Better (?) default left margin

copienator/commands/cutleft.py

@@ -26,6 +26,7 @@ from copienator import (
 from copienator.filesystem import staged_files
 from copienator.copy_errors import clear_copy_error, mark_copy_error, marked_copy_paths
 from copienator.configuration import SHOW_PERSONAL_STEPS
+from copienator.left_margin import analyse_margins
 
 DELIMITER_WIDTH = 5
 DELIMITER_COLOR = (0, 0, 0)
@@ -34,6 +35,7 @@ CROP_SHIFT_STEP = 50
 CROP_WIDTH_STEP = 50
 RENDER_DPI = 200
 NARROW_PAGE_RATIO = 0.8
+AUTO_STRIP_WIDTH_FACTOR = 1.10
 pdf_cache_lock = threading.Lock()
 
 # The requested spelling is kept as the public switch for this alternate layout.
@@ -116,6 +118,9 @@ def process_single_pdf(
     shift_offset: int = 0,
     max_per_file: int = 5,
     width_offset: int = 0,
+    *,
+    automatic_margin: bool = True,
+    margin_analysis: dict | None = None,
 ) -> tuple[Image.Image, list[Image.Image], dict[str, object]] | None:
     """Convert one PDF into a preview, full-resolution splits and metadata."""
     try:
@@ -124,7 +129,10 @@ def process_single_pdf(
         if not images:
             return None
         reference_width = normal_page_width(pdf_path, images)
+        if automatic_margin and max_per_file != 1 and margin_analysis is None:
+            margin_analysis = analyse_margins(images, RENDER_DPI, reference_width)
         full_width_pages = []
+        page_crops = []
         for page_index, image in enumerate(images):
             width, height = image.size
             if max_per_file == 1:
@@ -140,8 +148,17 @@ def process_single_pdf(
                     width,
                     width // 3 + 100 + shift_offset + max(0, width_offset),
                 )
+            if automatic_margin and max_per_file != 1:
+                left = min(width - 1, max(0, margin_analysis['starts_px'][page_index] + shift_offset))
+                strip_width = round(
+                    ((reference_width if width < NARROW_PAGE_RATIO * reference_width else width) // 3)
+                    * AUTO_STRIP_WIDTH_FACTOR
+                )
+                right = min(width, left + strip_width + max(0, width_offset))
             if right > left:
                 cropped_images.append(image.crop((left, 0, right, height)))
+                page_crops.append({'page': page_index + 1, 'left_px': left, 'right_px': right,
+                                   'width_px': width, 'height_px': height})
                 if left == 0 and right == width:
                     full_width_pages.append(page_index + 1)
         if not cropped_images:
@@ -165,6 +182,8 @@ def process_single_pdf(
             "number_of_files": len(split_images),
             "columns_per_file": distribution,
             "full_width_pages": full_width_pages,
+            "page_crops": page_crops,
+            "auto_margin": margin_analysis if automatic_margin and max_per_file != 1 else None,
         }
         return preview, split_images, schema
     except Exception as exc:  # noqa: BLE001 - interactive item failure
@@ -208,6 +227,7 @@ class ImageReviewer:
         files: list[Path],
         output_dir: Path,
         default_max_per_file: int = 5,
+        automatic_margin: bool = True,
     ) -> None:
         self.files = files
         self.output_dir = output_dir
@@ -216,6 +236,8 @@ class ImageReviewer:
         self.had_errors = False
         self.stop_prefetch = threading.Event()
         self.current_result = None
+        self.automatic_margin = automatic_margin
+        self.margin_analysis = None
         self.index = 0
         self.current_shift = 0
         self.current_width_offset = 0
@@ -283,6 +305,7 @@ class ImageReviewer:
         self.current_shift = 0
         self.current_width_offset = 0
         self.current_result = None
+        self.margin_analysis = None
         self.trigger_processing(self.files[self.index], self.current_shift)
 
     def trigger_processing(self, pdf_path: Path, shift: int) -> None:
@@ -299,6 +322,8 @@ class ImageReviewer:
                     shift,
                     self.current_max_per_file,
                     self.current_width_offset,
+                    automatic_margin=self.automatic_margin,
+                    margin_analysis=self.margin_analysis,
                 )
             )
 
@@ -325,6 +350,8 @@ class ImageReviewer:
     ) -> None:
         self.current_preview = result[0]
         self.current_result = result
+        if result[2].get('auto_margin') is not None:
+            self.margin_analysis = result[2]['auto_margin']
         self.update_display(pdf_path.name, result[2])
 
     def update_display(self, filename: str, schema: dict[str, object]) -> None:
@@ -338,6 +365,17 @@ class ImageReviewer:
             if personnal_binding
             else ""
         )
+        analysis = schema.get('auto_margin')
+        margin_info = ''
+        if analysis:
+            offsets = sorted({round(crop['left_px'] * 25.4 / RENDER_DPI, 1)
+                              for crop in schema['page_crops']})
+            needs_review = any(group['status'] in {'inconsistent', 'uncertain', 'insufficient-ink'}
+                               for group in analysis['groups'])
+            margin_info = (f"Auto margin: {offsets} mm from left"
+                           + (' | Check uncertain pages' if needs_review else '') + '\n')
+            if analysis.get('no_margin_from_page') is not None:
+                margin_info += f"No margin from page {analysis['no_margin_from_page']} onwards\n"
         self.label_info.configure(
             text=(
                 f"[{self.index + 1}/{len(self.files)}] {filename} | "
@@ -345,6 +383,7 @@ class ImageReviewer:
                 f"Extra width: {self.current_width_offset}px\n"
                 f"Files: {schema['number_of_files']} | "
                 f"Cols: {schema['columns_per_file']}\n"
+                f"{margin_info}"
                 "Arrows: left/right move by 50px | up/down widen/narrow by 50px\n"
                 "Enter: Save and next | !: flag error and skip | 1: use full pages"
                 f"{personal_instructions}"
@@ -428,6 +467,7 @@ def run(
     fullpage: bool = False,
     no_visual_check: bool = False,
     marked: bool = False,
+    automatic_margin: bool = True,
 ) -> ExitCode:
     files = marked_copy_paths(workspace) if marked else _selected_files(workspace, target)
     if not files:
@@ -463,6 +503,7 @@ def run(
         files,
         workspace.cutleft_dir,
         default_max_per_file=1 if fullpage else 5,
+        automatic_margin=automatic_margin,
     )
     if not reviewer.completed:
         return ExitCode.INTERRUPTED
@@ -472,6 +513,10 @@ def run(
 def build_parser() -> argparse.ArgumentParser:
     parser = target_parser("Interactively crop the label margin from PDF copies")
     parser.add_argument("--marked", action="store_true", help="Review flagged copies and clear only margin-cropping flags after saving with Enter")
+    parser.add_argument(
+        '--manual-margin', action='store_true',
+        help='Use the previous fixed margin instead of automatically estimating body-text positions',
+    )
     parser.add_argument(
         "--fullpage",
         action="store_true",
@@ -496,6 +541,7 @@ def main(argv: Sequence[str] | None = None) -> int:
             fullpage=args.fullpage,
             no_visual_check=args.no_visual_check,
             marked=args.marked,
+            automatic_margin=not args.manual_margin,
         )
 
     return execute(parser, argv, handle)

copienator/left_margin.py

@@ -0,0 +1,151 @@
+"""Conservative body-text margin suggestions for the cutleft reviewer."""
+from __future__ import annotations
+
+import cv2
+import numpy as np
+from PIL import Image
+
+from copienator.ink_detection import detect_bounds
+
+ANALYSIS_DPI = 100
+WIDTH_TOLERANCE = .08
+# Copie01's name/class block reaches about 27% of the full page height.
+# Start below it with clearance, including on later pages with repeated names.
+BODY_TOP = .30
+BODY_BOTTOM = .82
+DISCARD_LEFT_FRACTION = .10
+STANDARD_MARGIN_MM = 40
+MIN_MARGIN_MM = STANDARD_MARGIN_MM / 2
+
+
+def trimmed_left_edge(starts: list[float]) -> float | None:
+    """Discard the leftmost 10% of occupied bands; retain the next minimum."""
+    if not starts:
+        return None
+    ordered = sorted(starts)
+    return round(ordered[int(len(ordered) * DISCARD_LEFT_FRACTION)], 1)
+
+
+def handwriting_mask(rgb: np.ndarray) -> np.ndarray:
+    """Intersect rognage's foreground with a directional ruling filter."""
+    px = ANALYSIS_DPI / 25.4
+    gray = cv2.cvtColor(rgb, cv2.COLOR_RGB2GRAY)
+    darkness = 255 - rgb.min(axis=2)
+    length = round(5.2 * px) | 1
+    horizontal = cv2.morphologyEx(darkness, cv2.MORPH_OPEN, np.ones((1, length), np.uint8))
+    vertical = cv2.morphologyEx(darkness, cv2.MORPH_OPEN, np.ones((length, 1), np.uint8))
+    residual = cv2.subtract(darkness, np.maximum(horizontal, vertical))
+    binary = ((residual > 32) & (gray < 200)).astype(np.uint8)
+    count, labels, stats, _ = cv2.connectedComponentsWithStats(binary, 8)
+    keep = np.zeros(count, bool)
+    for index, (_x, _y, width, height, area) in enumerate(stats[1:], 1):
+        keep[index] = (area >= max(5, .35 * px * px) and height >= .5 * px
+                       and width >= .25 * px and height <= 6 * width and width <= 15 * height)
+    diagnostics = {}
+    detect_bounds(rgb, ANALYSIS_DPI, diagnostics=diagnostics)
+    return keep[labels] & diagnostics['masks']['foreground']
+
+
+def body_edge(mask: np.ndarray) -> dict:
+    """Estimate a left edge from occupied 4 mm bands within 30–82% height."""
+    height = mask.shape[0]
+    band_height = max(1, round(4 * ANALYSIS_DPI / 25.4))
+    bottom = round(BODY_BOTTOM * height)
+    starts = []
+    for y in range(round(BODY_TOP * height), bottom, band_height):
+        band = mask[y:min(y + band_height, bottom)]
+        xs = np.flatnonzero(band.sum(axis=0) >= 2)
+        if band.sum() >= 40 and len(xs) >= 12:
+            starts.append(float(xs[0]) * 25.4 / ANALYSIS_DPI)
+    return {
+        'edge_mm': trimmed_left_edge(starts) if len(starts) >= 6 else None,
+        'bands': len(starts),
+        'band_starts_mm': starts,
+    }
+
+
+def estimate_page(image: Image.Image, render_dpi: int) -> dict:
+    scale = ANALYSIS_DPI / render_dpi
+    small = image.convert('RGB').resize(
+        (max(1, round(image.width * scale)), max(1, round(image.height * scale))),
+        Image.Resampling.LANCZOS,
+    )
+    return body_edge(handwriting_mask(np.asarray(small)))
+
+
+def format_groups(images: list[Image.Image], reference_width: float) -> list[list[int]]:
+    """Group similar widths; keep secondary narrow subpages in separate groups."""
+    groups: list[list[int]] = []
+    for index, image in enumerate(images):
+        narrow = image.width < .8 * reference_width
+        for group in groups:
+            width = float(np.median([images[i].width for i in group]))
+            if (images[group[0]].width < .8 * reference_width) == narrow and abs(image.width - width) <= width * WIDTH_TOLERANCE:
+                group.append(index)
+                break
+        else:
+            groups.append([index])
+    return groups
+
+
+def analyse_margins(images: list[Image.Image], render_dpi: int, reference_width: float) -> dict:
+    """Pool occupied bands per format, trim outliers, and crop at half the edge."""
+    starts = [0] * len(images)
+    groups = []
+    page_estimates = {}
+
+    def page_estimate(index):
+        if index not in page_estimates:
+            page_estimates[index] = estimate_page(images[index], render_dpi)
+        return page_estimates[index]
+
+    for indices in format_groups(images, reference_width):
+        # Spread samples through the copy rather than relying on the header page.
+        sample = sorted({indices[0], indices[min(1, len(indices) - 1)],
+                         indices[len(indices) // 2], indices[-1]})
+        estimates = {index: page_estimate(index) for index in sample}
+        values = [item['edge_mm'] for item in estimates.values() if item['edge_mm'] is not None]
+        # Trim across the copy's format group, not separately on each page:
+        # one marginal comment must not veto the other pages' shared margin.
+        band_starts = [start for item in estimates.values() for start in item['band_starts_mm']]
+        status = 'insufficient-ink'
+        edge = trimmed_left_edge(band_starts)
+        left_mm = 0.0
+        if len(values) >= min(2, len(indices)):
+            status = 'near-edge' if edge < MIN_MARGIN_MM else 'margin'
+            left_mm = 0.0 if edge < MIN_MARGIN_MM else edge / 2
+        # Secondary subpages can start in the middle of an answer. Preserve
+        # their left edge even when the only visible writing is indented.
+        if images[indices[0]].width < .8 * reference_width:
+            status = 'narrow-page'
+            left_mm = 0.0
+        start = round(left_mm * render_dpi / 25.4)
+        for index in indices:
+            starts[index] = start
+        groups.append({
+            'pages': [index + 1 for index in indices],
+            'status': status,
+            'edge_mm': edge,
+            'left_mm': round(left_mm, 1),
+            'bands': len(band_starts),
+            'discarded_bands': int(len(band_starts) * DISCARD_LEFT_FRACTION),
+            'samples': [{'page': index + 1, **item} for index, item in estimates.items()],
+        })
+    # Inspect actual neighbours, including pages outside the format samples.
+    # Blank/uncertain pages do not establish or extend a no-margin run.
+    no_margin_run = 0
+    no_margin_from_page = None
+    for index in range(len(images)):
+        edge = page_estimate(index)['edge_mm']
+        no_margin_run = no_margin_run + 1 if edge is not None and edge < MIN_MARGIN_MM else 0
+        if no_margin_run >= 2:
+            run_start = index - 1
+            no_margin_from_page = run_start + 1
+            starts[run_start:] = [0] * (len(images) - run_start)
+            break
+
+    return {'version': 5, 'starts_px': starts, 'groups': groups,
+            'no_margin_from_page': no_margin_from_page,
+            'discard_left_fraction': DISCARD_LEFT_FRACTION, 'cut_fraction_of_margin': .5,
+            'standard_margin_mm': STANDARD_MARGIN_MM, 'minimum_margin_mm': MIN_MARGIN_MM,
+            'body_height_fraction': [BODY_TOP, BODY_BOTTOM]}

copienator_gui/workflow.py

@@ -289,6 +289,12 @@ def build_workflow(show_personal_steps: bool) -> list[StepDefinition]:
             "Prétraitement des copies",
             "Découper une partie à gauche pour détection des labels",
             "Produit les images de la partie gauche des copies. "
+            "La position est estimée automatiquement sur plusieurs pages de même largeur ; "
+            "Les 10 % de bandes d’écriture les plus à gauche sont ignorées ; "
+            "la découpe commence à la moitié de la marge restante. "
+            "Si la marge est inférieure à 20 mm, elle commence au bord gauche. "
+            "Deux pages consécutives sans marge imposent le bord gauche pour la suite. "
+            "La bande est élargie de 10 %. "
             "Pour les sous-pages étroites, la marge commence au bord gauche et est limitée à la largeur disponible. "
             "Dans la fenêtre de découpe, utilisez les flèches : ←/→ déplacent la zone "
             "de 50 px vers la gauche/droite, ↑ l’élargit de 50 px et ↓ la réduit de 50 px. "
@@ -299,6 +305,10 @@ def build_workflow(show_personal_steps: bool) -> list[StepDefinition]:
             (python("default", "Découpe", "crop-labels"),),
             arguments=(
                 arg_target(),
+                ArgumentSpec(
+                    "manual_margin", "Utiliser la marge fixe", "bool", "--manual-margin",
+                    help="Désactive l’estimation automatique et reprend la position fixe habituelle, ajustable avec les flèches.",
+                ),
                 ArgumentSpec(
                     "fullpage",
                     "Toujours utiliser la page entière",

docs/guide_utilisateur_gui.md

@@ -94,11 +94,32 @@ cette étape non interactive, cocher aussi `et ne pas vérifier visuellement`.
 Ce mode est plus cher : sans avoir vérifié, Astra estime que le coût
 passerait de 0,50€ à 2€.
 
-Sans quoi, pour chaque copie, il faut cadrer la bande pour qu'elle
+La position de la bande est proposée automatiquement à partir de plusieurs
-contienne bien tous les labels :
+pages de largeur comparable. L’estimation cherche le début de l’écriture
+dans la partie centrale des pages (30–82 % de la hauteur), en ignorant le haut
+et le bas. Les positions de début d’écriture des bandes sont regroupées pour
+les pages de même format. Les 10 % de bandes les plus à gauche sont ignorées ;
+le début le plus à gauche restant définit la marge. La découpe commence à la
+moitié de cette distance. Une marge inférieure à 20 mm (la moitié d’une marge
+de référence de 40 mm) commence au bord gauche. Les pages étroites et les groupes sans assez
+d’écriture restent au bord gauche. La même règle s’applique à toutes les pages,
+y compris la première.
+
+Deux pages consécutives dont la marge détectée est inférieure à 20 mm imposent
+le bord gauche pour ces deux pages et toutes les suivantes, même si les pages
+suivantes semblent plus indentées. Les pages sans assez d’écriture ne comptent
+pas comme des pages sans marge. La largeur automatique de la bande est augmentée
+de 10 %, puis les ajustements manuels de largeur sont ajoutés.
+
+La fenêtre indique les positions proposées et les cas à vérifier. Vérifiez
+pour chaque copie que la bande contient bien tous les labels :
 
 **←/→** la déplacent, **↑/↓** l’élargissent ou la réduisent ; **Entrée** valide.
 
+Ces réglages ajustent la proposition automatique sans relancer l’analyse.
+Pour reprendre l’ancienne position fixe, cochez **Utiliser la marge fixe**
+(option `--manual-margin` en ligne de commande).
+
 Si les labels sont dispersés, utilisez **1** pour utiliser les pages entières pour cette copie.
 
 Cette vérification visuelle permet toutefois de repérer des copies à

tests/test_cutleft_narrow_pages.py

@@ -17,7 +17,7 @@ class NarrowPageCropTests(unittest.TestCase):
         narrow.paste('blue', (0, 0, 300, 600))
         thin = Image.new('RGB', (200, 600), 'yellow')
         with patch.object(cutleft, 'get_pdf_pages', return_value=[normal, narrow, thin]):
-            result = cutleft.process_single_pdf(Path('/tmp/no-original/Copy.pdf'))
+            result = cutleft.process_single_pdf(Path('/tmp/no-original/Copy.pdf'), automatic_margin=False)
         _, images, schema = result
         image = images[0]
         self.assertEqual(schema['full_width_pages'], [3])
@@ -39,7 +39,7 @@ class NarrowPageCropTests(unittest.TestCase):
         narrow.paste('blue', (50, 0, 450, 600))
         with patch.object(cutleft, 'get_pdf_pages', return_value=[normal, narrow]):
             result = cutleft.process_single_pdf(
-                Path('/tmp/no-original/Copy.pdf'), shift_offset=50, width_offset=100)
+                Path('/tmp/no-original/Copy.pdf'), shift_offset=50, width_offset=100, automatic_margin=False)
         self.assertEqual(result[1][0].size, (805, 600))
         self.assertEqual(result[1][0].getpixel((405, 0)), (0, 0, 255))
         self.assertEqual(result[1][0].getpixel((804, 0)), (0, 0, 255))
@@ -54,14 +54,14 @@ class NarrowPageCropTests(unittest.TestCase):
                 document.save(originals / 'Copy.pdf')
             images = [Image.new('RGB', (900, 2339)), Image.new('RGB', (750, 2339))]
             with patch.object(cutleft, 'get_pdf_pages', return_value=images):
-                result = cutleft.process_single_pdf(root / 'Copies' / 'Copy.pdf')
+                result = cutleft.process_single_pdf(root / 'Copies' / 'Copy.pdf', automatic_margin=False)
             self.assertEqual(result[2]['full_width_pages'], [])
             self.assertEqual(result[1][0].width, 2 * int((1191 / 2 * 200 / 72) // 3) + 5)
 
     def test_small_main_split_adjustments_keep_margin_cropping(self):
         images = [Image.new('RGB', (1600, 2339)), Image.new('RGB', (1694, 2339))]
         with patch.object(cutleft, 'get_pdf_pages', return_value=images):
-            result = cutleft.process_single_pdf(Path('/tmp/no-original/Copy.pdf'))
+            result = cutleft.process_single_pdf(Path('/tmp/no-original/Copy.pdf'), automatic_margin=False)
         self.assertEqual(result[2]['full_width_pages'], [])
         self.assertEqual(result[1][0].width, 2 * (1694 // 3) + 5)
 

tests/test_left_margin.py

@@ -0,0 +1,184 @@
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+import numpy as np
+from PIL import Image, ImageDraw, ImageFont
+
+from copienator import left_margin
+from copienator.commands import cutleft
+
+
+class LeftMarginTests(unittest.TestCase):
+    def test_trim_discards_exactly_ten_percent_without_interpolation(self):
+        self.assertEqual(left_margin.trimmed_left_edge([100, 1, 9, 8, 7, 6, 5, 4, 3, 2]), 2)
+        self.assertEqual(left_margin.trimmed_left_edge([5] * 9), 5)
+        self.assertIsNone(left_margin.trimmed_left_edge([]))
+
+    def test_one_margin_comment_does_not_veto_other_pages(self):
+        images = [Image.new('RGB', (1654, 200)) for _ in range(4)]
+        bands = [[42] * 40, [42] * 40, [42] * 40, [4] * 8 + [42] * 12]
+        estimates = [{'edge_mm': left_margin.trimmed_left_edge(values),
+                      'bands': len(values), 'band_starts_mm': values} for values in bands]
+        with patch.object(left_margin, 'estimate_page', side_effect=estimates):
+            result = left_margin.analyse_margins(images, 200, 1654)
+        group = result['groups'][0]
+        self.assertEqual(group['edge_mm'], 42)
+        self.assertEqual(group['left_mm'], 21)
+        self.assertEqual(group['discarded_bands'], 14)
+        self.assertEqual(result['starts_px'], [round(21 * 200 / 25.4)] * 4)
+
+    def test_repeated_left_writing_beyond_ten_percent_is_retained(self):
+        self.assertEqual(left_margin.trimmed_left_edge([10] * 20 + [42] * 80), 10)
+
+    def test_disagreeing_pages_use_half_the_trimmed_margin(self):
+        result = self.analyse([25, 31, 68, 49])
+        self.assertEqual(result['groups'][0]['left_mm'], 12.5)
+        self.assertEqual(result['starts_px'][1:], [round(12.5 * 200 / 25.4)] * 3)
+
+    def test_small_margin_starts_at_left(self):
+        result = self.analyse([10, 15, 8, 13])
+        self.assertEqual(result['groups'][0]['left_mm'], 0)
+        self.assertEqual(result['starts_px'], [0] * 4)
+
+    def test_two_consecutive_no_margin_pages_enforce_left_edge_afterwards(self):
+        images = [Image.new('RGB', (1654, 200)) for _ in range(8)]
+        # Low pages have fewer occupied bands, so pooling still proposes a margin.
+        edges = [42, 42, 10, 12, 42, 42, 42, 42]
+        by_id = {id(image): edge for image, edge in zip(images, edges)}
+        def estimate(image, _dpi):
+            edge = by_id[id(image)]
+            bands = 6 if edge < 20 else 100
+            return {'edge_mm': edge, 'bands': bands, 'band_starts_mm': [edge] * bands}
+        with patch.object(left_margin, 'estimate_page', side_effect=estimate):
+            result = left_margin.analyse_margins(images, 200, 1654)
+        self.assertEqual(result['groups'][0]['left_mm'], 21)
+        self.assertEqual(result['no_margin_from_page'], 3)
+        self.assertEqual(result['starts_px'], [round(21 * 200 / 25.4)] * 2 + [0] * 6)
+
+    def test_isolated_low_or_blank_pages_do_not_trigger_no_margin_rule(self):
+        for edges in ([42, 10, 42, 12], [42, 10, None, 12]):
+            with self.subTest(edges=edges):
+                result = self.analyse(edges)
+                self.assertIsNone(result['no_margin_from_page'])
+
+    def test_two_no_margin_pages_enforce_rule_across_later_width_changes(self):
+        result = self.analyse([42, 42, 10, 12, 42, 42],
+                              widths=[1654, 1654, 1654, 1654, 2000, 2000],
+                              band_counts=[100, 100, 6, 6, 100, 100])
+        self.assertEqual(result['no_margin_from_page'], 3)
+        self.assertGreater(result['starts_px'][0], 0)
+        self.assertEqual(result['starts_px'][2:], [0] * 4)
+
+    def test_half_standard_margin_threshold(self):
+        for edge, expected in ((19.9, 0), (20, 10), (20.1, 10.05), (40, 20)):
+            with self.subTest(edge=edge):
+                result = self.analyse([edge] * 4)
+                self.assertEqual(result['starts_px'][1:], [round(expected * 200 / 25.4)] * 3)
+
+    def test_body_bands_ignore_header_footer_and_abstain_on_sparse_ink(self):
+        mask = np.zeros((1169, 827), bool)
+        mask[60:340, 2:120] = True  # Name/class block extends past 25% height.
+        mask[1000:1050, 2:50] = True  # Pagination.
+        for y in range(330, 940, 35):
+            mask[y:y + 5, 170:270] = True
+        result = left_margin.body_edge(mask)
+        self.assertAlmostEqual(result['edge_mm'], 43.2, delta=.3)
+        self.assertGreaterEqual(result['bands'], 6)
+        mask[300:960] = False
+        self.assertIsNone(left_margin.body_edge(mask)['edge_mm'])
+
+    def test_real_mask_rejects_ruling_and_preserves_handwriting(self):
+        image = Image.new('RGB', (827, 1169), 'white')
+        draw = ImageDraw.Draw(image)
+        for x in range(12, 827, 20):
+            draw.line((x, 0, x, 1169), fill=(170, 190, 210))
+        for y in range(10, 1169, 20):
+            draw.line((0, y, 827, y), fill=(170, 190, 210))
+        font = ImageFont.truetype('DejaVuSans.ttf', 16)
+        for y in range(330, 920, 35):
+            draw.text((170, y), '1) Pour tout x, f(x) = x + 1', font=font, fill=(20, 20, 100))
+        result = left_margin.estimate_page(image, 100)
+        self.assertAlmostEqual(result['edge_mm'], 43.2, delta=3)
+
+    def test_width_groups_tolerate_split_adjustments_and_separate_narrow_pages(self):
+        pages = [Image.new('RGB', (w, 200)) for w in (1600, 1694, 1654, 800, 810, 2000)]
+        self.assertEqual(left_margin.format_groups(pages, 1654), [[0, 1, 2], [3, 4], [5]])
+
+    def analyse(self, edges, widths=None, band_counts=None):
+        images = [Image.new('RGB', (w, 200)) for w in (widths or [1654] * len(edges))]
+        by_id = {id(image): edge for image, edge in zip(images, edges)}
+        counts = {id(image): count for image, count in zip(images, band_counts or [12] * len(edges))}
+        def estimate(image, _dpi):
+            edge = by_id[id(image)]
+            return {'edge_mm': edge, 'bands': counts[id(image)] if edge is not None else 0,
+                    'band_starts_mm': [edge] * counts[id(image)] if edge is not None else []}
+        with patch.object(left_margin, 'estimate_page', side_effect=estimate):
+            return left_margin.analyse_margins(images, 200, max(im.width for im in images))
+
+    def test_clear_margin_uses_same_conservative_edge_on_every_page(self):
+        result = self.analyse([43, 41, 46, 44])
+        self.assertEqual(result['groups'][0]['status'], 'margin')
+        self.assertEqual(result['starts_px'], [round(20.5 * 200 / 25.4)] * 4)
+        self.assertNotIn('first_page_body_start_px', result)
+
+    def test_sparse_groups_stay_at_left(self):
+        for edges in ([None] * 4, [40, None, None, None]):
+            with self.subTest(edges=edges):
+                self.assertEqual(self.analyse(edges)['starts_px'], [0] * 4)
+
+    def test_narrow_subpages_stay_at_left_even_with_indented_text(self):
+        result = self.analyse([45, 45, 70, 72], widths=[1654, 1654, 800, 800])
+        self.assertEqual(result['starts_px'][2:], [0, 0])
+        self.assertEqual(result['groups'][1]['status'], 'narrow-page')
+
+    def test_auto_crop_applies_manual_offsets_to_cached_estimate(self):
+        images = [Image.new('RGB', (900, 600), 'white') for _ in range(2)]
+        images[0].paste('blue', (150, 0, 190, 100))
+        analysis = {'starts_px': [150, 150], 'groups': []}
+        with patch.object(cutleft, 'get_pdf_pages', return_value=images), \
+                patch.object(cutleft, 'analyse_margins', return_value=analysis) as analyse:
+            result = cutleft.process_single_pdf(Path('/tmp/left-margin/Copy.pdf'))
+            shifted = cutleft.process_single_pdf(Path('/tmp/left-margin/Copy.pdf'),
+                                                shift_offset=-50, width_offset=50,
+                                                margin_analysis=analysis)
+        analyse.assert_called_once()
+        self.assertEqual(result[1][0].getpixel((0, 0)), (0, 0, 255))
+        self.assertEqual([p['left_px'] for p in result[2]['page_crops']], [150, 150])
+        self.assertEqual([p['left_px'] for p in shifted[2]['page_crops']], [100, 100])
+        self.assertEqual(result[2]['columns_per_file'], [2])
+        self.assertEqual(result[1][0].size, (665, 600))
+        self.assertEqual(shifted[1][0].size, (765, 600))
+
+    def test_fullpage_bypasses_analysis(self):
+        with patch.object(cutleft, 'get_pdf_pages', return_value=[Image.new('RGB', (900, 600))]), \
+                patch.object(cutleft, 'analyse_margins') as analyse:
+            result = cutleft.process_single_pdf(Path('/tmp/left-margin/Copy.pdf'), max_per_file=1)
+        analyse.assert_not_called()
+        self.assertIsNone(result[2]['auto_margin'])
+
+    def test_wider_auto_strip_is_capped_at_page_edge(self):
+        images = [Image.new('RGB', (900, 600)), Image.new('RGB', (200, 600))]
+        with patch.object(cutleft, 'get_pdf_pages', return_value=images):
+            result = cutleft.process_single_pdf(Path('/tmp/left-margin/Copy.pdf'),
+                                                margin_analysis={'starts_px': [750, 0], 'groups': []})
+        self.assertEqual([crop['right_px'] for crop in result[2]['page_crops']], [900, 200])
+        self.assertEqual(result[2]['full_width_pages'], [2])
+
+    def test_first_page_uses_same_crop_width_as_later_pages(self):
+        images = [Image.new('RGB', (900, 600), 'white') for _ in range(2)]
+        images[0].paste('blue', (420, 100, 450, 200))
+        analysis = {'starts_px': [150, 150], 'groups': []}
+        with patch.object(cutleft, 'get_pdf_pages', return_value=images):
+            result = cutleft.process_single_pdf(Path('/tmp/left-margin/Copy.pdf'),
+                                                margin_analysis=analysis)
+        self.assertEqual(result[2]['page_crops'][0]['right_px'], 480)
+        self.assertEqual(result[1][0].getpixel((290, 120)), (0, 0, 255))
+        self.assertEqual(result[1][0].size, (665, 600))
+        self.assertEqual([crop['left_px'] for crop in result[2]['page_crops']], [150, 150])
+
+    def test_manual_margin_cli_disables_automatic_positioning(self):
+        with patch.object(cutleft, 'run', return_value=0) as run, \
+                patch.object(cutleft, 'workspace_from_target', return_value=(None, Path('/tmp'))):
+            self.assertEqual(cutleft.main(['/tmp', '--manual-margin']), 0)
+        self.assertFalse(run.call_args.kwargs['automatic_margin'])

tests/test_marked_copies.py

@@ -115,9 +115,9 @@ class MarkedCopiesTests(unittest.TestCase):
     def test_process_single_pdf_enlarges_crop_and_caps_it_at_page_edge(self):
         page = Image.new("RGB", (900, 300), "white")
         with patch.object(cutleft, "get_pdf_pages", return_value=[page]):
-            regular = cutleft.process_single_pdf(self.files[0])
+            regular = cutleft.process_single_pdf(self.files[0], automatic_margin=False)
-            enlarged = cutleft.process_single_pdf(self.files[0], width_offset=200)
+            enlarged = cutleft.process_single_pdf(self.files[0], width_offset=200, automatic_margin=False)
-            capped = cutleft.process_single_pdf(self.files[0], width_offset=1000)
+            capped = cutleft.process_single_pdf(self.files[0], width_offset=1000, automatic_margin=False)
             fullpage = cutleft.process_single_pdf(
                 self.files[0], max_per_file=1, width_offset=200
             )