python-hwpx 4.0.0__py3-none-any.whl → 4.1.1__py3-none-any.whl

hwpx/form_fit/wordbox.py

@@ -681,6 +681,14 @@ def extract_cell_clips(pdf_path: str, *, page: int | None = None) -> list[Rect]:
     These are the overflow clips — same coordinate space as the glyph boxes — so
     no HWPX→PDF transform is needed. ``find_tables`` can raise on odd pages; such
     a page contributes no clips rather than crashing.
+
+    Borders-only (``strategy="lines_strict"``): a cell rectangle is defined by its
+    *drawn borders*, not by the text inside it. The ``find_tables`` default snaps
+    edges to glyph positions, so a filled value can appear to *escape* a clip that
+    was mis-sized around the very text it holds — a false overflow. S-094 P3
+    measured this: the only 3 corpus pairs that flagged ``newOverflow`` (the nts
+    forms) all had the fill value escaping a text-snapped clip under the default,
+    and 0 escapes under ``lines_strict``; there were no real escapes to mask.
     """
 
     if not fitz_available():
@@ -695,7 +703,7 @@ def extract_cell_clips(pdf_path: str, *, page: int | None = None) -> list[Rect]:
     with doc:
         for pno in _page_indices(doc, page):
             try:
-                tables = doc[pno].find_tables().tables
+                tables = doc[pno].find_tables(strategy="lines_strict").tables
             except Exception:
                 continue  # layout analysis failed on this page -> no clips, no crash
             for ti, table in enumerate(tables):
@@ -839,8 +847,21 @@ class LayoutSignature:
     lesson that absolute glyph-overlap is unusable on real forms). Page count is
     the dominant, high-confidence signal — a fill that spills onto a new page is
     the canonical layout collapse. Per-table ``(rows, cols)`` is the finer
-    row/column signal; ``find_tables`` is content-sensitive, so it is judged as a
-    multiset *delta* (blank vs filled), not an absolute count.
+    row/column signal, extracted from *drawn table borders only*
+    (``find_tables(strategy="lines_strict")``): a structural fingerprint must be
+    text-independent, so it is judged as a multiset *delta* (blank vs filled),
+    not an absolute count.
+
+    Why ``lines_strict`` and not the ``find_tables`` default (``"lines"``): the
+    default snaps table edges to *text* positions, so injecting a value into a
+    previously-empty cell makes it hallucinate a phantom ``(2,2)``/``(1,2)`` table
+    or wobble a real one by ±1 row/col even when the drawn grid and page count are
+    unchanged. S-094 P3 measured this on the frozen corpus (all 6 "table-shape"
+    differential failures were page-stable phantoms; under ``lines_strict`` blank
+    and filled agree, 0 regressions across 31 pairs; ``sen-24`` has byte-identical
+    drawn rectangles yet the default strategy still invented a ``(1,2)`` table from
+    the 3 added glyphs). Borders-only detection measures the grid the fill did not
+    touch.
     """
 
     page_count: int
@@ -879,7 +900,12 @@ def extract_layout_signature(pdf_path: str) -> LayoutSignature:
         for page in doc:
             sizes.append((float(page.rect.width), float(page.rect.height)))
             try:
-                tables = page.find_tables().tables
+                # Borders-only ("lines_strict"): a *structural* fingerprint must
+                # not react to cell text. The default strategy snaps to glyph
+                # positions and invents phantom tables from a fill's added text
+                # (S-094 P3). Bordered form grids are detected identically either
+                # way; the text-sensitivity is the only difference we drop.
+                tables = page.find_tables(strategy="lines_strict").tables
             except Exception:
                 continue  # layout analysis failed on this page -> no tables, no crash
             for table in tables:

{python_hwpx-4.0.0.dist-info → python_hwpx-4.1.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-hwpx
-Version: 4.0.0
+Version: 4.1.1
 Summary: 한글 없이 HWPX 문서를 열고, 편집하고, 생성하고, 검증하는 Python 자동화 라이브러리
 Author: python-hwpx Maintainers
 License-Expression: Apache-2.0

{python_hwpx-4.0.0.dist-info → python_hwpx-4.1.1.dist-info}/RECORD

@@ -108,7 +108,7 @@ hwpx/form_fit/measure.py,sha256=gnSfqccm7_3oTKv3HH6Zs77r2bToxzmfvUWN2zshDE8,2232
 hwpx/form_fit/policy.py,sha256=iXAatjS7ZycOdKgMCL4Q_kYg0qnogFPdRNlr_ecOvjc,3106
 hwpx/form_fit/report.py,sha256=h1NzQfbj1JYvZtKUi0DckJNWGnUHMZtQ3iCzbUxUjHA,3305
 hwpx/form_fit/seal.py,sha256=QsPzFagaEi7UKAXS-iFwiZs8rLG-_Fo9RixLc2OaqbU,17640
-hwpx/form_fit/wordbox.py,sha256=bStyzpcK4PF3pihRUSKiStwW_XwYHDeU0D3Or_Eh0nM,48301
+hwpx/form_fit/wordbox.py,sha256=n86BegyBZMybf6Qxol6hWCG1r0EWiaJRyZEno_YkpIE,50170
 hwpx/ingest/__init__.py,sha256=cjDQwMdaqbq2xiD32zmaFAyzZbENga_6MLbgo0sLGZE,636
 hwpx/ingest/base.py,sha256=95jXJ70Hkay70BgyryGOzI3UVTixUpR-6f71iF3eqn4,8134
 hwpx/ingest/hwpx_converter.py,sha256=HOUCFYO-HK4WRpWiN5sA_kZIvdv439mkZ6PMCFU3hLg,4269
@@ -209,10 +209,10 @@ hwpx/visual/page_qa.py,sha256=AuJCySfLIdXjPeekhAc3YnrIWOsN0sI35Pnz8BZ6Pro,7278
 hwpx/visual/qa_contracts.py,sha256=1cjoiRTAWgi7NgE8SPc6bKxdoUfKx9nyou1mgC47TxY,10506
 hwpx/visual/qa_metrics.py,sha256=I7RdybN-f1Fvcv2j-ceDPoek51nOkMui1Zrd7TEX_C8,9838
 hwpx/visual/report.py,sha256=2RhXN1KBYOZTim9FNpeUhaaDHR7oFxI6Z2DLUkDIiwE,1717
-python_hwpx-4.0.0.dist-info/licenses/LICENSE,sha256=_ubz4wv-BkkT3l3gu-QuH7JGeVjuRYGZoZK95eNsCHU,9688
-python_hwpx-4.0.0.dist-info/licenses/NOTICE,sha256=KJgtwIIzrXoA0j3PUhwp78ZtnucocEoB7jiuwoL4i7o,3060
-python_hwpx-4.0.0.dist-info/METADATA,sha256=bKvgcCFo1AJOMbkSGyeLDL6oP8z2blDqtP-JOElOgBo,9164
-python_hwpx-4.0.0.dist-info/WHEEL,sha256=K260EYznzXsJYBQGqmI8VTxEdiZYNvDZwW9cBh9-_MA,91
-python_hwpx-4.0.0.dist-info/entry_points.txt,sha256=PVUBTBQtBHiflQ9XBjfd004w-LrDibgZjwzxgu8Tx7w,480
-python_hwpx-4.0.0.dist-info/top_level.txt,sha256=R1iToqDh80Nf2oQhRjTN0rbN2X6kyDUizIocZjkhuxc,5
-python_hwpx-4.0.0.dist-info/RECORD,,
+python_hwpx-4.1.1.dist-info/licenses/LICENSE,sha256=_ubz4wv-BkkT3l3gu-QuH7JGeVjuRYGZoZK95eNsCHU,9688
+python_hwpx-4.1.1.dist-info/licenses/NOTICE,sha256=KJgtwIIzrXoA0j3PUhwp78ZtnucocEoB7jiuwoL4i7o,3060
+python_hwpx-4.1.1.dist-info/METADATA,sha256=5R5nvsB-mwhZxnwU0KQG3wXscXzEBa6TjdvA4iknZhg,9164
+python_hwpx-4.1.1.dist-info/WHEEL,sha256=K260EYznzXsJYBQGqmI8VTxEdiZYNvDZwW9cBh9-_MA,91
+python_hwpx-4.1.1.dist-info/entry_points.txt,sha256=PVUBTBQtBHiflQ9XBjfd004w-LrDibgZjwzxgu8Tx7w,480
+python_hwpx-4.1.1.dist-info/top_level.txt,sha256=R1iToqDh80Nf2oQhRjTN0rbN2X6kyDUizIocZjkhuxc,5
+python_hwpx-4.1.1.dist-info/RECORD,,