path, model_id, display_name in batches:
+        existing = manifest["jobs"].get(tier)
+        if tier in manifest["jobs"]:
+            if (not isinstance(existing, dict)
+                    or not isinstance(existing.get("name"), str)
+                    or not existing["name"].strip()):
+                raise CliError(f"Invalid batch job in {workspace.batch_jobs_file}: {tier}")
+            if not overwrite:
+                print(f"Skipping {tier}: already submitted as {existing['name']}.")
+                skipped += 1
+                continue
         if not file_path.is_file():
             print(f"Skipping {model_id}: {file_path.name} does not exist.")
             continue
         if file_path.stat().st_size == 0:
             print(f"Skipping {model_id}: {file_path.name} is empty.")
             continue
+        if client is None:
+            if not config.API_KEY:
+                raise CliError("GEMINI_API_KEY is not configured")
+            client = genai.Client(api_key=config.API_KEY)
         print(f"Uploading {file_path.name} for model {model_id}...")
         uploaded = client.files.upload(
             file=str(file_path),
@@ -77,19 +89,26 @@ def run(workspace: EvaluationWorkspace, *, client=None) -> ExitCode:
         }
         atomic_write_json(workspace.batch_jobs_file, manifest)
         print(f"Started batch job: {job.name}")
-    if not started:
+    if not started and not skipped:
         print("No non-empty batch request files were found.")
         return ExitCode.PARTIAL
     return ExitCode.SUCCESS
 
 
 def build_parser() -> argparse.ArgumentParser:
-    return evaluation_parser("Upload correction JSONL files and start Gemini batches")
+    parser = evaluation_parser("Upload correction JSONL files and start Gemini batches")
+    parser.add_argument(
+        "--overwrite", action="store_true",
+        help="Submit new jobs even for recorded tiers, replacing their saved job IDs",
+    )
+    return parser
 
 
 def main(argv: Sequence[str] | None = None) -> int:
     parser = build_parser()
-    return execute(parser, argv, lambda args: run(workspace_from_args(args)))
+    return execute(parser, argv, lambda args: run(
+        workspace_from_args(args), overwrite=args.overwrite,
+    ))
 
 
 if __name__ == "__main__":

copienator/utils.py

@@ -177,21 +177,23 @@ def compile_to_pdf(text, output_pdf_path):
                     text=True,
                     check=False)
 
+            generated_pdf = os.path.join(temp_dir, pdf_filename)
+            pdf_generated = os.path.exists(generated_pdf)
+            if pdf_generated:
+                shutil.move(generated_pdf, output_pdf_path)
+
             if result.returncode != 0:
                 error_lines = [
                     line.strip() for line in result.stdout.splitlines()
                     if line.lstrip().startswith("!")
                 ]
                 detail = f": {error_lines[0]}" if error_lines else ""
+                pdf_status = "a PDF was generated" if pdf_generated else "no PDF was generated"
                 print(
-                    f"Warning: LaTeX compilation failed for {output_pdf_path} "
-                    f"(exit code {result.returncode}){detail}"
+                    f"Warning: LaTeX compilation reported errors for {output_pdf_path} "
+                    f"(exit code {result.returncode}; {pdf_status}){detail}"
                 )
-
-            generated_pdf = os.path.join(temp_dir, pdf_filename)
-            if os.path.exists(generated_pdf):
-                shutil.move(generated_pdf, output_pdf_path)
-            else:
+            elif not pdf_generated:
                 print(f"Warning: LaTeX compilation produced no PDF for {output_pdf_path}")
         except Exception as e:
             print(f"Compilation error for {output_pdf_path}: {e}")

copienator_gui/app.py

@@ -10,7 +10,8 @@ from pathlib import Path
 from tkinter import filedialog, messagebox, ttk
 from typing import Any
 
-from copienator import CliError, EvaluationWorkspace, ExitCode, atomic_write_json
+from copienator import CliError, EvaluationWorkspace, ExitCode, atomic_write_json, read_json
+from copienator.commands.resolve_manual import pending_manual_conflicts
 from copienator.copy_errors import copy_errors
 from copienator.filesystem import staged_files
 from copienator.platform import (
@@ -244,7 +245,14 @@ def has_manual_conflicts(path: Path) -> bool:
         # If the file cannot be inspected, keep the step visible rather than
         # silently claiming that there is nothing to resolve.
         return True
-    return any(line.strip() and not line.lstrip().startswith("###") for line in lines)
+    if any(line.strip() and not line.lstrip().startswith("###") for line in lines):
+        return True
+    # An emptied manual file still needs execution to acknowledge deletions.
+    try:
+        results = read_json(path.parent / "correction.json", default={})
+        return bool(pending_manual_conflicts(results))
+    except (OSError, ValueError, TypeError, KeyError, AttributeError):
+        return True
 
 
 class CopienatorApp(tk.Tk):
@@ -591,7 +599,10 @@ class CopienatorApp(tk.Tk):
         opened = {self.tree.item(item, "text"): self.tree.item(item, "open") for item in self.tree.get_children()}
         self.tree.delete(*self.tree.get_children())
         section_items: dict[str, str] = {}
+        show_batches = self._has_batch_correction()
         for step in self.steps:
+            if step.skip_for_live_correction and not show_batches:
+                continue
             if step.section not in section_items:
                 section_id = f"section:{len(section_items)}"
                 section_items[step.section] = section_id
@@ -607,6 +618,25 @@ class CopienatorApp(tk.Tk):
             )
         if selected and self.tree.exists(selected):
             self.tree.selection_set(selected)
+        elif selected and self.tree.exists("correction"):
+            self.tree.selection_set("correction")
+
+    def _has_batch_correction(self) -> bool:
+        evaluation = self.state_store.evaluation
+        if evaluation is None:
+            return False
+        correction = self.state_store.step("correction")
+        if correction.get("batch_started") or correction.get("last_run_variant") in {"batch", "hybrid", "integrate"}:
+            return True
+        for name in ("batch_requests_flash.jsonl", "batch_requests_pro.jsonl",
+                     "batch_jobs.json", "batched_correction_result.jsonl"):
+            path = evaluation / name
+            try:
+                if path.is_file() and path.stat().st_size > 0:
+                    return True
+            except OSError:
+                continue
+        return False
 
     def _step_status(self, step: StepDefinition) -> str:
         if self.state_store.evaluation:
@@ -721,16 +751,6 @@ class CopienatorApp(tk.Tk):
         # trigger the same automatic behavior twice.
         self.state_store.update_step(step.id, visited=True)
 
-        if step.skip_for_live_correction:
-            correction = self.state_store.step("correction")
-            if correction.get("variant", "live") == "live":
-                self.after_idle(
-                    lambda step_id=step.id: self._automatic_skip(
-                        step_id, "correction immédiate sélectionnée"
-                    )
-                )
-                return
-
         if step.skip_without_manual_conflicts:
             evaluation = self.evaluation
             conflicts = evaluation / "manual_resolutions.txt" if evaluation else None
@@ -1217,7 +1237,9 @@ class CopienatorApp(tk.Tk):
 
     def _progression_ids(self, step_id: str) -> list[str]:
         redo = step_id.startswith("refaire_")
-        return [step.id for step in self.steps if step.id.startswith("refaire_") == redo]
+        show_batches = self._has_batch_correction()
+        return [step.id for step in self.steps if step.id.startswith("refaire_") == redo
+                and (not step.skip_for_live_correction or show_batches)]
 
     def _refaire_selection(self) -> list[list]:
         evaluation = self.evaluation
@@ -1445,6 +1467,10 @@ class CopienatorApp(tk.Tk):
         self._save_current_form()
         run_values = self._values()
         ordered_ids = self._progression_ids(step.id)
+        if step.id == "correction" and variant.id == "integrate":
+            # Integration consumes the retrieved batches; it does not invalidate them.
+            ordered_ids = [step_id for step_id in ordered_ids
+                           if not self.step_by_id[step_id].skip_for_live_correction]
         self.state_store.invalidate_after(ordered_ids, step.id)
         self.state_store.update_step(
             step.id,
@@ -1473,6 +1499,8 @@ class CopienatorApp(tk.Tk):
         self._append_console(f"\n$ {command_display(command)}\n")
         try:
             self.runner.start(command, self.repository, environment, log_path)
+            if step.id == "correction" and variant.id in {"batch", "hybrid"}:
+                self.state_store.update_step("correction", batch_started=True)
         except (OSError, RuntimeError) as exc:
             self.state_store.update_step(step.id, status="failed")
             self.state_store.add_history(
@@ -1551,7 +1579,6 @@ 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)
 
     def _poll_runner(self) -> None:
@@ -1615,15 +1642,18 @@ class CopienatorApp(tk.Tk):
         self._populate_tree()
         message = f"{evaluation.name} : tous les résultats batch sont prêts à récupérer."
         self.info_var.set(message)
-        try:
-            notify_desktop("Copienator — batchs terminés", message)
-        except (OSError, RuntimeError) as exc:
-            self.bell()
-            self._append_console(f"Notification de bureau indisponible : {exc}\n{message}\n")
+        self._notify_completion("Copienator — batchs terminés", message)
         if (self.current_step and self.current_step.id == "batch_status"
                 and not self.active_step_id and not self.runner.running):
             self._move_selection_from("batch_status", 1)
 
+    def _notify_completion(self, title: str, message: str) -> None:
+        try:
+            notify_desktop(title, message)
+        except (OSError, RuntimeError) as exc:
+            self.bell()
+            self._append_console(f"Notification de bureau indisponible : {exc}\n{message}\n")
+
     def _finish_process(self, return_code: int, interrupted: bool) -> None:
         step_id = self.active_step_id
         if not step_id:
@@ -1710,12 +1740,28 @@ class CopienatorApp(tk.Tk):
                 "sur « Marquer terminée »."
             )
             return
+        notified_completions = {
+            "labels": "détection des labels terminée",
+            "correction": "étape de correction terminée",
+            "annotation": "génération des copies annotées terminée",
+            "read_annotations": "lecture des annotations manuscrites terminée",
+        }
+        if step_id in notified_completions and self.state_store.evaluation:
+            completion = notified_completions[step_id]
+            outcome = STATUS_LABELS[status].lower()
+            self._notify_completion(
+                f"Copienator — {completion} ({outcome})",
+                f"{self.state_store.evaluation.name} : {completion} — {outcome} "
+                f"(code {return_code}).",
+            )
         if status == "success":
             if step_id == "batch_status" and self.batch_monitor.active:
                 self.batch_monitor.stop()
                 self._batch_results_ready()
                 return
-            if step_id == "manual_resolution":
+            if step_id == "fetch_batches":
+                self.after_idle(self._select_batch_integration)
+            elif step_id == "manual_resolution":
                 self.after_idle(self._select_refaire_correction)
             else:
                 self.after_idle(
@@ -1813,6 +1859,15 @@ class CopienatorApp(tk.Tk):
         self.tree.selection_set(ids[target])
         self.tree.see(ids[target])
 
+    def _select_batch_integration(self) -> None:
+        """Prepare integration after batch results have been retrieved."""
+        if self.runner.running or self.active_step_id or not self.state_store.evaluation:
+            return
+        self._save_current_form()
+        self.state_store.update_step("correction", variant="integrate")
+        self.tree.selection_set("correction")
+        self.tree.see("correction")
+
     def _select_refaire_correction(self) -> None:
         """Return from manual conflict resolution to the main redo correction."""
         if self.runner.running or self.active_step_id or not self.state_store.evaluation:

copienator_gui/manual_resolution.py

@@ -29,7 +29,8 @@ c{43}2x : couper la source à 43 %, conserver la fin et remplacer la cible par l
 Cut permet de placer la coupe, avec accrochage entre les pages. Entrée ferme l’aperçu et affiche la commande à recopier ci-dessus : aucun fichier n’est modifié. Une séparation entre pages conserve les pages entières. La source complète est archivée en _old ; les deux labels modifiés sont générés en _new et ajoutés à refaire.json.
 
 Pour les fusions : « Source x> Cible| » place la cible avant la source ; « Source x> |Cible » place la source avant la cible. Même règle avec -> et c{…}1>/c{…}2> (pour la partie transférée) ; sans |, la cible vient en premier.
-Vous pouvez changer l’opérateur, déplacer |, corriger les labels ou retirer une instruction. Les lignes vides et celles commençant par ### sont ignorées. Retirer/commenter une ligne ne résout pas son conflit.
+Si le PDF cible n’existe pas, il est considéré comme vide : la source (ou la partie transférée après une coupe) devient le nouveau PDF cible.
+Vous pouvez changer l’opérateur, déplacer |, corriger les labels ou retirer une instruction. Les lignes vides et celles commençant par ### sont ignorées. Retirer/commenter une ligne signifie conserver les PDF sans modification et considérer ce conflit comme résolu après Exécuter. Exécutez cette étape même si vous avez retiré toutes les instructions : ces conflits ne seront plus proposés lors de la reprise avec refaire.json.
 
 Enregistrez dans l’éditeur, puis cliquez sur Recharger et enfin sur Exécuter. Seul le fichier enregistré est appliqué. L’archivage utilise le suffixe _old ; les PDF créés utilisent _new. Après succès, manual_resolutions.txt est supprimé et correction.json est mis à jour. Si des PDF sont créés, refaire.json est généré : relancez la correction avec --refaire."""
 
@@ -46,7 +47,10 @@ class ManualResolutionPanel(ttk.Frame):
             actions, text="Ouvrir dans un éditeur de texte", command=self.open_editor
         )
         self.editor_button.pack(side="left")
-        ttk.Button(actions, text="Recharger", command=self.reload).pack(side="left", padx=6)
+        self.reload_button = ttk.Button(
+            actions, text="Recharger", command=self.reload
+        )
+        self.reload_button.pack(side="left", padx=6)
         self.cut_result = tk.StringVar()
         result = ttk.Frame(self)
         result.pack(fill="x", pady=4)
@@ -116,10 +120,10 @@ class ManualResolutionPanel(ttk.Frame):
                 else:
                     if instructions:
                         instruction = instructions[0]
-                        for title, label in (("PDF source", instruction.old_label),
-                                             ("PDF cible", instruction.new_label)):
+                        for title, label in (("Source", instruction.old_label),
+                                             ("Cible", instruction.new_label)):
                             button = ttk.Button(
-                                self.text, text=title,
+                                self.text, text=f"{title} : {label}",
                                 command=lambda label=label, copy_id=instruction.copy_id, root=path.parent: self.open_pdf(root, copy_id, label),
                             )
                             self.pdf_buttons.append(button)

default_config.py

@@ -44,7 +44,6 @@ MODEL_PRO_ID = "gemini-3.8-flash"
 PAGE_SPLITTER_KB = {
     "move_left": "<Left>",
     "move_right": "<Right>",
-    "confirm_next": "<Return>",
     "rotate_page": "r",
     "rotate_all_pages": "R",
     "rotate_all_files": ",",

tests/test_gui_convenience.py

@@ -133,7 +133,8 @@ class GuiConvenienceTests(unittest.TestCase):
         self.assertNotEqual(self.app.current_step.id, "giving_names")
 
     def test_batch_status_waits_until_all_jobs_are_ready(self):
-        self.app.state_store.update_step("correction", variant="batch")
+        self.app.state_store.update_step("correction", variant="batch", last_run_variant="batch")
+        self.app._populate_tree()
         self.app.state_store.update_step("batch_status", visited=True)
         self.app.tree.selection_set("batch_status")
         self.app.update()
@@ -149,8 +150,59 @@ class GuiConvenienceTests(unittest.TestCase):
         self.app.update()
         self.assertEqual(self.app.current_step.id, "fetch_batches")
 
+    def test_batch_retrieval_success_selects_integration_without_starting_it(self):
+        self.app.state_store.update_step("correction", variant="batch", last_run_variant="batch")
+        self.app._populate_tree()
+        self.app.tree.selection_set("fetch_batches")
+        self.app.update()
+        self.app.active_step_id = "fetch_batches"
+        with patch.object(self.app.runner, "start") as start:
+            self.app._finish_process(0, False)
+            self.app.update()
+        start.assert_not_called()
+        self.assertEqual(self.app.current_step.id, "correction")
+        self.assertEqual(self.app.variant_var.get(), "integrate")
+        self.assertEqual(self.app.state_store.step("correction")["variant"], "integrate")
+        self.assertIn("--deal-with-batched", self.app._make_command())
+
+    def test_integration_preserves_batch_statuses_but_invalidates_downstream_steps(self):
+        for variant in ("integrate", "batch"):
+            with self.subTest(variant=variant):
+                self.app.active_step_id = None
+                self.app.current_step = None
+                self.app.state_store.update_step("correction", variant=variant, batch_started=True)
+                statuses = {"submit_batches": "success", "batch_status": "detected",
+                            "fetch_batches": "success", "annotation": "success"}
+                for step_id, status in statuses.items():
+                    self.app.state_store.update_step(step_id, status=status)
+                self.app._populate_tree()
+                self.app.tree.selection_set("correction")
+                self.app.update()
+                with patch.object(self.app.runner, "start"), \
+                        patch.object(self.app, "_missing_requirements", return_value=[]):
+                    self.app._run_current_step()
+                self.assertEqual(self.app.active_step_id, "correction")
+                for step_id, status in statuses.items():
+                    expected = status if variant == "integrate" and step_id != "annotation" else "stale"
+                    self.assertEqual(self.app.state_store.step(step_id)["status"], expected)
+        self.app.active_step_id = None
+
+    def test_unsuccessful_batch_retrieval_stays_on_retrieval(self):
+        self.app.state_store.update_step("correction", variant="batch", last_run_variant="batch")
+        self.app._populate_tree()
+        self.app.tree.selection_set("fetch_batches")
+        self.app.update()
+        for code, interrupted in ((4, False), (1, False), (0, True)):
+            with self.subTest(code=code, interrupted=interrupted):
+                self.app.active_step_id = "fetch_batches"
+                self.app._finish_process(code, interrupted)
+                self.app.update()
+                self.assertEqual(self.app.current_step.id, "fetch_batches")
+                self.assertEqual(self.app.state_store.step("correction")["variant"], "batch")
+
     def test_batch_watch_buttons_use_loaded_evaluation_and_stop_on_reload(self):
-        self.app.state_store.update_step("correction", variant="batch")
+        self.app.state_store.update_step("correction", variant="batch", last_run_variant="batch")
+        self.app._populate_tree()
         self.app.tree.selection_set("batch_status")
         self.app.update()
         with patch.object(self.app.batch_monitor, "start") as start:
@@ -175,6 +227,97 @@ class GuiConvenienceTests(unittest.TestCase):
         self.assertEqual(self.app.current_step.id, "inputs")
         self.assertEqual(self.app.state_store.step("batch_status")["status"], "success")
 
+    def test_batch_steps_hidden_until_a_correction_is_started(self):
+        batch_ids = ("submit_batches", "batch_status", "fetch_batches")
+        self.app.state_store.update_step("correction", variant="batch")
+        self.app._populate_tree()
+        for step_id in batch_ids:
+            self.assertFalse(self.app.tree.exists(step_id))
+            self.assertNotIn(step_id, self.app._progression_ids("correction"))
+        self.app.state_store.update_step("correction", last_run_variant="hybrid", status="running")
+        self.app._populate_tree()
+        for step_id in batch_ids:
+            self.assertTrue(self.app.tree.exists(step_id))
+            self.assertIn(step_id, self.app._progression_ids("correction"))
+
+    def test_existing_batch_artifacts_reveal_steps_even_in_live_mode(self):
+        for filename in ("batch_requests_flash.jsonl", "batch_requests_pro.jsonl",
+                         "batch_jobs.json", "batched_correction_result.jsonl"):
+            with self.subTest(filename=filename):
+                path = self.evaluation / filename
+                path.write_text("")
+                self.app._populate_tree()
+                self.assertFalse(self.app.tree.exists("batch_status"))
+                path.write_text('{"jobs":{"flash":{"name":"batches/1"}}}\n')
+                self.app._populate_tree()
+                self.assertTrue(self.app.tree.exists("batch_status"))
+                self.app._handle_first_visit(self.app.step_by_id["batch_status"])
+                self.app.update()
+                self.assertNotEqual(self.app.state_store.step("batch_status").get("status"), "skipped")
+                path.unlink()
+
+    def test_started_batches_remain_visible_after_mode_change_and_reload(self):
+        self.app.state_store.update_step("correction", batch_started=True, last_run_variant="live")
+        self.app.state_store.load(self.evaluation)
+        self.app._populate_tree()
+        self.assertTrue(self.app.tree.exists("batch_status"))
+        fresh = self.repository / "Fresh"
+        fresh.mkdir()
+        self.app.state_store.load(fresh)
+        self.app._populate_tree()
+        self.assertFalse(self.app.tree.exists("batch_status"))
+
+    def test_navigation_skips_hidden_batch_steps(self):
+        with patch.object(self.app.tree, "selection_set") as select:
+            self.app._move_selection_from("correction", 1)
+        select.assert_called_once_with("manual_resolution")
+
+    def test_long_running_steps_notify_once_on_success(self):
+        expected_text = {
+            "labels": "labels",
+            "correction": "correction",
+            "annotation": "copies annotées",
+            "read_annotations": "annotations manuscrites",
+        }
+        for step_id, completion_text in expected_text.items():
+            with self.subTest(step=step_id), patch("copienator_gui.app.notify_desktop") as notify:
+                self.app.active_step_id = step_id
+                self.app._finish_process(0, False)
+                self.app._finish_process(0, False)
+                notify.assert_called_once()
+                self.assertIn(self.evaluation.name, notify.call_args.args[1])
+                self.assertIn(completion_text, notify.call_args.args[1])
+
+    def test_incomplete_failed_and_interrupted_steps_notify_the_outcome_once(self):
+        for step_id in ("labels", "correction", "annotation", "read_annotations"):
+            for code, interrupted, outcome in (
+                (4, False, "partielle"), (1, False, "échouée"),
+                (0, True, "interrompue"), (130, False, "interrompue"),
+            ):
+                with self.subTest(step=step_id, code=code, interrupted=interrupted), \
+                        patch("copienator_gui.app.notify_desktop") as notify:
+                    self.app.active_step_id = step_id
+                    self.app._finish_process(code, interrupted)
+                    self.app._finish_process(code, interrupted)
+                    notify.assert_called_once()
+                    title, message = notify.call_args.args
+                    self.assertIn(outcome, title)
+                    self.assertIn(outcome, message)
+                    self.assertIn(self.evaluation.name, message)
+                    self.assertIn(f"code {code}", message)
+                    self.assertNotIn("réussie", message)
+
+    def test_notification_failure_keeps_success_and_uses_existing_fallback(self):
+        self.app.active_step_id = "labels"
+        with patch("copienator_gui.app.notify_desktop", side_effect=RuntimeError("unavailable")), \
+                patch.object(self.app, "bell") as bell, \
+                patch.object(self.app, "_append_console") as console:
+            self.app._finish_process(0, False)
+        bell.assert_called_once()
+        self.assertTrue(any("Notification de bureau indisponible" in call.args[0]
+                            for call in console.call_args_list))
+        self.assertEqual(self.app.state_store.step("labels")["status"], "success")
+
     def test_optional_blank_crop_can_target_a_copy_or_be_skipped(self):
         copies = self.evaluation/"Copies"
         copies.mkdir()
@@ -190,7 +333,9 @@ 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 +352,11 @@ 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_gui_core.py

@@ -19,6 +19,7 @@ from PIL import Image
 from pypdf import PdfReader, PdfWriter
 
 from copienator import (
+    CliError,
     EvaluationWorkspace,
     WorkspaceNotFoundError,
     WorkspaceValidationError,
@@ -2058,7 +2059,23 @@ class StandardCliTests(unittest.TestCase):
             self.assertEqual(groups, [["Ex 1"], ["Ex 2"]])
             editor.assert_not_called()
 
-    def test_grouped_batching_does_not_split_one_student(self) -> None:
+    def test_grouped_batching_looks_ahead_to_keep_below_the_hard_limit(self) -> None:
+        module = self.modules["annotating_by_label"]
+        large = Image.new("RGB", (10, 80), "white")
+        small = Image.new("RGB", (10, 15), "white")
+        rendered = [
+            ("01", "Ex 1", large, 0, []),
+            ("02", "Ex 1", small, 0, []),
+            ("02", "Ex 2", small, 0, []),
+        ]
+        with patch.object(module, "MAX_HEIGHT_PX", 100):
+            batches = module.split_batches(rendered)
+        self.assertEqual([len(batch) for batch in batches], [1, 2])
+        self.assertTrue(
+            all(sum(item[2].height for item in batch) <= 100 for batch in batches)
+        )
+
+    def test_grouped_batching_never_splits_one_student(self) -> None:
         module = self.modules["annotating_by_label"]
         image = Image.new("RGB", (10, 60), "white")
         rendered = [
@@ -2066,9 +2083,16 @@ class StandardCliTests(unittest.TestCase):
             ("01", "Ex 2", image, 0, []),
             ("02", "Ex 1", image, 0, []),
         ]
-        with patch.object(module, "MAX_HEIGHT_PX", 100):
-            batches = module.split_batches(rendered)
-        self.assertEqual([len(batch) for batch in batches], [2, 1])
+        with patch.object(module, "MAX_HEIGHT_PX", 100), self.assertRaises(
+            CliError
+        ) as raised:
+            module.split_batches(rendered)
+        self.assertIn("Copie01", str(raised.exception))
+        self.assertIn("cannot be split", str(raised.exception))
+
+    def test_grouped_annotation_height_limit_is_32000_pixels(self) -> None:
+        module = self.modules["annotating_by_label"]
+        self.assertEqual(module.MAX_HEIGHT_PX, 32000)
 
 
 class WorkflowTests(unittest.TestCase):

tests/test_gui_manual_resolution.py

@@ -38,13 +38,31 @@ class ManualResolutionPanelTests(unittest.TestCase):
         self.assertEqual(panel.text.get("1.0", "end-1c"), content)
         self.assertEqual(len(panel.pdf_buttons), 4)
         self.assertEqual([button.cget("text") for button in panel.pdf_buttons],
-                         ["PDF source", "PDF cible"] * 2)
+                         ["Source : Ex 1", "Cible : Ex 2",
+                          "Source : Ex 3", "Cible : Ex 4"])
         with patch("copienator_gui.manual_resolution.open_path") as opened:
             panel.editor_button.invoke()
             for button in panel.pdf_buttons:
                 button.invoke()
         self.assertEqual(opened.call_args_list, [call(self.path), *map(call, paths)])
 
+    def test_reload_updates_the_displayed_target_and_its_pdf_action(self):
+        self.path.write_text("Copie01 A -> B|\n", encoding="utf-8")
+        copies = self.evaluation / "Copies" / "Copie01"
+        copies.mkdir(parents=True)
+        (copies / "A.pdf").touch()
+        (copies / "C.pdf").touch()
+        panel = self.panel()
+
+        self.path.write_text("Copie01 A -> C|\n", encoding="utf-8")
+        panel.reload_button.invoke()
+
+        self.assertEqual(panel.text.get("1.0", "end-1c"), "Copie01 A -> C|\n")
+        self.assertEqual(panel.pdf_buttons[1].cget("text"), "Cible : C")
+        with patch("copienator_gui.manual_resolution.open_path") as opened:
+            panel.pdf_buttons[1].invoke()
+        opened.assert_called_once_with(copies / "C.pdf")
+
     def test_reload_keeps_invalid_lines_visible_and_removes_stale_actions(self):
         self.path.write_text("Copie01 A -> B|\n", encoding="utf-8")
         panel = self.panel()

tests/test_manual_resolution_state.py

@@ -6,11 +6,40 @@ from unittest.mock import Mock, patch
 
 import pymupdf
 
-from copienator import EvaluationWorkspace, atomic_write_json, read_json
+from copienator import CliError, EvaluationWorkspace, atomic_write_json, read_json
 from copienator.commands import correction, resolve_manual
+from copienator_gui.app import has_manual_conflicts
 
 
 class ManualResolutionStateTests(unittest.TestCase):
+    def test_merge_into_missing_target_uses_source_and_schedules_redo(self):
+        for operator in ("->", "x>"):
+            for target in ("B|", "|B"):
+                with self.subTest(operator=operator, target=target), tempfile.TemporaryDirectory() as temporary:
+                    workspace = self.fixture(Path(temporary), operator)
+                    copies = workspace.copies_dir / "Copie01"
+                    (copies / "B.pdf").unlink()
+                    workspace.manual_resolutions_file.write_text(f"Copie01 A {operator} {target}\n")
+                    resolve_manual.resolve_manual(workspace)
+                    with pymupdf.open(copies / "B_new.pdf") as document:
+                        self.assertEqual(len(document), 1)
+                        self.assertIn("A", document[0].get_text())
+                    self.assertEqual((copies / "A_old.pdf").exists(), operator == "x>")
+                    self.assertEqual((copies / "A.pdf").exists(), operator == "->")
+                    self.assertFalse((copies / "B_old.pdf").exists())
+                    self.assertEqual(read_json(workspace.refaire_file), [["Copie01", ["B"]]])
+                    self.assertNotIn("delayed", read_json(workspace.correction_file)["A"][0][0]["result"])
+
+    def test_merge_still_requires_source_without_changing_state(self):
+        with tempfile.TemporaryDirectory() as temporary:
+            workspace = self.fixture(Path(temporary), "x>")
+            (workspace.copies_dir / "Copie01" / "A.pdf").unlink()
+            before = workspace.correction_file.read_bytes()
+            with self.assertRaises(CliError):
+                resolve_manual.resolve_manual(workspace)
+            self.assertEqual(workspace.correction_file.read_bytes(), before)
+            self.assertTrue(workspace.manual_resolutions_file.exists())
+
     def test_acknowledgement_clears_only_this_target_and_copy(self):
         for error in ("wrg-lbl:B?", "wrg-lbl:B?delayed", "wrg-lbl:B?exists", "al:(->)B?(->)C?", "al:(delayed)B"):
             with self.subTest(error=error):
@@ -63,3 +92,75 @@ class ManualResolutionStateTests(unittest.TestCase):
                 self.assertEqual(correction.run_configured(args), 0)
             self.assertFalse(workspace.manual_resolutions_file.exists())
             self.assertNotIn("delayed", correction.results["A"][0][0]["result"])
+
+    def test_deleted_conflict_stays_resolved_when_other_resolution_generates_refaire(self):
+        for kind in ("wrong-label", "add-label"):
+            with self.subTest(kind=kind), tempfile.TemporaryDirectory() as temporary:
+                workspace = self.fixture(Path(temporary), "x>")
+                other = workspace.copies_dir / "Copie02"
+                other.mkdir()
+                for label in ("A", "B"):
+                    (other / f"{label}.pdf").write_bytes(
+                        (workspace.copies_dir / "Copie01" / f"{label}.pdf").read_bytes())
+                before = {path.name: path.read_bytes() for path in other.iterdir()}
+                results = read_json(workspace.correction_file)
+                results["A"][0].append({"id": "02", "result": {
+                    "error": "wrg-lbl:B?" if kind == "wrong-label" else "al:(->)B?",
+                    "delayed": [[kind, "B"]], "score": 3,
+                }})
+                atomic_write_json(workspace.correction_file, results)
+                # Only Copie01 remains in the edited file; Copie02 was deleted.
+                resolve_manual.resolve_manual(workspace)
+                self.assertEqual(read_json(workspace.refaire_file), [["Copie01", ["B"]]])
+                self.assertEqual(before, {path.name: path.read_bytes() for path in other.iterdir()})
+                saved = read_json(workspace.correction_file)["A"][0][1]["result"]
+                self.assertNotIn("delayed", saved)
+                self.assertEqual(saved["score"], 3)
+                self.assertNotIn("B?", saved["error"])
+                (workspace.groups_dir / "B").mkdir(parents=True)
+                args = correction.build_parser().parse_args([str(workspace.root), "--refaire"])
+                correction.configure_runtime(workspace, [], args, api_client=Mock())
+                with patch.object(correction.grouping, "get_pdf_height", return_value=400), \
+                        patch.object(correction.grouping, "create_jpg"), \
+                        patch.object(correction, "process_single_task", return_value=[]):
+                    self.assertEqual(correction.run_configured(args), 0)
+                self.assertFalse(workspace.manual_resolutions_file.exists())
+                self.assertNotIn("delayed", correction.results["A"][0][1]["result"])
+
+    def test_deleting_all_lines_can_be_applied_and_survives_resume(self):
+        for text in ("", "### Instructions\n", "### Copie01 A x> B|\n"):
+            with self.subTest(text=text), tempfile.TemporaryDirectory() as temporary:
+                workspace = self.fixture(Path(temporary), "x>")
+                source = workspace.copies_dir / "Copie01"
+                before = {path.name: path.read_bytes() for path in source.iterdir()}
+                workspace.manual_resolutions_file.write_text(text)
+                self.assertTrue(has_manual_conflicts(workspace.manual_resolutions_file))
+                resolve_manual.resolve_manual(workspace)
+                self.assertEqual(before, {path.name: path.read_bytes() for path in source.iterdir()})
+                self.assertFalse(has_manual_conflicts(workspace.manual_resolutions_file))
+                atomic_write_json(workspace.refaire_file, [])
+                args = correction.build_parser().parse_args([str(workspace.root), "--refaire"])
+                correction.configure_runtime(workspace, [], args, api_client=Mock())
+                self.assertEqual(correction.run_configured(args), 0)
+                self.assertFalse(workspace.manual_resolutions_file.exists())
+
+    def test_invalid_remaining_instruction_does_not_commit_deletions(self):
+        with tempfile.TemporaryDirectory() as temporary:
+            workspace = self.fixture(Path(temporary), "x>")
+            before = workspace.correction_file.read_bytes()
+            workspace.manual_resolutions_file.write_text("invalid instruction\n")
+            with self.assertRaises(CliError):
+                resolve_manual.resolve_manual(workspace)
+            self.assertEqual(workspace.correction_file.read_bytes(), before)
+
+    def test_generated_first_comment_documents_deleting_lines(self):
+        with tempfile.TemporaryDirectory() as temporary:
+            workspace = self.fixture(Path(temporary), "x>")
+            atomic_write_json(workspace.refaire_file, [])
+            args = correction.build_parser().parse_args([str(workspace.root), "--refaire"])
+            correction.configure_runtime(workspace, [], args, api_client=Mock())
+            self.assertEqual(correction.run_configured(args), 0)
+            first_line = workspace.manual_resolutions_file.read_text().splitlines()[0]
+            self.assertTrue(first_line.startswith("###"))
+            self.assertIn("delete a line", first_line)
+            self.assertIn("resolve-manual", first_line)

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"

tests/test_utils_compile.py

@@ -27,9 +27,26 @@ class CompileToPdfTests(unittest.TestCase):
                 utils.compile_to_pdf("broken", output)
 
             self.assertEqual(output.read_bytes(), b"partial")
-            self.assertIn("Warning: LaTeX compilation failed", console.getvalue())
+            self.assertIn("Warning: LaTeX compilation reported errors", console.getvalue())
+            self.assertIn("a PDF was generated", console.getvalue())
             self.assertIn("missing \\item", console.getvalue())
 
+    def test_says_no_pdf_was_generated_when_latex_fails(self):
+        with tempfile.TemporaryDirectory() as temporary:
+            output = Path(temporary) / "result.pdf"
+            result = subprocess.CompletedProcess(
+                args=[], returncode=1, stdout="! Undefined control sequence.\n"
+            )
+
+            console = io.StringIO()
+            with patch.object(utils.subprocess, "run", return_value=result), \
+                    contextlib.redirect_stdout(console):
+                utils.compile_to_pdf("broken", output)
+
+            self.assertFalse(output.exists())
+            self.assertIn("no PDF was generated", console.getvalue())
+            self.assertIn("Undefined control sequence", console.getvalue())
+
     def test_warns_when_no_pdf_is_produced(self):
         with tempfile.TemporaryDirectory() as temporary:
             output = Path(temporary) / "result.pdf"