docvortex 0.4.14__tar.gz → 0.4.16__tar.gz

{docvortex-0.4.14/src/docvortex.egg-info → docvortex-0.4.16}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: docvortex
-Version: 0.4.14
+Version: 0.4.16
 Summary: A fast, multi-format document parsing and conversion engine
 License-Expression: MIT
 Project-URL: Repository, https://github.com/myhloli/DocVortex

{docvortex-0.4.14 → docvortex-0.4.16}/docs/sdk-0.4.md

@@ -1,6 +1,6 @@
 # DocVortex 0.4 公共 SDK 与宿主迁移
 
-MinerU 使用 `docvortex>=0.4.0,<0.5.0`。本轮调整 Python 模块边界，保持文档协议、解析路由、模型回退与渲染语义。
+本文记录 0.4 系列的公共模块边界。MinerU 的页面原语复用需要 `docvortex>=0.4.15,<1`，文档协议、解析路由、模型回退与渲染语义保持不变。
 
 ## 公开入口
 
@@ -30,6 +30,11 @@ MinerU 使用 `docvortex>=0.4.0,<0.5.0`。本轮调整 Python 模块边界，保
 
 `prepare_table_page` 仅在存在候选表格时调用；同页复用其物化数据。`recover_table_region` 返回最终 HTML、来源、置信度和诊断，`None` 表示没有可接受结果。结构恢复异常以 `PDFTableRecoveryError` 保留原始原因，交给宿主回退；HTML 物化异常继续传播，避免改变旧异常边界。页面几何可以复用，证据不持有 PDFium 裸句柄。
 
+从 0.4.15 起，`PDFPage.get_vector_geometry()` 一次遍历返回公开的 `PDFPageVectorGeometry`，其中
+`drawing_lines` 与 `path_infos` 是物化元组。`prepare_table_page` 和 `prepare_text_evidence` 都接受可选的
+`vector_geometry=`，可与 `geometry=` 一起复用。调用方按页、按处理窗口持有并释放这些数据；`PDFPage`
+本身不缓存。文字证据构造只读访问源字符，旋转和组行累加使用独立的局部对象。
+
 通用 `convert_bbox` 显式声明 `unit`、`pixel` 或 `point`；`page_size` 使用 point，`render_scale` 是每 point 的像素数。模型 bbox 的自动解释与三位小数规则保留在 MinerU。共享 PDFium 锁、进程池和资源关闭职责仍归 DocVortex。
 
 ## 升级与回退

{docvortex-0.4.14 → docvortex-0.4.16}/src/docvortex/analyzers/native/pdf/line_merging.py

@@ -7,6 +7,8 @@ from .layout_evidence import build_layout_evidence
 import re
 import statistics
 import unicodedata
+import heapq
+import math
 
 from ....schema import BBox
 from .geometry import (
@@ -35,6 +37,59 @@ def _caption_crosses_left_text(members: list[_LineItem]) -> bool:
     )
 
 
+def _same_baseline_candidate_pairs(
+    lines: list[_LineItem],
+    local_bboxes: list[BBox],
+    compatible_indices: dict[tuple[int, bool, str | None], list[int]],
+) -> list[list[int]] | None:
+    """仅筛除原规则不可能接受的纵向远距行对；异常或密集页面回退原遍历。"""
+    bounds: list[tuple[float, float]] = []
+    for line, box in zip(lines, local_bboxes, strict=True):
+        try:
+            height = _line_effective_height(line, box)
+            if not all(math.isfinite(value) for value in (*box, height)) or box[2] <= box[0] or box[3] <= box[1]:
+                return None
+            # 同行规则既接受框交叠，也接受底边差不超过 0.25 * max(height) 的行。
+            # 两行各自扩展底边容差形成超集，最终仍由原判定函数决定是否合并。
+            low, high = min(box[1], box[3] - 0.25 * height), box[3] + 0.25 * height
+            if line.angle == 0 and line.source_bbox is not None:
+                source = line.source_bbox
+                if not all(math.isfinite(value) for value in source) or source[2] <= source[0] or source[3] <= source[1]:
+                    return None
+                source_height = source[3] - source[1]
+                low = min(low, source[1], source[3] - 0.25 * source_height)
+                high = max(high, source[3] + 0.25 * source_height)
+            if not math.isfinite(low) or not math.isfinite(high):
+                return None
+        except (TypeError, ValueError, IndexError):
+            return None
+        bounds.append((low, high))
+
+    candidates: list[list[int]] = [[] for _ in lines]
+    pair_count = 0
+    pair_limit = max(4096, 32 * len(lines))
+    for indices in compatible_indices.values():
+        active: set[int] = set()
+        ends: list[tuple[float, int]] = []
+        for index in sorted(indices, key=lambda item: (bounds[item][0], item)):
+            low, high = bounds[index]
+            # 保留相等端点，避免漏掉原规则恰好落在容差边界上的行对。
+            while ends and ends[0][0] < low:
+                _, expired = heapq.heappop(ends)
+                active.remove(expired)
+            pair_count += len(active)
+            if pair_count > pair_limit:
+                return None
+            for previous in active:
+                left, right = (previous, index) if previous < index else (index, previous)
+                candidates[left].append(right)
+            active.add(index)
+            heapq.heappush(ends, (high, index))
+    for partners in candidates:
+        partners.sort()
+    return candidates
+
+
 def _merge_same_baseline_text_lines(
     lines: list[_LineItem],
     page_size: tuple[float, float],
@@ -67,8 +122,14 @@ def _merge_same_baseline_text_lines(
     compatible_indices: dict[tuple[int, bool, str | None], list[int]] = {}
     for index, line in enumerate(lines):
         compatible_indices.setdefault((line.angle, line.formula_candidate_only, line.semantic_type), []).append(index)
+    candidates = _same_baseline_candidate_pairs(lines, local_bboxes, compatible_indices)
     for left_index, left_line in enumerate(lines):
-        for right_index in compatible_indices[(left_line.angle, left_line.formula_candidate_only, left_line.semantic_type)]:
+        right_indices = (
+            candidates[left_index]
+            if candidates is not None
+            else compatible_indices[(left_line.angle, left_line.formula_candidate_only, left_line.semantic_type)]
+        )
+        for right_index in right_indices:
             if right_index <= left_index:
                 continue
             right_line = lines[right_index]
@@ -222,14 +283,14 @@ def _overlapping_inline_cluster_pair_is_connected(
         return False
     if first_line.semantic_type != second_line.semantic_type:
         return False
+    if _bbox_axis_overlap_ratio(first_bbox, second_bbox, axis="y") < 0.55:
+        return False
     if _connection_crosses_table(
         first_line.bbox,
         second_line.bbox,
         table_bboxes,
     ):
         return False
-    if _bbox_axis_overlap_ratio(first_bbox, second_bbox, axis="y") < 0.55:
-        return False
 
     first_guard_bbox = first_line.source_bbox or first_bbox
     second_guard_bbox = second_line.source_bbox or second_bbox

{docvortex-0.4.14 → docvortex-0.4.16}/src/docvortex/analyzers/pdf.py

@@ -3,11 +3,10 @@
 from __future__ import annotations
 
 from collections.abc import Sequence
-from copy import deepcopy
 from dataclasses import dataclass, replace
 from typing import Any
 
-from ..document.pdf import PDFPage, PDFPageTextGeometry
+from ..document.pdf import PDFPage, PDFPageTextGeometry, PDFPageVectorGeometry
 from ..schema import BBox
 from .native.pdf._script_geometry import ScriptRole, classify_char_script_roles
 from .native.pdf._table_recovery.contracts import NativeTableRectangle, NativeTableRule, PDFTableRecoveryError
@@ -57,6 +56,7 @@ def prepare_text_evidence(
     page: PDFPage,
     *,
     geometry: PDFPageTextGeometry | None = None,
+    vector_geometry: PDFPageVectorGeometry | None = None,
     supported_angles: Sequence[float] = (0.0,),
     table_regions: Sequence[BBox] = (),
     excluded_script_regions: Sequence[BBox] = (),
@@ -70,11 +70,12 @@ def prepare_text_evidence(
 
     geometry = geometry if geometry is not None else page.get_chars_with_geometry()
     page_size = tuple(float(value) for value in page.size)
-    chars = deepcopy(geometry.chars)
+    # 源字符只读；组行使用独立累加框，旋转等变换在消费处创建局部字符副本。
+    chars = geometry.chars
     lines = _build_native_line_items(
         get_lines_from_chars(chars), page_size, page_rotation=page.rotation, supported_angles=supported_angles
     )
-    drawing_lines = page.get_drawing_lines()
+    drawing_lines = vector_geometry.drawing_lines if vector_geometry is not None else page.get_drawing_lines()
     styles = detect_pdf_text_style_lines(lines, drawing_lines)
     links = detect_pdf_text_link_lines(lines, page.get_link_annotations())
     script_lines = merge_text_line_clusters(list(lines), page_size, list(table_regions))
@@ -124,16 +125,22 @@ def apply_text_evidence(
     )
 
 
-def prepare_table_page(page: PDFPage, *, geometry: PDFPageTextGeometry | None = None) -> PDFTablePage:
+def prepare_table_page(
+    page: PDFPage,
+    *,
+    geometry: PDFPageTextGeometry | None = None,
+    vector_geometry: PDFPageVectorGeometry | None = None,
+) -> PDFTablePage:
     """在调用方确认存在候选表格后物化页面原语，复用已有字符几何。"""
     from .native.pdf._table_recovery.engine import coerce_native_table_rectangles, coerce_native_table_rules
 
     geometry = geometry if geometry is not None else page.get_chars_with_geometry()
+    vector_geometry = vector_geometry if vector_geometry is not None else page.get_vector_geometry()
     return PDFTablePage(
         tuple(float(value) for value in page.size),
         geometry,
-        coerce_native_table_rules(page.get_drawing_lines()),
-        coerce_native_table_rectangles(page.get_path_infos()),
+        coerce_native_table_rules(vector_geometry.drawing_lines),
+        coerce_native_table_rectangles(vector_geometry.path_infos),
     )
 
 

{docvortex-0.4.14 → docvortex-0.4.16}/src/docvortex/document/pdf/__init__.py

@@ -1,6 +1,6 @@
 """PDF 文档访问、显式分类及共享文本契约。"""
 
-from ._document import PDFDocument, PDFPage, PDFPageTextGeometry, get_lines_from_chars
+from ._document import PDFDocument, PDFPage, PDFPageTextGeometry, PDFPageVectorGeometry, get_lines_from_chars
 from .pdfium import PdfiumFontError, PdfiumRuntimeInfo, initialize_pdfium_runtime
 from .text import Bbox, Char, Line, Span
 
@@ -8,6 +8,7 @@ __all__ = [
     "PDFDocument",
     "PDFPage",
     "PDFPageTextGeometry",
+    "PDFPageVectorGeometry",
     "get_lines_from_chars",
     "Bbox",
     "Char",

{docvortex-0.4.14 → docvortex-0.4.16}/src/docvortex/document/pdf/_document.py

@@ -104,6 +104,7 @@ from .native_contracts import (
 from .native_contracts import (
     PDFPageTextGeometry as PDFPageTextGeometry,
 )
+from .native_contracts import PDFPageVectorGeometry as PDFPageVectorGeometry
 from .native_contracts import (
     PDFPathInfo as PDFPathInfo,
 )
@@ -289,6 +290,10 @@ class PDFPage:
 
         return self.pdf_doc.get_page_path_infos(self._idx)
 
+    def get_vector_geometry(self) -> PDFPageVectorGeometry:
+        """一次读取当前页的绘图线和路径摘要，由调用方显式管理复用范围。"""
+        return self.pdf_doc.get_page_vector_geometry(self._idx)
+
     def get_link_annotations(self) -> list[PDFLinkAnnotation]:
         """读取当前页已规范到视觉页面坐标的外部 URI Link 注解。"""
 
@@ -599,6 +604,17 @@ class PDFDocument:
                 page_rotation = 0
             return _extract_page_drawing_lines(page, page_bbox, page_rotation)
 
+    def get_page_vector_geometry(self, page_idx: int) -> PDFPageVectorGeometry:
+        """在同一页面和共享锁内遍历一次 Path，返回独立物化的矢量几何。"""
+        with self._open_page(page_idx) as page:
+            page_bbox = _normalize_pdf_page_bbox(page.get_bbox())
+            try:
+                page_rotation = int(page.get_rotation()) % 360
+            except Exception:
+                page_rotation = 0
+            drawings, paths = _extract_page_paths_and_lines(page, page_bbox, page_rotation)
+            return PDFPageVectorGeometry(tuple(drawings), tuple(paths))
+
     def get_page_path_infos(self, page_idx: int) -> list[PDFPathInfo]:
         """提取页面及嵌套 Form 中的 Path 几何和绘制特征。"""
         with self._open_page(page_idx) as page:

{docvortex-0.4.14 → docvortex-0.4.16}/src/docvortex/document/pdf/native_contracts.py

@@ -125,6 +125,14 @@ class PDFImageInfo:
     fingerprint: str | None
 
 
+@dataclass(frozen=True, slots=True)
+class PDFPageVectorGeometry:
+    """保存一次 Path 遍历物化的线与路径摘要，不持有 PDFium 句柄。"""
+
+    drawing_lines: tuple[PDFDrawingLine, ...] = ()
+    path_infos: tuple[PDFPathInfo, ...] = ()
+
+
 @dataclass(frozen=True, slots=True)
 class _PDFPageSnapshot:
     """保存单次页面生命周期提取的纯 Python 数据，不持有 PDFium 子对象。"""

{docvortex-0.4.14 → docvortex-0.4.16}/src/docvortex/document/properties.py

@@ -29,7 +29,7 @@ def property_date(value: object, *, warnings: list[str] | None = None) -> str |
     if not text:
         return None
     if text.startswith("D:"):
-        match = re.fullmatch(r"D:(\d{4})(\d{2})?(\d{2})?(\d{2})?(\d{2})?(\d{2})?(Z|[+-]\d{2}'?\d{2}'?)?", text)
+        match = re.fullmatch(r"D:(\d{4})(\d{2})?(\d{2})?(\d{2})?(\d{2})?(\d{2})?(Z|[+-]\d{2}'?(?:\d{2}'?)?)?", text)
         if match is None:
             if warnings is not None:
                 warnings.append(f"Invalid declared date: {text[:120]}")
@@ -41,7 +41,7 @@ def property_date(value: object, *, warnings: list[str] | None = None) -> str |
                 text += separator + part
         if zone:
             zone = zone.replace("'", "")
-            text += zone if zone == "Z" else zone[:3] + ":" + zone[3:]
+            text += zone if zone == "Z" else zone[:3] + (":" + zone[3:] if zone[3:] else ":00")
     try:
         if re.fullmatch(r"\d{4}", text):
             date(int(text), 1, 1)

{docvortex-0.4.14 → docvortex-0.4.16}/src/docvortex/public_api.py

@@ -131,6 +131,7 @@ PUBLIC_API: dict[str, tuple[str, ...]] = {
         "PDFDocument",
         "PDFPage",
         "PDFPageTextGeometry",
+        "PDFPageVectorGeometry",
         "PdfiumFontError",
         "PdfiumRuntimeInfo",
         "Span",

{docvortex-0.4.14 → docvortex-0.4.16}/src/docvortex/version.py

@@ -1,3 +1,3 @@
-__version__ = "0.4.14"
+__version__ = "0.4.16"
 
 __all__ = ["__version__"]

{docvortex-0.4.14 → docvortex-0.4.16/src/docvortex.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: docvortex
-Version: 0.4.14
+Version: 0.4.16
 Summary: A fast, multi-format document parsing and conversion engine
 License-Expression: MIT
 Project-URL: Repository, https://github.com/myhloli/DocVortex

{docvortex-0.4.14 → docvortex-0.4.16}/src/docvortex.egg-info/SOURCES.txt

@@ -452,6 +452,7 @@ tests/test_history_geometry_tools.py
 tests/test_html_namespace.py
 tests/test_mixed_font_script_reference.py
 tests/test_pdf_analysis_api.py
+tests/test_pdf_cpu_evidence.py
 tests/test_pdf_font_geometry.py
 tests/test_pdf_font_runtime.py
 tests/test_pdf_model_text_normalization.py

{docvortex-0.4.14 → docvortex-0.4.16}/tests/test_document_metadata.py

@@ -156,7 +156,16 @@ def test_rtf_explicit_time_count_and_application() -> None:
 
 @pytest.mark.parametrize(
     "value,expected",
-    [("D:2024", "2024"), ("D:202402", "2024-02"), ("2024-03-04", "2024-03-04"), ("D:202413", None), ("yesterday", None)],
+    [
+        ("D:2024", "2024"),
+        ("D:202402", "2024-02"),
+        ("2024-03-04", "2024-03-04"),
+        ("D:202413", None),
+        ("yesterday", None),
+        ("D:20131014124204-04'", "2013-10-14T12:42:04-04:00"),
+        ("D:20240304050607+08", "2024-03-04T05:06:07+08:00"),
+        ("D:20240304050607+0", None),
+    ],
 )
 def test_dates_never_invent_missing_parts(value: str, expected: str | None) -> None:
     """非法日期保持未知，年和月精度不能补成任意日期。"""

{docvortex-0.4.14 → docvortex-0.4.16}/tests/test_pdf_analysis_api.py

@@ -23,7 +23,7 @@ from docvortex.analyzers.pdf import (
     prepare_text_evidence,
     recover_table_region,
 )
-from docvortex.document.pdf import PDFPageTextGeometry
+from docvortex.document.pdf import PDFPageTextGeometry, PDFPageVectorGeometry
 from docvortex.public_api import PUBLIC_API
 
 
@@ -112,6 +112,7 @@ def test_table_page_is_reused_and_returns_materialized_html(monkeypatch: pytest.
     page = MagicMock(size=(100.0, 200.0))
     page.get_drawing_lines.return_value = []
     page.get_path_infos.return_value = []
+    page.get_vector_geometry.return_value = PDFPageVectorGeometry()
     geometry = PDFPageTextGeometry([], {}, {})
     prepared = prepare_table_page(page, geometry=geometry)
     recover = MagicMock(
@@ -125,8 +126,9 @@ def test_table_page_is_reused_and_returns_materialized_html(monkeypatch: pytest.
     assert all(result.html == "<table><tr><td>x<sup>2</sup></td></tr></table>" for result in results)
     assert [call.args[0].angle for call in recover.call_args_list] == [0, 90]
     page.get_chars_with_geometry.assert_not_called()
-    page.get_drawing_lines.assert_called_once()
-    page.get_path_infos.assert_called_once()
+    page.get_vector_geometry.assert_called_once()
+    page.get_drawing_lines.assert_not_called()
+    page.get_path_infos.assert_not_called()
     assert prepared.geometry is geometry
     monkeypatch.setattr(table_materialization, "recover_table_result", lambda *_args: None)
     assert recover_table_region(prepared, (10, 20, 90, 80)) is None

docvortex-0.4.16/tests/test_pdf_cpu_evidence.py

@@ -0,0 +1,164 @@
+"""验证页面原语复用、源数据只读和行对筛选的保守等价性。"""
+
+from __future__ import annotations
+
+import json
+import pickle
+import random
+from copy import deepcopy
+from dataclasses import asdict
+from io import BytesIO
+from unittest.mock import MagicMock
+
+import pytest
+from reportlab.pdfgen.canvas import Canvas
+
+from docvortex.analyzers import pdf as analysis
+from docvortex.analyzers.native.pdf import line_merging as merging
+from docvortex.analyzers.native.pdf.models import _LineItem
+from docvortex.analyzers.native.pdf.native_text import _build_native_line_items
+from docvortex.document.pdf import Bbox, PDFDocument, PDFPageVectorGeometry, get_lines_from_chars
+from docvortex.document.pdf import _document
+
+
+def _pdf() -> bytes:
+    """生成含链接、旋转文本、上下标和矢量线的独立 PDF。"""
+    buffer = BytesIO()
+    canvas = Canvas(buffer, pagesize=(400, 600))
+    for index in range(25):
+        canvas.setFont("Helvetica", 12)
+        canvas.drawString(30, 570 - index * 20, f"Row {index}: x")
+        canvas.setFont("Helvetica", 7)
+        canvas.drawString(130, 574 - index * 20, "2")
+        canvas.line(20, 566 - index * 20, 180, 566 - index * 20)
+    canvas.linkURL("https://example.com", (30, 550, 90, 565), relative=0)
+    canvas.saveState()
+    canvas.translate(300, 200)
+    canvas.rotate(90)
+    canvas.drawString(0, 0, "Rotated text")
+    canvas.restoreState()
+    canvas.rect(20, 50, 170, 540)
+    canvas.save()
+    return buffer.getvalue()
+
+
+def test_vector_snapshot_matches_independent_queries_and_outlives_document(monkeypatch: pytest.MonkeyPatch) -> None:
+    """一次组合遍历与旧查询逐值一致，文档关闭后仍可安全使用。"""
+    with PDFDocument(_pdf()) as document:
+        page = document[0]
+        drawings, paths = page.get_drawing_lines(), page.get_path_infos()
+        extract = MagicMock(wraps=_document._extract_page_paths_and_lines)
+        monkeypatch.setattr(_document, "_extract_page_paths_and_lines", extract)
+        vectors = page.get_vector_geometry()
+        assert vectors == PDFPageVectorGeometry(tuple(drawings), tuple(paths))
+        extract.assert_called_once()
+    assert pickle.loads(pickle.dumps(vectors)) == vectors
+
+
+def test_shared_primitives_preserve_characters_and_text_evidence() -> None:
+    """共享字符与矢量几何不改动源对象，并保留区域相关的最终证据。"""
+    with PDFDocument(_pdf()) as document:
+        page = document[0]
+        geometry = page.get_chars_with_geometry()
+        vectors = page.get_vector_geometry()
+        before = pickle.dumps(geometry)
+        expected = analysis.prepare_text_evidence(page, geometry=geometry, table_regions=[(20, 30, 190, 300)])
+        materialized_page = MagicMock(size=page.size, rotation=page.rotation)
+        materialized_page.get_link_annotations.return_value = page.get_link_annotations()
+        actual = analysis.prepare_text_evidence(
+            materialized_page, geometry=geometry, vector_geometry=vectors, table_regions=[(20, 30, 190, 300)]
+        )
+        assert (actual.styles, actual.links, actual.scripts) == (expected.styles, expected.links, expected.scripts)
+        assert actual.geometry is geometry
+        analysis.prepare_table_page(materialized_page, geometry=geometry, vector_geometry=vectors)
+        materialized_page.get_chars_with_geometry.assert_not_called()
+        materialized_page.get_vector_geometry.assert_not_called()
+        materialized_page.get_drawing_lines.assert_not_called()
+        materialized_page.get_path_infos.assert_not_called()
+        assert pickle.dumps(geometry) == before
+
+
+def _groups(lines: list[_LineItem]) -> dict[tuple[int, bool, str | None], list[int]]:
+    """按原合并规则的必要分类条件建立索引。"""
+    groups: dict[tuple[int, bool, str | None], list[int]] = {}
+    for index, line in enumerate(lines):
+        groups.setdefault((line.angle, line.formula_candidate_only, line.semantic_type), []).append(index)
+    return groups
+
+
+def test_candidate_screening_never_drops_original_acceptance() -> None:
+    """覆盖来源框偏移和字体尺度大于可见框的情况，逐对验证原判定的接受超集。"""
+    rng = random.Random(918)
+    accepted = 0
+    for _ in range(30):
+        lines = []
+        for index in range(80):
+            x, y = rng.uniform(0, 200), rng.uniform(0, 200)
+            width, height = rng.uniform(0.1, 30), rng.uniform(0.1, 25)
+            box = (x, y, x + width, y + height)
+            lines.append(
+                _LineItem(
+                    "x",
+                    box,
+                    0,
+                    index,
+                    source_bbox=(x - 5, y - 10, x + width + 5, y + height + 10),
+                    baseline=y + height,
+                    chars=[{"char_idx": index, "source_indices": (index,), "bbox": Bbox(list(box)), "char": "x"}],
+                    visual_row_id=index,
+                    effective_height=rng.uniform(0.1, 80),
+                    font_signature=("test", 0),
+                    font_coverage=1,
+                )
+            )
+        candidates = merging._same_baseline_candidate_pairs(lines, [line.bbox for line in lines], _groups(lines))
+        assert candidates is not None
+        for left, first in enumerate(lines):
+            assert candidates[left] == sorted(set(candidates[left]))
+            for right in range(left + 1, len(lines)):
+                second = lines[right]
+                if merging._can_merge_same_baseline_pair(first, first.bbox, second, second.bbox, []):
+                    accepted += 1
+                    assert right in candidates[left]
+    assert accepted > 100
+
+
+@pytest.mark.parametrize("invalid", [float("nan"), float("inf"), -float("inf")])
+def test_invalid_geometry_uses_exhaustive_fallback(invalid: float) -> None:
+    """异常源框不能悄悄缩小候选，交回已有算法处理。"""
+    lines = [_LineItem("a", (0, 0, 20, 10), 0, 0), _LineItem("b", (21, 0, 40, 10), 0, 1)]
+    lines[1].source_bbox = (0, invalid, 40, 10)
+    assert merging._same_baseline_candidate_pairs(lines, [line.bbox for line in lines], _groups(lines)) is None
+
+
+def test_dense_page_bounds_candidate_storage() -> None:
+    """候选数量超过线性预算时回退，不分配整个平方级行对集合。"""
+    lines = [_LineItem("a", (0, 0, 20, 10), 0, index) for index in range(128)]
+    assert merging._same_baseline_candidate_pairs(lines, [line.bbox for line in lines], _groups(lines)) is None
+
+
+def test_bottom_edge_tolerance_is_included() -> None:
+    """可见框不交叠但有效字体尺度允许底边对齐时仍保留原规则候选。"""
+    lines = [
+        _LineItem("a", (0, 0, 10, 1), 0, 0, effective_height=40, font_signature=("font", 0), font_coverage=1),
+        _LineItem("b", (11, 10, 21, 11), 0, 1, effective_height=40, font_signature=("font", 0), font_coverage=1),
+    ]
+    assert merging._can_merge_same_baseline_pair(lines[0], lines[0].bbox, lines[1], lines[1].bbox, [])
+    assert merging._same_baseline_candidate_pairs(lines, [line.bbox for line in lines], _groups(lines)) == [[1], []]
+
+
+@pytest.mark.parametrize("angle", [0, 90, 180, 270])
+@pytest.mark.parametrize("tables", [[], [(0, 100, 400, 180)]])
+def test_merged_output_matches_exhaustive_order(monkeypatch: pytest.MonkeyPatch, angle: int, tables: list) -> None:
+    """比较完整行对象，确认筛选不会改变并查集认领、字符或结果排序。"""
+    with PDFDocument(_pdf()) as document:
+        page = document[0]
+        lines = _build_native_line_items(get_lines_from_chars(page.get_chars_with_geometry().chars), page.size)
+        for line in lines:
+            line.angle = angle
+        actual = merging.merge_text_line_clusters(deepcopy(lines), page.size, tables)
+        monkeypatch.setattr(merging, "_same_baseline_candidate_pairs", lambda *_args: None)
+        expected = merging.merge_text_line_clusters(deepcopy(lines), page.size, tables)
+        assert json.dumps([asdict(line) for line in actual], default=repr, sort_keys=True) == json.dumps(
+            [asdict(line) for line in expected], default=repr, sort_keys=True
+        )