fix(magazine): derive figure boundaries from the page render

A vision OCR engine grounds a figure far less precisely than it
transcribes one. Over all 138 figure regions of a 204-page issue the
transcribed contents were right while the boxes were routinely offset,
undersized, or sitting on the neighbouring text column -- page 23's card
grid was reported at x 0.24-0.54 where the graphic runs 0.41-0.95.

snap_visual_bboxes() reduces the page render to an ink-coverage grid,
seeds a connected-component search with the engine's box, and re-joins
tiles a gutter apart so a card grid comes back as one figure. Text
blocks are excluded from the mask, which is what stops a figure
swallowing the column beside it. A candidate must be solidly inked or
framed on nearly all of its rows and columns (admitting a white-interior
screenshot), may not shrink below 30% of the seed, and reverts when two
figures land on the same graphic; anything else keeps the box it had.

Runs before the figure crop in both OCR paths, so the images shown
during playback are framed correctly too. Replaces refine_picture_
bounds(), which only recognised coloured panels and had no text fence.

Also adds "Fix figure boundaries" in review and POST
/api/magazine/{id}/figures/refine behind it, so an issue ingested before
this pass can be corrected without re-running OCR. Verified live against
c't KI Wissen 2026: 78 of 138 boxes corrected, the rest declined.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

CHANGELOG.md

@@ -5,6 +5,23 @@ Follows [Keep a Changelog](https://keepachangelog.com/en/1.0.0/) · versioned wi
 
 ---
 
+## [1.27.0] — 2026-09-12
+
+### Fixed
+- **Figure boundaries are now derived from the scanned page, not from the OCR engine's guess.**
+  - A vision OCR engine locates a photo far less precisely than it reads type: over all 138 figure regions of a 204-page issue, figure boxes were routinely offset, undersized, or sitting on the neighbouring text column while the transcribed text of the same regions was correct.
+  - New `snap_visual_bboxes()` pass reduces the page render to an ink-coverage grid, seeds a connected-component search with the engine's box, and re-joins tiles a gutter apart — so a grid of cards or a chart beside its key comes back as one figure. Text blocks are excluded from the mask, which stops a figure swallowing the column beside it.
+  - Conservative by design: a candidate must be solidly inked *or* framed on nearly all of its rows and columns (which is what admits a white-interior UI screenshot), may not shrink below 30% of the engine's box, and reverts when two figures are pulled onto the same graphic. Anything it cannot place confidently keeps the box it had.
+  - Runs before the figure crop in both OCR paths, so the images shown during playback are framed correctly too. Corrected 78 of 138 figure boxes on the reference issue.
+  - Replaces `refine_picture_bounds()`, which only recognised coloured panels and had no text fence.
+
+### Added
+- **"Fix figure boundaries" in magazine review**, and `POST /api/magazine/{id}/figures/refine` behind it.
+  - Re-snaps every figure box in an issue as a restart-safe background job. No OCR engine is called, so an issue ingested before this release is corrected without paying for its transcription again.
+  - Text, article grouping and existing approvals are untouched — safe to run halfway through a review. Reports how many boundaries moved.
+
+---
+
 ## [1.26.14] — 2026-09-10
 
 ### Added

Manual.md

@@ -47,7 +47,14 @@ turned into a podcast discussion between two or three voices you pick.
    mistaken for unclaimed text. If OCR quality is poor, **"OCR again"**
    re-reads just the current page through the configured vision engine
    (or **"OCR again (whole magazine)"** from the contents screen for every
-   page, which also resets any articles already reviewed). If the layout
+   page, which also resets any articles already reviewed). If a figure's
+   box is drawn in the wrong place or at the wrong size — a common miss,
+   because an OCR engine locates a photo far less precisely than it reads
+   type — **"Fix figure boundaries"** re-derives every figure box in the
+   issue from the scanned pages themselves. It calls no OCR engine, changes
+   no text and touches no approval, so it is safe to run at any point,
+   including on an issue you are halfway through reviewing; boxes it cannot
+   place confidently are left exactly as they were. If the layout
    engine missed a region entirely, **"Draw it"** lets you click-drag a box
    directly on the page image and add it as text (typed or OCRed) or as a
    new figure. Confirm the boundaries, correct the title, then **Approve**.
@@ -99,6 +106,8 @@ unreviewed boundaries may include a neighbouring article or an advert.
 | "No LLM model configured" | `llm_model` is empty. | Set it in Settings. |
 | "LLM did not return usable JSON" | The model ignored the format. | Retry, or pick a stronger model. |
 | "Article has too little text" | Under 200 characters — usually a stub. | Check the boundaries in review. |
+| A figure box sits over the wrong part of the page | The OCR engine grounded the figure loosely. | Press **"Fix figure boundaries"**; if the graphic is too pale to detect, draw the region by hand. |
+| "Fix figure boundaries" reports nothing corrected | Every box already matched its page, or none could be placed with confidence. | Nothing is wrong; correct any remaining box by drawing it. |
 
 ### Privacy
 

PROJECT_MAP.md

@@ -78,6 +78,11 @@ which is the writable volume.
   real measured audio duration, not true forced alignment.
 - No formal cross-engine OCR accuracy benchmark exists (two engines wired:
   ocr-baidu-unlimited-ocr primary, KDL Frontier Parser nano fallback).
+- Figure boundaries are re-derived from the page render (`snap_visual_bboxes`)
+  because vision OCR grounds figures loosely; it corrected 78 of 138 figure
+  boxes on `c't KI Wissen 2026` and deliberately declines the rest. A pale
+  chart or line drawing on white is below its ink threshold and keeps the
+  engine's box — draw those by hand in review.
 - No `npm run qa` aggregate command yet; Python tests run via `unittest`,
   browser tests via `playwright test`.
 - The app has no authentication for the browser UI and no per-user data

VERSION

@@ -1 +1 @@
-1.26.14
+1.27.0

architecture/magazine-ingestion.md

@@ -34,6 +34,9 @@ PDF
                           the vision call failed/timed out/was unreachable
  → build_vision_blocks()  vision regions -> Blocks; a direct label bypasses
                           classify_blocks entirely (see below)
+ → snap_visual_bboxes()   figure/chart/table/code boxes re-derived from the page
+                          render, because grounding is the weakest part of a
+                          vision engine's answer (see below)
  → classify_blocks()      title / deck / byline / body / caption / pullquote /
                           sidebar / furniture / ad / figure -- for native/tesseract
                           blocks, and any vision region left ambiguous ("text")
@@ -357,6 +360,58 @@ fallback and not the default.
 
 A formal cross-engine accuracy benchmark remains TODO P2.
 
+### Figure bounds come from the render, not the engine
+
+A vision OCR engine grounds a figure far less precisely than it transcribes
+one. Measured over all 138 figure regions of `c't KI Wissen 2026` (204 pages,
+Baidu Unlimited-OCR): text boxes matched the page render closely enough to
+use as-is, while figure boxes were routinely offset, undersized, or landed on
+the neighbouring text column — page 23's card grid was reported at
+`x 0.24–0.54` of the page where the graphic actually runs `0.41–0.95`, and
+page 20's full-width illustration was cut off at a third of its height. The
+transcribed *contents* of those same regions were right, which is the useful
+signal: the model reliably says **which** graphic it means, so the render is
+asked **where** that graphic ends.
+
+`snap_visual_bboxes()` (core/magazine.py) does that, on a coarse grid of the
+rendered page — `page_ink_grid()` reduces it to the inked fraction of each
+10 px cell, which separates graphics from type cleanly: body copy covers
+roughly a quarter of the cells it runs through, a photo or filled panel
+covers nearly all of them. The engine's box seeds a connected-component
+search over the "solid" cells (scipy `ndimage.label`); components a gutter
+apart that line up are re-joined, so a grid of cards or a chart beside its
+key comes back as one figure rather than four.
+
+Three fences keep it from making things worse, because a box confidently
+snapped onto the wrong thing is worse than a loose one the user can see:
+
+- **Text blocks are excluded from the mask up front.** They are the one
+  trustworthy boundary on the page, and without them a figure happily
+  swallows the column beside it.
+- **Acceptance is explicit.** The result must be either solidly inked
+  (`_SNAP_MIN_FILL`) or framed on nearly all of its rows *and* columns — the
+  second test is what admits a UI screenshot, which is white inside and inked
+  only at its edge — and it may not shrink below `_SNAP_MIN_SEED_AREA` of the
+  engine's box. Anything else keeps the engine's box untouched.
+- **Two figures pulled onto one graphic revert.** Either the page is one
+  picture the engine read as two or the reverse; neither reading can be
+  trusted over the other.
+
+Result on that issue: 78 of 138 boxes corrected, none made worse on visual
+inspection of the sample, the rest left alone (pale charts on white, mostly,
+which the ink test correctly declines). The pass is idempotent — an
+already-correct box seeds the same graphic again — which is what allows it to
+run outside ingestion: `refresh_visual_regions()` re-snaps on every
+re-detection, and `POST /api/magazine/{id}/figures/refine` re-snaps a whole
+issue as a background job, so a document ingested before this pass existed is
+corrected without paying for its transcription a second time. It replaced
+`refine_picture_bounds()`, which only recognised *coloured* panels and had no
+text fence.
+
+The snap runs before `crop_visual_regions()` in both OCR paths: the figure
+files shown during playback are cropped from these boxes, so a box corrected
+late would leave a mis-framed image on disk.
+
 The Active OCR Engine panel (Settings → Engines → OCR & Document Parsing)
 mirrors the Active Language Model panel exactly — same endpoint/model
 selects, same `/api/conversation/llm-models` fetch (it just proxies to

core/magazine.py

@@ -1143,6 +1143,201 @@ def build_vision_blocks(regions: list[dict[str, Any]], page_no: int,
     return out
 
 
+# ── Figure bounds from the page render ───────────────────────────────────────
+#
+# A vision OCR engine grounds a figure far less precisely than it transcribes
+# one: on c't KI Wissen 2026 the label and the read-out contents were right
+# while the four numbers landed on the neighbouring text column at two thirds
+# of the real size (page 23), or covered only the top third of a full-width
+# illustration (page 20).  The rendered page itself carries the answer -- a
+# photo, a chart or a UI screenshot is a densely inked rectangle fenced off
+# from the type area by white gutters -- so the model's box is used only to
+# say WHICH graphic is meant and the render says where that graphic ends.
+#
+# Deliberately conservative: every rule below can decline, and declining
+# leaves the model's box untouched.  A wrong box the user can see and redraw
+# is a nuisance; a box confidently snapped onto the wrong thing is worse.
+
+_SNAP_CELL_PX = 10        # grid resolution -- ~2 mm at the 110 dpi page render
+_SNAP_SOLID = 0.55        # inked fraction marking a cell as graphic, not type
+_SNAP_GAP_CELLS = 3       # white gutter a tiled graphic is allowed to span
+_SNAP_MIN_FILL = 0.45     # an accepted box is this solidly inked ...
+_SNAP_MIN_FRAME = 0.9     # ... or framed on this share of its rows AND columns
+_SNAP_MIN_SEED_AREA = 0.30   # ... and never shrinks below this much of the seed
+_SNAP_MIN_OVERLAP = 0.02  # a graphic must touch the seed by at least this much
+
+
+def page_ink_grid(png_path: Path, cell: int = _SNAP_CELL_PX):
+    """Coarse map of the page render: inked fraction of each `cell`-px square.
+
+    "Inked" means visibly darker than the paper or visibly coloured, so a
+    grey scan background does not count as ink.  Body type covers roughly a
+    quarter of the cells it runs through; a photo or a filled panel covers
+    nearly all of them, which is the separation snap_visual_bboxes() relies
+    on.  Returns None when the render cannot be read.
+    """
+    try:
+        import numpy as np
+        from PIL import Image
+        with Image.open(png_path) as im:
+            rgb = np.asarray(im.convert("RGB"), dtype=np.float32)
+    except Exception:
+        return None
+    lightness = rgb.max(axis=2)
+    saturation = lightness - rgb.min(axis=2)
+    paper = float(np.percentile(lightness, 90)) or 255.0
+    ink = (lightness < paper * 0.88) | (saturation > 40)
+    rows, cols = (n // cell for n in ink.shape)
+    if rows < 4 or cols < 4:
+        return None
+    return ink[:rows * cell, :cols * cell].reshape(
+        rows, cell, cols, cell).mean(axis=(1, 3))
+
+
+def _snap_seed(coverage, seed: Sequence[int], blocked) -> list[int] | None:
+    """Grow the graphic the seed cell-box points at, or None to keep the seed.
+
+    Cells already claimed by a text block are excluded up front: text boxes
+    come back from the same engine far more accurately than figure boxes
+    (verified against the page render for every block of the issue above),
+    so they are the one trustworthy fence against a figure swallowing the
+    column beside it.
+    """
+    from scipy import ndimage
+
+    solid = (coverage >= _SNAP_SOLID) & ~blocked
+    labels, count = ndimage.label(solid)
+    if not count:
+        return None
+    slices = ndimage.find_objects(labels)
+    boxes = [[s[1].start, s[0].start, s[1].stop - 1, s[0].stop - 1]
+             for s in slices]
+    sizes = ndimage.sum_labels(solid, labels, range(1, count + 1))
+
+    sx0, sy0, sx1, sy1 = seed
+    seed_area = max(1, (sx1 - sx0 + 1) * (sy1 - sy0 + 1))
+    overlaps = [_cell_overlap(b, seed) for b in boxes]
+    best = max(range(count), key=overlaps.__getitem__)
+    if overlaps[best] < _SNAP_MIN_OVERLAP * seed_area:
+        return None
+
+    # A figure split by white gutters (a grid of cards, a chart beside its
+    # key) arrives as several components; re-join the ones that line up with
+    # what we have so far, then re-test -- joining one may bring the next
+    # within reach.
+    box = list(boxes[best])
+    joined = {best}
+    growing = True
+    while growing:
+        growing = False
+        for i, other in enumerate(boxes):
+            if i in joined or sizes[i] < 4 or not _cells_adjacent(box, other):
+                continue
+            box = [min(box[0], other[0]), min(box[1], other[1]),
+                   max(box[2], other[2]), max(box[3], other[3])]
+            joined.add(i)
+            growing = True
+
+    area = (box[2] - box[0] + 1) * (box[3] - box[1] + 1)
+    if area < _SNAP_MIN_SEED_AREA * seed_area:
+        return None
+    inside = coverage[box[1]:box[3] + 1, box[0]:box[2] + 1]
+    if inside.mean() >= _SNAP_MIN_FILL:
+        return box
+    # A screenshot or a boxed insert is mostly white inside but inked right
+    # around its edge, so a solid frame counts as a graphic too.
+    edge = solid[box[1]:box[3] + 1, box[0]:box[2] + 1]
+    if (edge.any(axis=1).mean() >= _SNAP_MIN_FRAME
+            and edge.any(axis=0).mean() >= _SNAP_MIN_FRAME):
+        return box
+    return None
+
+
+def _cell_overlap(a: Sequence[int], b: Sequence[int]) -> int:
+    """Area, in grid cells, shared by two inclusive [x0, y0, x1, y1] boxes."""
+    return (max(0, min(a[2], b[2]) - max(a[0], b[0]) + 1)
+            * max(0, min(a[3], b[3]) - max(a[1], b[1]) + 1))
+
+
+def _cells_adjacent(box: Sequence[int], other: Sequence[int]) -> bool:
+    """True when `other` is a gutter away from `box` and lines up with it."""
+    gap = _SNAP_GAP_CELLS
+    if not (other[0] <= box[2] + gap and other[2] >= box[0] - gap
+            and other[1] <= box[3] + gap and other[3] >= box[1] - gap):
+        return False
+    # Side by side or stacked, not merely diagonal: a tile of the same figure
+    # shares most of a row or most of a column with it.
+    shared_rows = max(0, min(other[3], box[3]) - max(other[1], box[1]) + 1)
+    shared_cols = max(0, min(other[2], box[2]) - max(other[0], box[0]) + 1)
+    return (shared_rows >= 0.5 * min(other[3] - other[1] + 1, box[3] - box[1] + 1)
+            or shared_cols >= 0.5 * min(other[2] - other[0] + 1, box[2] - box[0] + 1))
+
+
+def snap_visual_bboxes(png_path: Path, blocks: Sequence[Block],
+                       page_meta: dict[str, Any]) -> int:
+    """Correct every visual block's bbox against the page render, in place.
+
+    Returns how many boxes moved.  Silent no-op when the render is missing or
+    the imaging stack is unavailable -- the layout stays exactly as the OCR
+    engine reported it, which is what happened before this pass existed.
+    """
+    visuals = [b for b in blocks if b.type in VISUAL_TYPES]
+    if not visuals:
+        return 0
+    try:
+        import numpy as np
+        import scipy.ndimage  # noqa: F401  -- required by _snap_seed
+    except ImportError:
+        return 0
+    coverage = page_ink_grid(png_path)
+    if coverage is None:
+        return 0
+    rows, cols = coverage.shape
+    page_w = float(page_meta.get("width") or 0) or 1.0
+    page_h = float(page_meta.get("height") or 0) or 1.0
+    cells_x, cells_y = cols / page_w, rows / page_h
+
+    def to_cells(bbox: Sequence[float]) -> list[int]:
+        x, y, w, h = bbox
+        return [min(cols - 1, max(0, int(x * cells_x))),
+                min(rows - 1, max(0, int(y * cells_y))),
+                min(cols - 1, max(0, int((x + w) * cells_x))),
+                min(rows - 1, max(0, int((y + h) * cells_y)))]
+
+    blocked = np.zeros(coverage.shape, dtype=bool)
+    for b in blocks:
+        if b.type in VISUAL_TYPES or not b.text.strip():
+            continue
+        x0, y0, x1, y1 = to_cells(b.bbox)
+        blocked[y0:y1 + 1, x0:x1 + 1] = True
+
+    snapped = [_snap_seed(coverage, to_cells(b.bbox), blocked) for b in visuals]
+    # Two figures pulled onto the same graphic means the page was read as
+    # one picture where the engine saw two (or the reverse); neither reading
+    # can be trusted over the other, so both keep their original box.
+    for i, a in enumerate(snapped):
+        for j in range(i + 1, len(snapped)):
+            c = snapped[j]
+            if not a or not c:
+                continue
+            smaller = min((a[2] - a[0] + 1) * (a[3] - a[1] + 1),
+                          (c[2] - c[0] + 1) * (c[3] - c[1] + 1))
+            if _cell_overlap(a, c) > 0.5 * smaller:
+                snapped[i] = snapped[j] = None
+                a = None
+
+    moved = 0
+    for block, box in zip(visuals, snapped):
+        if box is None:
+            continue
+        bbox = [box[0] / cells_x, box[1] / cells_y,
+                (box[2] + 1 - box[0]) / cells_x, (box[3] + 1 - box[1]) / cells_y]
+        if any(abs(a - b) > 1.0 for a, b in zip(bbox, block.bbox)):
+            moved += 1
+        block.bbox = bbox
+    return moved
+
+
 def ocr_page(png_path: Path, lang: str = "deu+eng",
              min_confidence: float = 40.0) -> list[dict[str, Any]]:
     """OCR one rendered page, returning paragraph boxes from tesseract's TSV.
@@ -1355,8 +1550,13 @@ def extract_pages(pdf_path: Path, out_dir: Path, render_dpi: int = 110,
                 # conflicting reading on top of it.
                 vision_page_count += 1
                 blocks = vision_blocks
+                page_meta = {"width": rect.width, "height": rect.height}
+                # Before cropping: the crop is taken from the block's bbox,
+                # so a box the engine placed loosely would be saved as a
+                # mis-framed figure file as well as a mis-drawn review box.
+                snap_visual_bboxes(png, blocks, page_meta)
                 detected_images = crop_visual_regions(out_dir, page_no,
-                    {"width": rect.width, "height": rect.height}, blocks)
+                    page_meta, blocks)
                 if detected_images:
                     page_images = detected_images
             elif needs_ocr:
@@ -1604,6 +1804,7 @@ def reocr_page(doc_dir: Path, layout: dict[str, Any], page_no: int,
 
     scale_x = page_meta["width"] / max(1.0, _png_width(png, fitz))
     new_blocks = build_vision_blocks(regions, page_no, scale_x, id_prefix="r")
+    snap_visual_bboxes(png, new_blocks, page_meta)
 
     all_blocks = [Block(**b) for b in layout.get("blocks", [])]
     kept = [b for b in all_blocks if b.page != page_no]
@@ -1734,54 +1935,23 @@ def _crop_page_png(doc_dir: Path, page_no: int, page_meta: dict[str, Any],
         im.crop(box_px).save(dest)
 
 
-def refine_picture_bounds(png: Path, page_meta: dict, bbox: Sequence[float]) -> list[float]:
-    """Snap an approximate picture box to a large, solid coloured panel.
-
-    Text and small decorations do not qualify. Grayscale or fragmented
-    artwork keeps the model's box instead of guessing a new boundary.
-    """
-    import numpy as np
-    from PIL import Image
-    from scipy import ndimage
-    with Image.open(png) as image:
-        rgb = np.asarray(image.convert("RGB"), dtype=np.int16)
-    height, width = rgb.shape[:2]
-    mask = (rgb.max(axis=2) - rgb.min(axis=2) > 28) & (rgb.min(axis=2) < 225)
-    mask = ndimage.binary_closing(mask, iterations=3)
-    labels, _ = ndimage.label(mask)
-    candidates = []
-    x, y, w, h = bbox
-    for i, slices in enumerate(ndimage.find_objects(labels), 1):
-        if slices is None:
-            continue
-        sy, sx = slices
-        area = (sy.stop - sy.start) * (sx.stop - sx.start)
-        if area < width * height * .025 or area > width * height * .85:
-            continue
-        if np.count_nonzero(labels[slices] == i) / area < .7:
-            continue
-        box = [sx.start / width * page_meta["width"], sy.start / height * page_meta["height"],
-               (sx.stop - sx.start) / width * page_meta["width"], (sy.stop - sy.start) / height * page_meta["height"]]
-        overlap = max(0, min(x+w, box[0]+box[2])-max(x, box[0])) * max(0, min(y+h, box[1]+box[3])-max(y, box[1]))
-        score = overlap / max(1, min(w*h, box[2]*box[3]))
-        if score > .6:
-            candidates.append((score, box))
-    return max(candidates, key=lambda c: c[0])[1] if candidates else list(bbox)
-
-
 def refresh_visual_regions(doc_dir: Path, layout: dict, page: int | None = None) -> None:
-    """Regenerate semantic crops and avoid treating a scanned page as a figure."""
+    """Re-snap figure bounds, regenerate the crops taken from them, and avoid
+    treating a whole scanned page as one figure."""
     for meta in layout.get("pages", []):
         page_no = meta["page"]
         if page is not None and page_no != page:
             continue
         blocks = [Block(**b) for b in layout.get("blocks", []) if b["page"] == page_no]
-        for block in blocks:
+        # Re-snapping here, not only at OCR time, is what lets an issue
-            if block.type == BLOCK_FIGURE and block.source == "llm_layout":
+        # ingested before this pass existed pick up corrected figure boxes
-                block.bbox = refine_picture_bounds(doc_dir / "pages" / f"{page_no:04d}.png", meta, block.bbox)
+        # without paying for a second OCR run of the whole document. It is
+        # idempotent: an already-correct box seeds the same graphic again.
+        if snap_visual_bboxes(doc_dir / "pages" / f"{page_no:04d}.png", blocks, meta):
+            snapped = {b.id: b.bbox for b in blocks if b.type in VISUAL_TYPES}
             for saved in layout["blocks"]:
-                    if saved["id"] == block.id:
+                if saved["id"] in snapped:
-                        saved["bbox"] = block.bbox
+                    saved["bbox"] = snapped[saved["id"]]
         crops = crop_visual_regions(doc_dir, page_no, meta, blocks)
         kept = []
         for img in layout.get("images", []):

routes/magazine.py

@@ -646,6 +646,53 @@ async def magazine_reocr_page(doc_id: str, page: int, request: Request):
     return {"job_id": job_id}
 
 
+def _refine_figures_sync(doc_id: str, job_id: str) -> None:
+    """Re-snap every figure box in a document against its page renders.
+
+    Restart-safe like the other layout jobs: a 200-page issue means 200 page
+    renders read and re-cropped, far too long for a request/response cycle.
+    Cheap compared with re-ingesting, though -- no OCR engine is called, so
+    an issue read before the snapping pass existed can be corrected without
+    paying for the transcription again, and without disturbing the text,
+    the article grouping, or any approval the user has already given.
+    """
+    d = _LIB / doc_id
+    try:
+        layout = mag.load_layout(d)
+        if not layout:
+            raise FileNotFoundError("No layout yet -- run ingestion first")
+        pages = layout.get("pages", [])
+        before = {b["id"]: list(b["bbox"]) for b in layout.get("blocks", [])}
+        for i, meta in enumerate(pages, 1):
+            background_job_update(job_id, stage="figures",
+                                  detail=f"Checking figure boundaries on page {i} of {len(pages)}")
+            mag.refresh_visual_regions(d, layout, meta["page"])
+        moved = sum(1 for b in layout.get("blocks", [])
+                    if before.get(b["id"]) and b["bbox"] != before[b["id"]])
+        mag.save_layout(d, layout)
+        background_job_update(
+            job_id, stage="done", done=True, result={"moved": moved},
+            detail=f"{moved} figure boundar{'y' if moved == 1 else 'ies'} corrected"
+                   if moved else "Every figure boundary already matched its page")
+    except Exception as exc:                       # noqa: BLE001 - surfaced to UI
+        background_job_update(job_id, stage="failed", done=True, error=str(exc)[:1000])
+
+
+@router.post("/api/magazine/{doc_id}/figures/refine")
+async def magazine_refine_figures(doc_id: str):
+    """Re-check every figure boundary in this issue against the page renders.
+
+    For an issue whose figure boxes came back loose from the OCR engine: the
+    boxes are re-derived from the rendered pages and the figure crops taken
+    from them are rebuilt. Text, articles and approvals are untouched.
+    """
+    _require_magazine(doc_id)
+    job_id = uuid.uuid4().hex
+    background_job_create(job_id, "magazine-refine-figures", fingerprint=doc_id)
+    asyncio.create_task(asyncio.to_thread(_refine_figures_sync, doc_id, job_id))
+    return {"job_id": job_id}
+
+
 def _add_region_sync(doc_id: str, job_id: str, page: int, bbox: list[float], kind: str,
                      text: str, vision_ocr_url: str | None, vision_ocr_model: str | None,
                      vision_ocr_fallback_url: str | None, vision_ocr_fallback_model: str | None) -> None:

static/js/magazine.js

@@ -740,9 +740,12 @@ function renderReview() {
     '<span class="mdi mdi-advertisements-outline" aria-hidden="true"></span> Advertisement</button>' +
     '<button type="button" class="btn-sm" onclick="magReviewClearCurrentPage()"><span class="mdi mdi-checkbox-multiple-blank-outline" aria-hidden="true"></span> Clear all fields</button></div>' +
     '<span class="mag-review-pagenav-spacer"></span>' +
+    '<button type="button" class="btn-sm" id="mag-refine-figures-btn" onclick="magRefineFigures()" ' +
+      'title="Re-check every figure boundary in this issue against the scanned pages">' +
+      '<span class="mdi mdi-image-auto-adjust" aria-hidden="true"></span> Fix figure boundaries</button>' +
     '<button type="button" class="btn-sm" id="mag-reocr-btn" onclick="magReocrPage()">' +
       '<span class="mdi mdi-text-recognition"></span> OCR again (this page)</button>' +
-    '<span id="mag-reocr-status" class="muted"></span>' +
+    '<span id="mag-reocr-status" class="muted" role="status"></span>' +
     '</div>';
 
   html += '<div class="mag-review-fields">';
@@ -1173,6 +1176,45 @@ async function magReocrPage() {
   renderReview();
 }
 
+// Re-derives every figure box in the issue from the page renders. Separate
+// from "OCR again" because no OCR engine is involved: it only corrects where
+// the figures are, never what the page says, so it is safe to run on an issue
+// that is half reviewed already.
+async function magRefineFigures() {
+  const btn = $('mag-refine-figures-btn');
+  const statusEl = $('mag-reocr-status');
+  if (btn) btn.disabled = true;
+  if (statusEl) statusEl.textContent = 'Checking figure boundaries…';
+  let jobId;
+  try {
+    const r = await api('/api/magazine/' + state.doc.id + '/figures/refine', { method: 'POST' });
+    jobId = r.job_id;
+  } catch (e) {
+    if (btn) btn.disabled = false;
+    if (statusEl) statusEl.textContent = '';
+    return say('Could not start the check: ' + e.message, 'err');
+  }
+  let failed = false;
+  for (;;) {
+    let job;
+    try { job = await api('/api/jobs/' + jobId); } catch (_) { break; }
+    if (job.done) {
+      if (job.error) { failed = true; say('Figure check failed: ' + job.error, 'err'); }
+      else say(job.detail || 'Figure boundaries checked', 'ok');
+      break;
+    }
+    if (statusEl && job.detail) statusEl.textContent = job.detail;
+    await new Promise((r) => setTimeout(r, 2000));
+  }
+  if (btn) btn.disabled = false;
+  if (statusEl) statusEl.textContent = '';
+  if (failed) return;
+  state.layout = null;
+  await ensureLayout();
+  _magReviewChecked = null;
+  renderReview();
+}
+
 async function magApprove() {
   const a = state.article;
   if (!a) return;
@@ -1805,6 +1847,7 @@ root.magReviewFocusBlock = magReviewFocusBlock;
 root.magReviewHighlightBox = magReviewHighlightBox;
 root.magReviewToggleBox = magReviewToggleBox;
 root.magReocrPage = magReocrPage;
+root.magRefineFigures = magRefineFigures;
 root.magReviewToggleImage = magReviewToggleImage;
 root.magReviewToggleDraw = magReviewToggleDraw;
 root.magReviewCancelDrawPanel = magReviewCancelDrawPanel;

tests/test_magazine.py

@@ -19,6 +19,9 @@ os.environ.setdefault("CONFIG_DIR", tempfile.mkdtemp(prefix="ttvc-magazine-test-
 from core.magazine import (
     ocr_page,
     cluster_running_furniture,
+    snap_visual_bboxes,
+    page_ink_grid,
+    BLOCK_FIGURE,
     looks_like_real_title,
     Article,
     BLOCK_AD,
@@ -1435,5 +1438,102 @@ class ManualRegionTests(unittest.TestCase):
         self.assertEqual(len(names), 2)
 
 
+class SnapVisualBboxesTests(unittest.TestCase):
+    """A vision OCR engine grounds a figure loosely; the render is the truth.
+
+    The synthetic page mirrors the real failure seen on c't KI Wissen 2026:
+    a solid graphic on the right, a text column on the left, and a reported
+    figure box that straddles both at a fraction of the real size.
+    """
+
+    PAGE_W, PAGE_H = 582.0, 678.0          # PDF points
+    PX_W, PX_H = 890, 1037                 # rendered pixels
+    FIGURE_PX = (400, 100, 850, 400)       # x0, y0, x1, y1
+    TEXT_PX = (60, 100, 350, 500)
+
+    def setUp(self):
+        from PIL import Image, ImageDraw
+        self.dir = pathlib.Path(tempfile.mkdtemp())
+        (self.dir / "pages").mkdir()
+        page = Image.new("RGB", (self.PX_W, self.PX_H), (255, 255, 255))
+        draw = ImageDraw.Draw(page)
+        draw.rectangle(self.FIGURE_PX, fill=(40, 90, 200))
+        x0, y0, x1, y1 = self.TEXT_PX
+        for y in range(y0, y1, 12):        # type covers ~a third of its cells
+            draw.rectangle((x0, y, x1, y + 3), fill=(20, 20, 20))
+        self.png = self.dir / "pages" / "0001.png"
+        page.save(self.png)
+        self.meta = {"page": 1, "width": self.PAGE_W, "height": self.PAGE_H}
+
+    def tearDown(self):
+        shutil.rmtree(self.dir, ignore_errors=True)
+
+    def _pt(self, px_box):
+        """Pixel [x0, y0, x1, y1] as a PDF-point [x, y, w, h] bbox."""
+        sx, sy = self.PAGE_W / self.PX_W, self.PAGE_H / self.PX_H
+        x0, y0, x1, y1 = px_box
+        return [x0 * sx, y0 * sy, (x1 - x0) * sx, (y1 - y0) * sy]
+
+    def _text_block(self):
+        return Block(id="t1", page=1, bbox=self._pt(self.TEXT_PX),
+                     type=BLOCK_BODY, text="body copy " * 40)
+
+    def _figure_block(self, px_box):
+        return Block(id="f1", page=1, bbox=self._pt(px_box), type=BLOCK_FIGURE,
+                     text="figure", source="vision_label")
+
+    def test_grid_separates_graphic_from_type(self):
+        grid = page_ink_grid(self.png)
+        self.assertIsNotNone(grid)
+        rows, cols = grid.shape
+        fig = grid[int(200 / self.PX_H * rows), int(600 / self.PX_W * cols)]
+        txt = grid[int(200 / self.PX_H * rows), int(200 / self.PX_W * cols)]
+        self.assertGreater(fig, 0.9)
+        self.assertLess(txt, 0.5)
+
+    def test_offset_box_snaps_onto_the_graphic(self):
+        block = self._figure_block((200, 60, 600, 250))
+        moved = snap_visual_bboxes(self.png, [block, self._text_block()], self.meta)
+        self.assertEqual(moved, 1)
+        want = self._pt(self.FIGURE_PX)
+        for got, expected in zip(block.bbox, want):
+            self.assertAlmostEqual(got, expected, delta=12.0)
+
+    def test_snap_does_not_swallow_the_text_column(self):
+        block = self._figure_block((200, 60, 600, 250))
+        snap_visual_bboxes(self.png, [block, self._text_block()], self.meta)
+        self.assertGreater(block.bbox[0], self._pt(self.TEXT_PX)[0] + 20)
+
+    def test_snapping_is_idempotent(self):
+        block = self._figure_block((200, 60, 600, 250))
+        blocks = [block, self._text_block()]
+        snap_visual_bboxes(self.png, blocks, self.meta)
+        settled = list(block.bbox)
+        self.assertEqual(snap_visual_bboxes(self.png, blocks, self.meta), 0)
+        self.assertEqual(block.bbox, settled)
+
+    def test_box_over_type_only_is_left_alone(self):
+        block = self._figure_block((60, 100, 350, 500))
+        before = list(block.bbox)
+        self.assertEqual(
+            snap_visual_bboxes(self.png, [block, self._text_block()], self.meta), 0)
+        self.assertEqual(block.bbox, before)
+
+    def test_two_figures_on_one_graphic_keep_their_boxes(self):
+        left = self._figure_block((410, 110, 600, 380))
+        right = Block(id="f2", page=1, bbox=self._pt((610, 110, 840, 380)),
+                      type=BLOCK_FIGURE, text="figure", source="vision_label")
+        before = [list(left.bbox), list(right.bbox)]
+        self.assertEqual(snap_visual_bboxes(self.png, [left, right], self.meta), 0)
+        self.assertEqual([left.bbox, right.bbox], before)
+
+    def test_missing_render_is_a_no_op(self):
+        block = self._figure_block((200, 60, 600, 250))
+        before = list(block.bbox)
+        self.assertEqual(
+            snap_visual_bboxes(self.dir / "pages" / "0099.png", [block], self.meta), 0)
+        self.assertEqual(block.bbox, before)
+
+
 if __name__ == "__main__":
     unittest.main()

tests/test_magazine_routes.py

@@ -160,3 +160,61 @@ class AdvertisementReviewTests(unittest.TestCase):
         self.assertEqual(len(layout['blocks']), 1)
         self.assertEqual(layout['blocks'][0]['type'], 'ad')
         self.assertEqual(layout['blocks'][0]['bbox'], [0, 0, 600, 800])
+
+
+class RefineFiguresJobTests(unittest.TestCase):
+    """Document-wide figure re-snap: reports what moved, keeps text untouched."""
+
+    def _layout(self):
+        return {
+            "pages": [{"page": 1, "width": 600, "height": 800},
+                      {"page": 2, "width": 600, "height": 800}],
+            "blocks": [
+                {"id": "fig", "page": 1, "type": "figure", "bbox": [10, 10, 50, 50], "text": ""},
+                {"id": "copy", "page": 2, "type": "body", "bbox": [0, 0, 100, 100], "text": "Body"},
+            ],
+        }
+
+    def test_moved_boxes_are_counted_and_saved(self):
+        from unittest.mock import patch
+        from routes.magazine import _refine_figures_sync
+        layout = self._layout()
+
+        def refresh(_dir, doc, page=None):
+            if page == 1:
+                doc["blocks"][0]["bbox"] = [20, 20, 90, 90]
+
+        with patch('routes.magazine.mag.load_layout', return_value=layout), \
+             patch('routes.magazine.mag.refresh_visual_regions', side_effect=refresh), \
+             patch('routes.magazine.mag.save_layout') as save, \
+             patch('routes.magazine.background_job_update') as update:
+            _refine_figures_sync('a' * 32, 'job')
+        save.assert_called_once()
+        done = [c.kwargs for c in update.call_args_list if c.kwargs.get('done')]
+        self.assertEqual(done[-1]['stage'], 'done')
+        self.assertEqual(done[-1]['result'], {'moved': 1})
+        self.assertEqual(layout['blocks'][1]['bbox'], [0, 0, 100, 100])
+
+    def test_nothing_to_correct_still_finishes_cleanly(self):
+        from unittest.mock import patch
+        from routes.magazine import _refine_figures_sync
+        with patch('routes.magazine.mag.load_layout', return_value=self._layout()), \
+             patch('routes.magazine.mag.refresh_visual_regions'), \
+             patch('routes.magazine.mag.save_layout'), \
+             patch('routes.magazine.background_job_update') as update:
+            _refine_figures_sync('a' * 32, 'job')
+        done = [c.kwargs for c in update.call_args_list if c.kwargs.get('done')]
+        self.assertEqual(done[-1]['result'], {'moved': 0})
+        self.assertNotIn('error', done[-1])
+
+    def test_a_document_with_no_layout_reports_failure(self):
+        from unittest.mock import patch
+        from routes.magazine import _refine_figures_sync
+        with patch('routes.magazine.mag.load_layout', return_value=None), \
+             patch('routes.magazine.mag.save_layout') as save, \
+             patch('routes.magazine.background_job_update') as update:
+            _refine_figures_sync('a' * 32, 'job')
+        save.assert_not_called()
+        done = [c.kwargs for c in update.call_args_list if c.kwargs.get('done')]
+        self.assertEqual(done[-1]['stage'], 'failed')
+        self.assertIn('No layout', done[-1]['error'])