Provide Gemini with the list of IDs in correction request

copienator/commands/correction.py

@@ -532,6 +532,19 @@ def process_single_task(task_tuple, precomputed_response=None):
 
             json_data = parse_correction_response(full_response_text)
 
+            # Validate the whole group before moving PDFs or applying corrections.
+            returned_ids = [entry["id"] for entry in json_data]
+            unexpected = sorted(set(returned_ids) - set(d_data))
+            missing = sorted(set(d_data) - set(returned_ids))
+            duplicates = sorted({pid for pid in returned_ids if returned_ids.count(pid) > 1})
+            if unexpected or missing or duplicates:
+                raise ValueError(
+                    f"Correction ID mismatch: expected {list(d_data)}, "
+                    f"received {returned_ids}; unexpected IDs: {unexpected}; "
+                    f"missing IDs: {missing}; duplicate IDs: {duplicates}. "
+                    "No corrections from this response were applied."
+                )
+
             # Ensure consistency of answer placements
             for p in json_data:
                 pid = p["id"]
@@ -899,7 +912,10 @@ def run_configured(args: argparse.Namespace) -> ExitCode:
                                             "data": b64_img,
                                         }
                                     },
-                                    {"text": prompting.make_prompt(INPUT_DIR, label)},
+                                    {"text": prompting.make_prompt(
+                                        INPUT_DIR, label,
+                                        expected_ids=[row[0] for row in group_data],
+                                    )},
                                 ],
                             }
                         ],

copienator/commands/cutleft.py

@@ -10,6 +10,7 @@ from pathlib import Path
 from queue import Empty, Queue
 from threading import Thread
 
+import pymupdf
 from pdf2image import convert_from_path
 from PIL import Image, ImageTk
 
@@ -31,6 +32,8 @@ DELIMITER_COLOR = (0, 0, 0)
 OUTPUT_SIZE = (1800, 1000)
 CROP_SHIFT_STEP = 50
 CROP_WIDTH_STEP = 50
+RENDER_DPI = 200
+NARROW_PAGE_RATIO = 0.8
 pdf_cache_lock = threading.Lock()
 
 # The requested spelling is kept as the public switch for this alternate layout.
@@ -52,14 +55,17 @@ def distribute_pages(total_pages: int, max_per_file: int = 5) -> list[int]:
 def stitch_images(image_list: list[Image.Image]) -> Image.Image | None:
     if not image_list:
         return None
-    total_width = sum(image.width for image in image_list)
+    # Downstream label coordinates assume equally spaced page columns. Pad
+    # narrower strips instead of scaling their content or vertical coordinates.
+    column_width = max(image.width for image in image_list)
+    total_width = column_width * len(image_list)
     total_width += (len(image_list) - 1) * DELIMITER_WIDTH
     max_height = max(image.height for image in image_list)
     combined = Image.new("RGB", (total_width, max_height), color="white")
     x_offset = 0
     for index, image in enumerate(image_list):
         combined.paste(image, (x_offset, 0))
-        x_offset += image.width
+        x_offset += column_width
         if index < len(image_list) - 1:
             delimiter = Image.new(
                 "RGB", (DELIMITER_WIDTH, max_height), color=DELIMITER_COLOR
@@ -73,7 +79,7 @@ def stitch_images(image_list: list[Image.Image]) -> Image.Image | None:
 def _get_pdf_pages_cached(pdf_path: Path) -> list[Image.Image]:
     # Label coordinates use the full, displayed MediaBox, including on PDFs
     # previously cropped by crop-margins. Keep this in sync with split-answers.
-    return convert_from_path(pdf_path, use_cropbox=False)
+    return convert_from_path(pdf_path, use_cropbox=False, dpi=RENDER_DPI)
 
 
 def get_pdf_pages(pdf_path: Path) -> list[Image.Image]:
@@ -82,6 +88,29 @@ def get_pdf_pages(pdf_path: Path) -> list[Image.Image]:
         return _get_pdf_pages_cached(pdf_path)
 
 
+def normal_page_width(pdf_path: Path, images: list[Image.Image]) -> float:
+    """Infer the unsplit sheet width, including for legacy PDFs without metadata."""
+    reference = max(image.width for image in images)
+    original = pdf_path.parent.parent / "Copies Originales" / pdf_path.name
+    try:
+        if not original.is_file():
+            return reference
+        with pymupdf.open(original) as document:
+            # Landscape scans contain two normal booklet pages; portrait scans
+            # contain one. MediaBox dimensions match get_pdf_pages above.
+            widths = []
+            for page in document:
+                rect = page.mediabox * page.rotation_matrix
+                width = rect.width / 2 if rect.width > rect.height else rect.width
+                widths.append(width * RENDER_DPI / 72)
+            if widths:
+                reference = max(widths)
+    except (OSError, RuntimeError, ValueError):
+        # A missing/unreadable backup must not prevent reviewing a valid copy.
+        pass
+    return reference
+
+
 def process_single_pdf(
     pdf_path: Path,
     shift_offset: int = 0,
@@ -91,10 +120,20 @@ def process_single_pdf(
     """Convert one PDF into a preview, full-resolution splits and metadata."""
     try:
         cropped_images = []
-        for image in get_pdf_pages(pdf_path):
+        images = get_pdf_pages(pdf_path)
+        if not images:
+            return None
+        reference_width = normal_page_width(pdf_path, images)
+        full_width_pages = []
+        for page_index, image in enumerate(images):
             width, height = image.size
             if max_per_file == 1:
                 left, right = 0, width
+            elif width < NARROW_PAGE_RATIO * reference_width:
+                # Secondary subpages start at their own left edge. Keep the
+                # normal margin width, capped to the available page content.
+                left = min(width - 1, max(0, shift_offset))
+                right = min(width, left + int(reference_width // 3) + max(0, width_offset))
             else:
                 left = max(0, 100 + shift_offset)
                 right = min(
@@ -103,6 +142,8 @@ def process_single_pdf(
                 )
             if right > left:
                 cropped_images.append(image.crop((left, 0, right, height)))
+                if left == 0 and right == width:
+                    full_width_pages.append(page_index + 1)
         if not cropped_images:
             return None
 
@@ -123,6 +164,7 @@ def process_single_pdf(
             "total_pages": len(cropped_images),
             "number_of_files": len(split_images),
             "columns_per_file": distribution,
+            "full_width_pages": full_width_pages,
         }
         return preview, split_images, schema
     except Exception as exc:  # noqa: BLE001 - interactive item failure

copienator/prompting.py

@@ -1,6 +1,8 @@
 from pathlib import Path
 import io
+import json
 from . import utils
+from . import read_json
 from . import configuration as settings
 from .gemini_config import thinking_config
 
@@ -103,7 +105,7 @@ sur les autres questions."""
 
 from .utils import get_label_text_content, get_label_sol_content, get_label_persp_content
 
-def make_prompt(input_dir,full_label):
+def make_prompt(input_dir,full_label, *, expected_ids=None):
     text = get_label_text_content(input_dir, full_label) or ""
     corr = get_label_sol_content(input_dir, full_label) or ""
     persp = get_label_persp_content(input_dir, full_label) or ""
@@ -117,7 +119,18 @@ def make_prompt(input_dir,full_label):
             + persp
             + "\n```\n"
         )
-    return main_prompt.replace("<<text>>", text).replace("<<corr>>", corr).replace("<<persp>>", persp).replace("<<label>>", full_label)
+    prompt = main_prompt.replace("<<text>>", text).replace("<<corr>>", corr).replace("<<persp>>", persp).replace("<<label>>", full_label)
+    if expected_ids is not None:
+        prompt += (
+            "\n\nIdentifiants exacts présents dans cette image, de haut en bas : "
+            + json.dumps(list(expected_ids), ensure_ascii=False)
+            + ".\nRenvoie exactement un résultat par identifiant de cette liste, "
+            "sans omission, doublon ni identifiant supplémentaire. Conserve les zéros initiaux. "
+            "Utilise uniquement les en-têtes imprimés « ID: … » pour identifier les réponses ; "
+            "ignore les numéros manuscrits. Tout le contenu jusqu’au prochain en-tête "
+            "imprimé appartient au même identifiant, même s’il contient plusieurs feuilles."
+        )
+    return prompt
 
 
 from pydantic import BaseModel, Field, TypeAdapter
@@ -190,8 +203,9 @@ def batch_generation_config(thinking_level):
 # The root model for parsing is be: List[EvaluationEntry]
 def generate_request(input_dir, file, full_label, *, thinking_level=settings.THINKING_LEVEL_PRO):
     """Generates request for Gemini."""
-    prompt = make_prompt(input_dir, full_label)
     image_path = Path(file)
+    group_data = read_json(image_path.with_suffix(".json"))
+    prompt = make_prompt(input_dir, full_label, expected_ids=[row[0] for row in group_data])
 
     contents = [
         types.Content(