Fix context duplication in pdf output

copienator/commands/reading_annotations.py

@@ -9,20 +9,26 @@ import numpy as np
 from pdf2image import convert_from_path
 from PIL import Image, ImageChops, ImageDraw, ImageFilter
 
-from copienator.commands import annotating
-from copienator import configuration, utils
 from copienator import (
     EvaluationWorkspace,
     ExitCode,
     atomic_write_json,
+    configuration,
     evaluation_parser,
     execute,
     read_json,
+    utils,
     workspace_from_args,
 )
 from copienator.annotation_actions import apply_checkbox_actions, apply_score_overrides
-from copienator.answer_info import build_answer_info
 from copienator.annotation_data import AnnotationData, load_annotation_data
+from copienator.annotation_modifications import (
+    BASELINE,
+    output_removals,
+    stage_baseline,
+)
+from copienator.answer_info import build_answer_info
+from copienator.commands import annotating
 from copienator.filesystem import staged_files
 from copienator.return_answers import save_return_answer_options
 
@@ -168,6 +174,8 @@ def apply_actions_and_regenerate(
     scores = dict.fromkeys(all_labels, "")
     answer_labels: list[str] = []
     dirty_images: dict[str, Image.Image] = {}
+    header_heights: dict[str, int] = {}
+    filtered_labels: list[str] = []
     concatenated: list[Image.Image] = []
     filtered: list[Image.Image] = []
     incomplete = False
@@ -221,15 +229,18 @@ def apply_actions_and_regenerate(
                 final_image.paste(body, (0, new_header_height), mask=body)
 
         dirty_images[label] = final_image
+        header_heights[label] = new_header_height
         answer_labels.append(label)
         concatenated.append(final_image)
         if float(scores[label]) != 4.0 or result.get("feedback", []):
             filtered.append(final_image)
+            filtered_labels.append(label)
 
+    if incomplete:
+        return ExitCode.PARTIAL
     concat_image = concatenate(concatenated)
     filtered_image = concatenate(filtered)
-    with staged_files(output_dir, remove=("Concat.jpg", "Concat_F.jpg", "Concat_F.pdf",
-                                         "touched.json", "answer_labels.json")) as staging:
+    with staged_files(output_dir, remove=(*output_removals(output_dir), BASELINE)) as staging:
         for label, image in dirty_images.items():
             image.save(staging / f"{label}.jpg")
         if not preserve_score_file:
@@ -241,6 +252,9 @@ def apply_actions_and_regenerate(
             concat_image.save(staging / "Concat.jpg")
         if filtered_image is not None:
             filtered_image.save(staging / "Concat_F.jpg")
+        stage_baseline(workspace.root, "Bnot", output_dir, staging, dirty_images,
+                       header_heights, build_answer_info(scores, labels_data, answer_labels),
+                       filtered_labels, scores)
 
     if preserve_score_file:
         print(f"  Preserved existing score.json in {output_dir}")

copienator/commands/reading_grouped_annotations.py

@@ -2,18 +2,21 @@ from __future__ import annotations
 
 import argparse
 import concurrent.futures
+import unicodedata
+from hashlib import sha256
 from collections import defaultdict
 from collections.abc import Sequence
 from pathlib import Path
 from typing import Any
 
 from PIL import Image, ImageDraw
+import pymupdf
 
-from copienator import configuration
 from copienator import (
     EvaluationWorkspace,
     ExitCode,
     atomic_write_json,
+    configuration,
     evaluation_parser,
     execute,
     read_json,
@@ -21,8 +24,15 @@ from copienator import (
     workspace_from_args,
 )
 from copienator.annotation_actions import apply_checkbox_actions, apply_score_overrides
-from copienator.answer_info import build_answer_info
 from copienator.annotation_data import AnnotationData, RefaireList, load_annotation_data
+from copienator.annotation_modifications import (
+    BASELINE,
+    baseline,
+    load_image,
+    output_removals,
+    stage_baseline,
+)
+from copienator.answer_info import build_answer_info
 from copienator.commands import annotating
 from copienator.commands.reading_annotations import (
     concatenate,
@@ -37,24 +47,51 @@ ScanResult = tuple[dict[str, list[dict[str, Any]]], dict[str, LabelNotes]]
 SCAN_WORKERS = 2
 
 
+def context_content_key(path: str | Path) -> tuple[str, str]:
+    """Identify context by its text, independently of filename and PDF metadata."""
+    with pymupdf.open(path) as document:
+        text = "\n".join(page.get_text(sort=True) for page in document)
+    text = " ".join(unicodedata.normalize("NFC", text).split())
+    if text:
+        return "text", text
+    # Image-only contexts must not all collapse to the same empty-text key.
+    return "file", sha256(Path(path).read_bytes()).hexdigest()
+
+
 def get_extra_pdfs_as_images(
     root_dir: str | Path,
     label: str,
     annotating_module: Any,
     all_labels: list[str],
+    *,
+    seen_contexts: set[tuple[str, str]] | None = None,
 ) -> list[Image.Image]:
-    """Convert the context, question and solution PDFs associated with a label."""
+    """Render supplements with semantic borders and optional per-copy deduplication."""
     paths = [
-        *utils.pdf_images_of_contexts(root_dir, label, all_labels),
-        utils.pdf_image_of_enonce(root_dir, label),
-        utils.pdf_image_of_solution(root_dir, label),
+        *(("context", path) for path in utils.pdf_images_of_contexts(root_dir, label, all_labels)),
+        ("question", utils.pdf_image_of_enonce(root_dir, label)),
+        ("solution", utils.pdf_image_of_solution(root_dir, label)),
     ]
     images = []
-    for path in paths:
+    for role, path in paths:
         if path:
+            key = None
+            if role == "context" and seen_contexts is not None:
+                key = context_content_key(path)
+                if key in seen_contexts:
+                    continue
             image, _, _ = annotating_module.make_base_image(path)
             if image is not None:
+                if role in {"question", "solution"}:
+                    image = image.copy()
+                    ImageDraw.Draw(image).rectangle(
+                        [0, 0, image.width - 1, image.height - 1],
+                        outline="black" if role == "question" else "blue",
+                        width=int((0.2 / 2.54) * 100),
+                    )
                 images.append(image)
+                if key is not None:
+                    seen_contexts.add(key)
     return images
 
 
@@ -67,7 +104,6 @@ def save_paginated_pdf(
         return
     max_width = max(image.width for group in non_empty for image in group)
     max_page_height = int(max_width * 1.414 * 1.25)
-    border = int((0.2 / 2.54) * 100)
     left_margin = int((0.3 / 2.54) * 100)
     vertical_margin = int((0.2 / 2.54) * 100)
     max_content_height = max_page_height - 2 * vertical_margin
@@ -94,21 +130,10 @@ def save_paginated_pdf(
         page_height = 0
 
     for group in non_empty:
-        processed: list[Image.Image] = []
-        for index, image in enumerate(group):
-            if index in (0, 1):
-                image = image.copy()
-                color = "black" if index == 0 else "blue"
-                ImageDraw.Draw(image).rectangle(
-                    [0, 0, image.width - 1, image.height - 1],
-                    outline=color,
-                    width=border,
-                )
-            processed.append(image)
-        group_height = sum(image.height for image in processed)
+        group_height = sum(image.height for image in group)
         if page_images and page_height + group_height > max_content_height:
             finish_page()
-        page_images.extend(processed)
+        page_images.extend(group)
         page_height += group_height
     finish_page()
     pages[0].save(
@@ -247,8 +272,12 @@ def apply_actions_and_regenerate_grouped(
     touched = dict.fromkeys(all_labels, False)
     answer_labels: list[str] = []
     dirty_images: dict[str, Image.Image] = {}
+    header_heights: dict[str, int] = {}
+    filtered_labels: list[str] = []
+    previous_baseline = baseline(output_dir) if (output_dir / BASELINE).is_file() else None
     concat_images: list[Image.Image] = []
     filtered_groups: list[list[Image.Image]] = []
+    seen_contexts: set[tuple[str, str]] = set()
     incomplete = False
 
     for label, content in sorted(
@@ -266,7 +295,17 @@ def apply_actions_and_regenerate_grouped(
         if result.get("error") == "empty-answer":
             continue
         saved_image = output_dir / f"{label}.jpg"
-        if selected_labels and label not in selected_labels and saved_image.is_file():
+        baseline_entry = (previous_baseline or {}).get("images", {}).get(label)
+        if selected_labels and label not in selected_labels and (saved_image.is_file() or baseline_entry):
+            if annotation_dir in {"BGnot", "Bnot"}:
+                if previous_baseline is not None and label in previous_baseline["images"]:
+                    entry = previous_baseline["images"][label]
+                    final_image = load_image(output_dir, entry)
+                    header_heights[label] = entry["header_height"]
+                else:
+                    with Image.open(saved_image) as saved:
+                        final_image = saved.convert("RGB").copy()
+            else:
                 with Image.open(saved_image) as saved:
                     final_image = saved.convert("RGB").copy()
             if (
@@ -282,10 +321,11 @@ def apply_actions_and_regenerate_grouped(
             scores[label] = str(old_scores.get(label, scores[label]))
             concat_images.append(final_image)
             answer_labels.append(label)
+            filtered_labels.append(label)
             # Keep previously reviewed content, including handwriting.
             if annotation_dir == "BGnot":
                 extras = get_extra_pdfs_as_images(
-                    workspace.root, label, annotating, all_labels
+                    workspace.root, label, annotating, all_labels, seen_contexts=seen_contexts
                 )
                 filtered_groups.append([*extras, final_image])
                 touched[label] = True
@@ -327,6 +367,7 @@ def apply_actions_and_regenerate_grouped(
 
         # Persist every final block, including unchanged answers, for returns.
         dirty_images[label] = final_image
+        header_heights[label] = new_header_height
         answer_labels.append(label)
         concat_images.append(final_image)
 
@@ -335,8 +376,10 @@ def apply_actions_and_regenerate_grouped(
             feedback.get("to_delete", False) for feedback in feedbacks
         )
         if not perfect or has_notes:
+            filtered_labels.append(label)
             extras = (
-                get_extra_pdfs_as_images(workspace.root, label, annotating, all_labels)
+                get_extra_pdfs_as_images(workspace.root, label, annotating, all_labels,
+                                        seen_contexts=seen_contexts)
                 if annotation_dir == "BGnot"
                 else []
             )
@@ -346,8 +389,7 @@ def apply_actions_and_regenerate_grouped(
     concat_image = concatenate(concat_images)
     if incomplete:
         return ExitCode.PARTIAL, "\n".join(logs)
-    with staged_files(output_dir, remove=("Concat.jpg", "Concat_F.pdf", "Concat_F.jpg",
-                                         "touched.json", "answer_labels.json")) as staging:
+    with staged_files(output_dir, remove=(*output_removals(output_dir), BASELINE)) as staging:
         if simple_layout is not None:
             simple_layout["replaced"] = sorted(
                 set(simple_layout["replaced"]) | selected_labels
@@ -370,6 +412,9 @@ def apply_actions_and_regenerate_grouped(
                     [image for group in filtered_groups for image in group]
                 )
                 filtered_image.save(staging / "Concat_F.jpg")
+        stage_baseline(workspace.root, annotation_dir, output_dir, staging, dirty_images,
+                       header_heights, build_answer_info(scores, labels_data, answer_labels, touched),
+                       filtered_labels, scores)
     if preserve_score_file:
         logs.append(f"  Preserved existing score.json in {output_dir}")
     logs.append(f"  Saved regenerated files in {output_dir}")

copienator/dispatcher.py

@@ -40,6 +40,7 @@ COMMANDS: dict[str, Command] = {
     "import": Command("import_annotations", "Import handwritten annotations"),
     "read-annotations": Command("reading_annotations", "Read checkable annotations"),
     "read-grouped": Command("reading_grouped_annotations", "Read grouped annotations"),
+    "modify-annotations": Command("modify_annotations", "Modify reviewed answer bodies"),
     "giving-names": Command("giving_names", "Name copies and prepare A Rendre"),
     "update-ods": Command("update_ods", "Update the configured score spreadsheet"),
     "add-final-score": Command("add_final_score", "Stamp final scores on copies"),

copienator/filesystem.py

@@ -92,3 +92,45 @@ def staged_files(
                 if not destination_path.exists():
                     saved.replace(destination_path)
             _remove_path(backup)
+
+
+def publish_file_sets(file_sets: dict[Path, tuple[Path, Iterable[str]]]) -> None:
+    """Commit several staged file sets together, rolling every set back on error.
+
+    Staging and destination directories must be on the same filesystem. Unlike
+    nesting staged_files, failure in a later directory restores earlier ones too.
+    """
+    token = uuid.uuid4().hex
+    backups: list[tuple[Path, Path]] = []
+    committed: list[Path] = []
+    try:
+        for target, (staging, remove) in file_sets.items():
+            target.mkdir(parents=True, exist_ok=True)
+            backup = target / f".publish-{token}.backup"
+            backup.mkdir()
+            backups.append((target, backup))
+            staged = sorted(path for path in staging.iterdir() if path.is_file())
+            names = {path.name for path in staged} | set(remove)
+            for name in sorted(names):
+                if Path(name).name != name:
+                    raise ValueError(f"Invalid output filename: {name}")
+                destination = target / name
+                if destination.exists() or destination.is_symlink():
+                    destination.replace(backup / name)
+            for source in staged:
+                destination = target / source.name
+                source.replace(destination)
+                committed.append(destination)
+    except Exception:
+        for destination in reversed(committed):
+            _remove_path(destination)
+        for target, backup in reversed(backups):
+            for saved in backup.iterdir():
+                saved.replace(target / saved.name)
+        raise
+    finally:
+        for _, backup in backups:
+            if backup.exists() and not any(backup.iterdir()):
+                backup.rmdir()
+    for _, backup in backups:
+        _remove_path(backup)

copienator/utils.py

@@ -70,7 +70,7 @@ def pdf_images_of_contexts(root_dir, label, all_labels):
 
     pertinent_contexts = []
 
-    for filename in os.listdir(text2_dir):
+    for filename in sorted(os.listdir(text2_dir)):
         if filename.startswith("CTXT ") and filename.endswith(".pdf"):
             # Extract "first_label -> last_label" from "CTXT first_label -> last_label.pdf"
             core = filename[5:-4]

copienator_gui/app.py

@@ -30,6 +30,7 @@ from .scores import ScoreEditorPanel
 from .notifications import notify_desktop
 from .syncthing import SyncthingPanel
 from .manual_resolution import ManualResolutionPanel
+from .annotation_modifications import AnnotationModificationsPanel, annotation_mode_for_reader
 from .recorrection import prepare_selection_file
 from .refaire import SECTION as REFAIRE_SECTION
 from .refaire import (
@@ -824,6 +825,7 @@ class CopienatorApp(tk.Tk):
             "labels",
             "plotting",
             "manual_resolution",
+            "modify_annotations",
             "giving_names",
             "edit_scores",
             "statement_persp",
@@ -888,6 +890,9 @@ class CopienatorApp(tk.Tk):
                 value = False  # Deletion is a one-run choice, never a saved default.
             if step.id == 'statement' and spec.name == 'restart':
                 value = False
+            if step.id == "modify_annotations" and spec.name == "annotation_dir" and spec.name not in values:
+                reader = self.state_store.step("read_annotations")
+                value = annotation_mode_for_reader(reader)
             choices = spec.choices
             if spec.name == "annotation_dir" and step.id in {"export", "import"}:
                 choices, value = self._annotation_directory_choices(step.id)
@@ -939,12 +944,19 @@ class CopienatorApp(tk.Tk):
             self.edit_recorrection_button.grid(row=row, column=0, columnspan=2, sticky="w", pady=8)
             row += 1
 
-        if step.id == "manual_resolution":
+        if step.id in {"manual_resolution", "modify_annotations"}:
             def manual_evaluation():
                 target = self.arg_vars.get("target")
                 value = target.get().strip() if target else ""
                 return (self.repository / value).resolve() if value else self.evaluation
 
+            if step.id == "modify_annotations":
+                self.manual_panel = AnnotationModificationsPanel(
+                    self.form, manual_evaluation,
+                    lambda: self.arg_vars["annotation_dir"].get(),
+                )
+                self.arg_vars["annotation_dir"].trace_add("write", lambda *_: self.manual_panel.reload())
+            else:
                 self.manual_panel = ManualResolutionPanel(self.form, manual_evaluation)
             self.manual_panel.grid(row=row, column=0, columnspan=2, sticky="nsew", pady=8)
             if "target" in self.arg_vars:
@@ -966,7 +978,7 @@ class CopienatorApp(tk.Tk):
             attach_tooltip(extra_entry, step.extra_arguments_help)
             row += 1
 
-        self.run_button.configure(text="Enregistrer la sélection" if step.id == "refaire_selection" else ("Marquer terminée" if step.is_manual else "Exécuter"))
+        self.run_button.configure(text="Make modifications" if step.id == "modify_annotations" else "Enregistrer la sélection" if step.id == "refaire_selection" else ("Marquer terminée" if step.is_manual else "Exécuter"))
         if step.id == "personal_create":
             self.run_button.configure(text="Marquer Réussie")
         if step.id == "personal_read":
@@ -1953,7 +1965,7 @@ class CopienatorApp(tk.Tk):
                 self._append_console(f"Impossible de lancer la copie suivante : {exc}\n")
                 status, return_code = "failed", 1
         self.pending_refaire_commands = []
-        if step_id == "refaire_merge" and status == "success":
+        if step_id in {"refaire_merge", "modify_annotations"} and status == "success":
             for downstream in ("giving_names", "update_ods", "final_score", "personal_deploy", "personal_sent"):
                 entry = self.state_store.step(downstream)
                 if entry.get("status") in {"success", "detected", "skipped"}:
@@ -1995,6 +2007,12 @@ class CopienatorApp(tk.Tk):
                 "manual": False, "reason": reason,
             })
             self._append_console(f"Remplacer les barèmes par Gemini : ignorée — {reason}\n")
+        if step_id == "read_annotations" and status == "success":
+            reader = self.state_store.step(step_id)
+            mode = annotation_mode_for_reader(reader)
+            values = dict(self.state_store.step("modify_annotations").get("values", {}))
+            values["annotation_dir"] = mode
+            self.state_store.update_step("modify_annotations", values=values)
         if step_id == "annotation" and status == "success":
             generation_variant = self.state_store.step("annotation").get(
                 "last_run_variant"

copienator_gui/cut_helper.py

@@ -45,21 +45,29 @@ def cut_operator(percent: float, keep: int, mode: str) -> str:
 
 
 class CutHelper(tk.Toplevel):
-    def __init__(self, parent, path: Path, on_accept, initial=None):
-        # Load the source before creating a window so invalid PDFs leave no dialog.
+    def __init__(self, parent, path: Path | Image.Image, on_accept, initial=None):
+        # Body edits pass a raster without headers or supplementary pages.
+        width = min(850, parent.winfo_screenwidth() - 100)
+        if isinstance(path, Image.Image):
+            self.heights = [path.height]
+            self.scale = min(1.5, width / path.width)
+            rendered = [path.resize((round(path.width * self.scale),
+                                     round(path.height * self.scale)), Image.Resampling.LANCZOS)]
+            title = "Corps de la réponse annotée"
+        else:
             with pymupdf.open(path) as document:
                 if not len(document):
                     raise ValueError("Le PDF est vide.")
                 self.heights = [page.rect.height for page in document]
-            width = min(850, parent.winfo_screenwidth() - 100)
                 self.scale = min(1.5, width / max(page.rect.width for page in document))
                 rendered = []
                 for page in document:
                     pix = page.get_pixmap(matrix=pymupdf.Matrix(self.scale, self.scale), alpha=False,
                                           colorspace=pymupdf.csRGB)
                     rendered.append(Image.frombytes("RGB", (pix.width, pix.height), pix.samples))
+            title = path.name
         super().__init__(parent)
-        self.title(f"Cut — {path.name}")
+        self.title(f"Cut — {title}")
         self.geometry(f"{width + 45}x{min(850, parent.winfo_screenheight() - 100)}")
         self.transient(parent.winfo_toplevel())
         self.on_accept = on_accept
@@ -125,7 +133,7 @@ class CutHelper(tk.Toplevel):
         operator = cut_operator(self.percent, self.keep.get(), self.mode.get())
         rounded = float(operator.split("{")[1].split("}")[0])
         if not 0 < rounded < 100:
-            messagebox.showerror("Coupe invalide", "Chaque partie doit contenir une portion du PDF.", parent=self)
+            messagebox.showerror("Coupe invalide", "Chaque partie doit contenir une portion de la réponse.", parent=self)
             return
         self.destroy()
         self.on_accept(operator)

copienator_gui/manual_resolution.py

@@ -37,9 +37,10 @@ Enregistrez dans l’éditeur, puis cliquez sur Recharger et enfin sur Exécuter
 
 
 class ManualResolutionPanel(ttk.Frame):
-    def __init__(self, parent, get_evaluation):
+    def __init__(self, parent, get_evaluation, *, help_text=HELP, show_provenance=True):
         super().__init__(parent)
         self.get_evaluation = get_evaluation
+        self.show_provenance = show_provenance
         self.pdf_buttons = []
         self.cut_buttons = []
         self.copy_buttons = []
@@ -67,7 +68,7 @@ class ManualResolutionPanel(ttk.Frame):
         self.text.configure(yscrollcommand=scroll.set)
         scroll.pack(side="right", fill="y")
         self.text.pack(fill="both", expand=True)
-        ttk.Label(self, text=HELP, wraplength=650, justify="left").pack(fill="x", pady=8)
+        ttk.Label(self, text=help_text, wraplength=650, justify="left").pack(fill="x", pady=8)
         self.reload()
 
     def manual_path(self) -> Path | None:
@@ -123,7 +124,7 @@ class ManualResolutionPanel(ttk.Frame):
                 raise FileNotFoundError("Chargez une évaluation.")
             content = path.read_text(encoding="utf-8")
         except (OSError, UnicodeError) as exc:
-            self.status.configure(text=f"manual_resolutions.txt indisponible : {exc}")
+            self.status.configure(text=f"{path.name if path else 'Fichier'} indisponible : {exc}")
         else:
             self.editor_button.configure(state="normal")
             malformed = []
@@ -142,7 +143,7 @@ class ManualResolutionPanel(ttk.Frame):
                         key = (instruction.copy_id, instruction.old_label)
                         if key not in provenance:
                             try:
-                                provenance[key] = populated_labels(path.parent, *key)
+                                provenance[key] = populated_labels(path.parent, *key) if self.show_provenance else []
                             except (OSError, ValueError) as exc:
                                 provenance[key] = []
                                 provenance_error = f" — Provenance indisponible : {exc}"

copienator_gui/workflow.py

@@ -668,6 +668,23 @@ def build_workflow(show_personal_steps: bool) -> list[StepDefinition]:
                 ),
             ) + return_answer_arguments,
         ),
+        StepDefinition(
+            "modify_annotations",
+            "Finalisation",
+            "Modifier les réponses annotées",
+            "Coupez ou transférez les corps de réponses après lecture des annotations. "
+            "Les en-têtes, les scores et les documents complémentaires restent associés à chaque label.",
+            (python("default", "Modifications", "modify-annotations"),),
+            arguments=(
+                arg_target("le dossier de l’évaluation"),
+                ArgumentSpec("annotation_dir", "Dossier d’annotations", "choice", "--annotation-dir",
+                             default="BGnot", choices=("BGnot", "Bnot"),
+                             help="Choisissez le mode utilisé pour lire les annotations manuscrites ; les coupes portent uniquement sur les corps des réponses."),
+            ),
+            optional=True,
+            requires=("labels",),
+            artifacts=("annotation_modifications.txt",),
+        ),
         StepDefinition(
             "giving_names",
             "Finalisation",
@@ -778,15 +795,6 @@ def build_workflow(show_personal_steps: bool) -> list[StepDefinition]:
             (manual("deploy"),),
             personal=True,
         ),
-        StepDefinition(
-            "personal_print",
-            "Étapes personnelles",
-            "Imprimer une copie",
-            "Étape manuelle via le lecteur PDF.",
-            (manual("print"),),
-            personal=True,
-            optional=True,
-        ),
         StepDefinition(
             "clean",
             "Archivage",

docs/final_output.md

@@ -177,17 +177,22 @@ For each retained answer, the PDF stacks the following available material:
 3. The model solution from `Sol2/<label>.pdf`.
 4. The same compiled answer block used for the JPEG.
 
-Missing supplementary PDFs are skipped. Contexts can repeat for successive
-questions. Each question's complete group stays together on one page; groups
+Missing supplementary PDFs are skipped. Each distinct context appears only once
+per student PDF, before its first retained question. Duplicate detection uses
+extracted PDF text with Unicode and whitespace normalization, including contexts
+stored under different filenames or validity ranges. Numbers, mathematical
+symbols and case remain significant. Contexts without extractable text use a
+PDF-content hash instead, so unrelated image-only contexts are not collapsed.
+The same rule applies to selective redo and manual annotation modifications.
+Each question's complete group stays together on one page; groups
 are packed until the next would exceed the target height. An oversized group
 gets its own taller page, rather than being split. Pages are raster images saved
 as PDF at 100 dpi, with variable heights, not fixed A4 sheets or searchable text.
 
 The target height is `int(max_image_width * 1.414 * 1.25)` pixels. Small white
-margins are added on the left and above/below the page. The first image of each
-group receives a black border and the second a blue border. These borders are
-assigned by position, so they do not consistently identify question and solution
-when contexts are present or supplementary files are missing.
+margins are added on the left and above/below the page. The question receives
+a black border and the model solution a blue border. Context and student answer
+blocks receive neither border, including when supplementary files are missing.
 
 If nothing survives filtering, `read-grouped` removes old `Concat_F` outputs and
 does not create a new PDF. During a selective `--refaire` merge, saved answer
@@ -256,3 +261,33 @@ ones. Weighting and the final overall mark belong to the separate grading flow.
 - [ODS export](../copienator/commands/update_ods.py)
 - [Final-mark stamping](../copienator/commands/add_final_score.py)
 - [Cleanup retention](../copienator/commands/clean.py)
+
+
+## Manual modifications after annotation reading
+
+`modify-annotations EVALUATION --annotation-dir BGnot` (or `Bnot`) applies
+`annotation_modifications.txt`. The GUI exposes it between annotation reading
+and return preparation. Its `### Source: BGnot`/`Bnot` line must match the mode.
+Instruction operators follow manual resolution syntax but act on reviewed raster
+**bodies**, excluding headers and all supplementary pages. Headers and scores stay
+with their labels. Missing target bodies are allowed; their new header identifies
+the target and its existing score, without inventing feedback or a score.
+
+Both readers record `annotation_baseline.json` and lossless hidden
+`.annotation-base-*.png` images in each copy's annotation directory. The JSON
+stores the exact header height, baseline answer metadata, and filtered membership.
+Older outputs require one full annotation read before body editing. Legacy
+selective redo still works, but cannot infer missing header boundaries.
+
+Each application rebuilds from that baseline and publishes all affected copies
+and the mode's `annotation_modifications_applied.txt` in one rollback-capable
+transaction. Reapplying is idempotent; removing an instruction restores its prior
+content and membership. Annotation rereads refresh the baseline and replay the
+applied snapshot, not the draft. Selective redo preserves unaffected baseline
+layers so previous transfers are never included twice.
+
+The operation rebuilds per-label JPEGs, `info.json`, `Concat.jpg`, and the filtered
+BGnot PDF or Bnot JPEG. Every changed non-empty body is included in the filtered
+output, with its destination label's supplementary material. `score.json` is not
+modified. The GUI marks downstream return preparation stale. Rerun `giving-names`
+without `--update` to refresh both full outputs and individual answer exports.

docs/guide_utilisateur_gui.md

@@ -368,6 +368,56 @@ score.json** sert à mettre à jour les scores sur les images finales à
 partir d'une modification manuelle des fichiers `score.json`.
 
 
+### Modifier les réponses annotées
+
+Cette étape facultative se trouve après la lecture des annotations, pour **Bnot**
+et **BGnot**. Pour une ancienne évaluation, relancez d’abord **Lire les annotations
+manuscrites** : cette lecture enregistre maintenant la limite exacte de chaque
+en-tête et une référence sans perte des images annotées.
+
+Saisissez la **Copie**, le **Label source** et le **Label cible**, puis **Ajouter**.
+Les deux labels doivent exister ; la source doit avoir une réponse annotée. Une
+cible sans réponse est acceptée. Vous pouvez ajouter plusieurs instructions.
+
+Les boutons **Source**, **Cible** et **Cut** montrent uniquement le **corps de la
+réponse** : copie originale, annotations Gemini conservées et annotations
+manuscrites situées sur ce corps. Ils excluent le score, les commentaires globaux
+de l’en-tête, le contexte, l’énoncé et le corrigé. Les annotations manuscrites de
+l’en-tête restent dans leur en-tête. Le pourcentage de coupe porte sur la hauteur
+du corps uniquement. Aucun fragment n’est redimensionné lors des transferts.
+
+Les instructions sont enregistrées dans `annotation_modifications.txt` :
+
+```text
+### Source: BGnot
+Copie43 Ex 1 c{43}1> Ex 2|
+```
+
+Cet exemple conserve le début de la source (43 %) et ajoute la fin après le corps
+de la cible. `2` conserve la fin, `x` remplace le corps cible au lieu d’ajouter, et
+`|Ex 2` place le fragment avant la cible. Les opérateurs sans coupe de la résolution
+manuelle sont aussi disponibles et décrits dans le panneau. Une instruction par
+source ; une cible ne peut pas être source dans le même fichier. Plusieurs sources
+peuvent alimenter une cible, dans l’ordre du fichier.
+
+**Cut** prépare une commande à recopier. Utilisez **Ouvrir dans un éditeur de
+texte**, enregistrez puis **Recharger**. **Aperçu du résultat** montre les corps
+résultants pour la copie saisie. Cliquez sur **Make modifications** pour appliquer
+le fichier enregistré. Chaque label conserve son en-tête ; les scores restent
+inchangés. Les images individuelles, les images combinées, le PDF filtré BGnot
+(ou l’image filtrée Bnot) et les informations de présence sont reconstruits.
+Les réponses modifiées non vides sont incluses dans le résultat filtré, avec les
+compléments propres à leur label.
+
+Le fichier reste éditable après application. Réappliquer ne cumule pas les
+transferts : le traitement repart de la référence issue de la lecture des
+annotations. Retirer une ligne puis appliquer annule son effet. Une nouvelle
+lecture des annotations renouvelle la référence et réapplique les dernières
+instructions **appliquées**, sans prendre en compte un brouillon non appliqué.
+
+Relancez ensuite **Attribuer les noms et préparer A Rendre**, sans cocher l’option
+limitée aux images `answers`, afin d’actualiser tous les livrables.
+
 ### Attribuer les noms et préparer A Rendre
 
 **Effet :** rassemble les résultats de chaque élève dans le dossier `A Rendre`.