docvortex 0.4.18__tar.gz → 0.4.20__tar.gz

{docvortex-0.4.18/src/docvortex.egg-info → docvortex-0.4.20}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: docvortex
-Version: 0.4.18
+Version: 0.4.20
 Summary: A fast, multi-format document parsing and conversion engine
 License-Expression: MIT
 Project-URL: Repository, https://github.com/myhloli/DocVortex

{docvortex-0.4.18 → docvortex-0.4.20}/docs/sdk-0.4.md

@@ -1,6 +1,6 @@
 # DocVortex 0.4 公共 SDK 与宿主迁移
 
-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。
+本文记录 0.4 系列的公共模块边界。MinerU 的共享图片导出需要 `docvortex>=0.4.19,<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
 
@@ -22,6 +22,29 @@
 
 跨库调用方可通过 `docvortex.export.materialize_middle(middle_json, assets=None)` 获取图片外置后的文档副本与 `AssetStore`，再通过 `validate_materialized_assets(document, assets)` 验证所有图片引用。直接图片和视觉 HTML 内嵌图片使用安全相对路径，原始图片字节及布局方向扩展保持不变；代码字面量不作为 HTML 素材处理。
 
+### 0.4.19 图片命名与宿主回调
+
+新物化图片统一沿用 MinerU Gradio 的命名：直接载荷为
+`images/page_{page_idx}_{owner.type}_{owner.index}.{ext}`，其中 owner 是所属
+image/table/chart 父块，独立公式使用 equation。视觉正文内的图片分别使用
+`image_{index}_{ordinal}`、`table_image_{index}_{ordinal}`、`chart_image_{index}_{ordinal}`。
+页号为从 0 开始的原始页索引；序号按每个正文的所有 `<img src=...>` 从 1 计数，外链也占序号。
+扩展名转小写，缺省 jpg；图片字节不重新编码。同名同字节复用，同名异字节追加
+`_duplicate_n`，最多检查 10000 个候选名称；不同语义名称不因字节相同而合并。
+
+`materialize_middle` 新增两个可选的仅限关键字参数：
+
+- `image_resolver: Callable[[ImagePayloadBlock, int], tuple[bytes, str] | None]`：接收副本中的载荷块和原始页号，返回字节及扩展名；返回 None 时保留载荷，不再触发默认解析。
+- `asset_resolver: Callable[[str], bytes]`：接收经过实体/URL 解码及安全校验的 HTML 图片相对路径，返回字节；扩展名从引用路径取得。仅在明确传入回调时读取宿主资源，回调负责自身根目录限制。
+
+不提供回调时保持内嵌图片的严格解析，不自动读文件或裁 PDF；已有相对路径保持原样。
+输入 `assets` 保留为独立副本并参与冲突检查，重复物化不改变路径或字节。
+新命名同步更新直接引用和 HTML src，保留 HTML 的属性、引号及其他正文。
+
+此变更会改变新生成的 CLI/API ZIP 及 DocVortex 导出图片名称，消费者应读取文档实际引用，
+不再自行拼接旧的 `*_body_*` 名称。历史结果包仍按已有引用读取，不自动迁移、不生成旧名别名；
+MiddleJson 仍为 2.0，文件写出层的 overwrite 语义不变。
+
 调用方负责将该副本及素材写入自己的结果包，不应先省略图片再从源 PDF 重裁。缺失引用、未物化网络图片或同名素材冲突会显式失败。物化不修改输入对象、不读取源 PDF、不下载网络资源；已有外部素材须通过 `AssetStore` 显式提供。
 
 ## 行为与资源所有权

{docvortex-0.4.18 → docvortex-0.4.20}/src/docvortex/analyzers/native/pdf/_script_geometry.py

@@ -30,6 +30,7 @@ SCRIPT_COMPONENT_SEED_MAX_POSITION_DISTANCE = 4
 SCRIPT_COMPONENT_MIN_OFFSET_RATIO = 0.08
 SCRIPT_COMPONENT_MAX_HEIGHT_RATIO = 1.1
 CONTROL_LINE_BREAK_CHARS = {"\r", "\n"}
+_NUMERIC_SCRIPT_SEPARATORS = frozenset({",", "，", "-", "–", "—"})
 
 ScriptRole = Literal["body", "sup", "sub"]
 ScriptMarkRole = Literal["sup", "sub"]
@@ -675,6 +676,91 @@ def paired_script_roles(
     return paired
 
 
+def _numeric_superscript_indices(
+    features: list[ScriptCharFeature],
+    roles: list[ScriptRole],
+    protected_body_indices: set[int],
+) -> set[int]:
+    """用左侧同一视觉行的稳定正文确认完整数字上标，允许中间有排版空白。"""
+    accepted: set[int] = set()
+    start = 0
+    while start < len(features):
+        if not features[start].text.isdecimal():
+            start += 1
+            continue
+        end = start + 1
+        while end < len(features) and (features[end].text.isdecimal() or features[end].text in _NUMERIC_SCRIPT_SEPARATORS):
+            end += 1
+        # 尾随标点属于正文，只有数字之间且共享上标基线的分隔符才随引用提升。
+        while end > start and not features[end - 1].text.isdecimal():
+            end -= 1
+        candidate = features[start:end]
+        if any(0 <= neighbor < len(features) and features[neighbor].text in {"/", "⁄"} for neighbor in (start - 1, end)):
+            # 不能把分式的分子或分母独立当成引用，否则会制造仅半个指数带上标的结果。
+            start = end
+            continue
+        previous = start - 1
+        while previous >= 0 and features[previous].text.isspace():
+            if features[previous].text in CONTROL_LINE_BREAK_CHARS:
+                break
+            previous -= 1
+        reference_end = previous + 1
+        while previous >= 0 and features[previous].is_valid and roles[previous] == "body":
+            if previous in protected_body_indices:
+                break
+            previous -= 1
+        reference = list(range(previous + 1, reference_end))
+        start = end
+        if not reference or any(
+            not item.is_valid or item.index in protected_body_indices or roles[item.index] == "sub" for item in candidate
+        ):
+            continue
+        clusters, _ = _cluster_baselines(features, reference)
+        body_result = _choose_body_band(features, clusters)
+        if body_result is None:
+            continue
+        band, body_cluster = body_result
+        if len(band.member_indices) < 2 or band.tight_height <= 0 or band.loose_height <= 0:
+            continue
+        edge = features[reference[-1]]
+        first = candidate[0]
+        assert first.tight_bbox is not None and edge.tight_bbox is not None
+        gap = first.tight_bbox[0] - edge.tight_bbox[2]
+        if not -0.15 * band.tight_height <= gap <= max(2.5, 1.2 * band.tight_height):
+            continue
+        digits = [item for item in candidate if item.text.isdecimal()]
+        baseline = statistics.median(item.origin[1] for item in digits if item.origin is not None)
+        shift = band.baseline - baseline
+        if (
+            not max(SCRIPT_ORIGIN_MIN_SHIFT_ABSOLUTE, SCRIPT_CONSENSUS_ORIGIN_SHIFT_RATIO * band.tight_height)
+            <= shift
+            <= (0.8 * band.tight_height)
+        ):
+            continue
+        tolerance = max(SCRIPT_BASELINE_ABSOLUTE_TOLERANCE, SCRIPT_BASELINE_LOOSE_HEIGHT_RATIO * band.loose_height)
+        if any(item.origin is None or abs(item.origin[1] - baseline) > tolerance for item in candidate):
+            continue
+        tight_ratio = max(item.tight_height for item in digits) / band.tight_height
+        if tight_ratio > SCRIPT_CONSENSUS_TIGHT_HEIGHT_RATIO:
+            # 纯小写词的 x-height 可接近缩小数字；强基线位移还须有 loose 缩小证据，不能仅放宽 tight 阈值。
+            loose_ratio = max(item.loose_height for item in digits) / band.loose_height
+            if not (
+                shift >= SCRIPT_STRONG_SHIFT_RATIO * band.tight_height
+                and tight_ratio <= SCRIPT_STRONG_MAX_HEIGHT_RATIO
+                and loose_ratio <= SCRIPT_CONSENSUS_TIGHT_HEIGHT_RATIO
+            ):
+                continue
+        digit_cluster = ScriptBaselineCluster(baseline, tuple(item.index for item in digits))
+        if (
+            _cluster_tight_center(features, digit_cluster) >= _cluster_tight_center(features, body_cluster) - 0.05
+            or _cluster_loose_center(features, digit_cluster) >= _cluster_loose_center(features, body_cluster) - 0.05
+        ):
+            continue
+        # 只补样式，不合并组件或移动字符；下一候选也不能把本次提升结果当作正文。
+        accepted.update(item.index for item in candidate)
+    return accepted
+
+
 def classify_char_script_roles(
     chars: list[Char],
     *,
@@ -689,6 +775,8 @@ def classify_char_script_roles(
     roles: list[ScriptRole] = ["body"] * len(features)
     for component_indices in split_script_visual_components(features):
         _assign_component(features, component_indices, protected, roles, font_keys)
+    for index in _numeric_superscript_indices(features, roles, protected):
+        roles[index] = "sup"
     return roles
 
 

{docvortex-0.4.18 → docvortex-0.4.20}/src/docvortex/analyzers/native/pdf/inline/scripts.py

@@ -7,7 +7,13 @@ import unicodedata
 from typing import Any, Literal, Sequence
 
 from .....schema import BBox
-from .._script_geometry import ScriptRole, classify_char_script_roles, paired_script_roles
+from .._script_geometry import (
+    ScriptRole,
+    _numeric_superscript_indices,
+    build_script_features,
+    classify_char_script_roles,
+    paired_script_roles,
+)
 from ..geometry import _rotate_bbox_to_upright
 from .common import _coerce_bbox, _normalize_match_fragment, _ordered_line_chars
 from .types import (
@@ -411,6 +417,11 @@ def _refine_math_script_tokens(
     """以最左稳定簇保护 base，并对弱单字符和复杂未分段 token 保守拒识。"""
     refined = list(roles)
     citation_indices = _citation_script_indices(chars, refined)
+    numeric_indices: set[int] = set()
+    if any(role == "sup" and _script_char_text(char).isdecimal() for char, role in zip(chars, roles)):
+        # 数字引用的基字符在 token 外；保护双 bbox/origin 已确认的整串，不能重置首位数字。
+        candidates = _numeric_superscript_indices(build_script_features(chars, tight_bboxes, origins, set()), roles, set())
+        numeric_indices = {index for index in candidates if roles[index] == "sup"}
     complex_unsegmented_token = False
     tokens = _iter_math_script_tokens(chars)
     token_alnum_positions = {
@@ -443,6 +454,9 @@ def _refine_math_script_tokens(
     for token in tokens:
         if any(index in citation_indices for index in token):
             continue
+        if all(index in numeric_indices for index in token):
+            # 只跳过纯数字引用的 base 重选，后续分式及复杂数学过滤仍按原规则执行。
+            continue
         token_key = tuple(token)
         alnum_positions = token_alnum_positions[token_key]
         if not alnum_positions or not any(refined[index] != "body" for index in token):

docvortex-0.4.20/src/docvortex/export/_images.py

@@ -0,0 +1,134 @@
+"""各导出入口共享的图片命名、冲突分配与引用改写。"""
+
+from __future__ import annotations
+
+import html
+import re
+from collections.abc import Callable
+from pathlib import PurePosixPath
+from urllib.parse import unquote
+
+from ..assets import AssetStore, parse_image_data_uri_strict, validate_image_sidecar_path
+from ..content.tree import iter_child_blocks
+from ..schema import (
+    AlgorithmBodyBlock,
+    BlockBase,
+    ChartBlock,
+    ChartBodyBlock,
+    CodeBlock,
+    CodeBodyBlock,
+    ImageBlock,
+    ImageBodyBlock,
+    ImagePayloadBlock,
+    MiddleJson,
+    TableBlock,
+    TableBodyBlock,
+)
+
+_HTML_IMAGE_SOURCE_RE = re.compile(
+    r"(?P<prefix><img\b(?:[^>\"']|\"[^\"]*\"|'[^']*')*?\s+src\s*=\s*)"
+    r"(?:(?P<quote>[\"'])(?P<quoted>.*?)(?P=quote)|(?P<unquoted>[^\s>]+))",
+    re.IGNORECASE | re.DOTALL,
+)
+_IMAGE_EXTENSIONS = frozenset({"jpg", "jpeg", "png", "gif", "webp", "bmp", "tiff", "svg"})
+
+
+def _register_image(
+    assets: AssetStore,
+    payload: bytes,
+    extension: str,
+    *,
+    page_idx: int,
+    owner: BlockBase,
+    ordinal: int | None = None,
+) -> str:
+    """按原始页和视觉父块命名，冲突后缀独立于正文图片序号。"""
+    extension = extension.lower().lstrip(".") or "jpg"
+    if extension not in _IMAGE_EXTENSIONS:
+        raise ValueError(f"Unsupported image extension: {extension}")
+    if owner.index is None:
+        raise ValueError("Image owner must have a block index")
+    kind = str(owner.type)
+    if ordinal is not None and kind != "image":
+        kind += "_image"
+    suffix = f"_{ordinal}" if ordinal is not None else ""
+    stem = f"page_{page_idx}_{kind}_{owner.index}{suffix}"
+    for attempt in range(10000):
+        duplicate = f"_duplicate_{attempt}" if attempt else ""
+        path = validate_image_sidecar_path(f"images/{stem}{duplicate}.{extension}")
+        if path not in assets or assets[path] == payload:
+            assets.add(path, payload)
+            return path
+    raise ValueError("Too many conflicting image assets")
+
+
+def _materialize_markup(
+    content: str,
+    assets: AssetStore,
+    *,
+    page_idx: int,
+    owner: BlockBase,
+    asset_resolver: Callable[[str], bytes] | None,
+) -> str:
+    """按所有 img src 的出现顺序编号，仅替换图片引用的值。"""
+    ordinal = 0
+
+    def replace_source(match: re.Match[str]) -> str:
+        """保留外链及原 HTML 属性，外链同样占用正文图片序号。"""
+        nonlocal ordinal
+        ordinal += 1
+        group = "quoted" if match.group("quote") else "unquoted"
+        source = html.unescape(match.group(group)).strip()
+        if source.startswith("data:"):
+            payload, extension = parse_image_data_uri_strict(source)
+        elif re.match(r"^[A-Za-z][A-Za-z0-9+.-]*:", source) or asset_resolver is None:
+            return match.group(0)
+        else:
+            safe_path = validate_image_sidecar_path(unquote(source))
+            payload = asset_resolver(safe_path)
+            extension = PurePosixPath(safe_path).suffix.lstrip(".")
+        path = _register_image(assets, payload, extension, page_idx=page_idx, owner=owner, ordinal=ordinal)
+        start, end = match.span(group)
+        original = match.group(0)
+        return original[: start - match.start()] + path + original[end - match.start() :]
+
+    return _HTML_IMAGE_SOURCE_RE.sub(replace_source, content)
+
+
+def _materialize_images(
+    middle_json: MiddleJson,
+    assets: AssetStore | None = None,
+    *,
+    image_resolver: Callable[[ImagePayloadBlock, int], tuple[bytes, str] | None] | None = None,
+    asset_resolver: Callable[[str], bytes] | None = None,
+) -> tuple[MiddleJson, AssetStore]:
+    """复制文档及素材索引，在同一次树遍历中完成命名和引用回填。"""
+    document = middle_json.model_copy(deep=True)
+    result = assets.copy() if assets is not None else AssetStore()
+    for page in document.pages:
+        pending = [(block, block) for block in reversed(page.blocks)]
+        while pending:
+            block, owner = pending.pop()
+            if isinstance(block, (CodeBlock, CodeBodyBlock, AlgorithmBodyBlock)):
+                continue
+            if isinstance(block, (ImageBlock, TableBlock, ChartBlock)):
+                owner = block
+            if isinstance(block, ImagePayloadBlock):
+                resolved = None
+                if image_resolver is not None:
+                    resolved = image_resolver(block, page.page_idx)
+                elif block.image_base64 is not None:
+                    resolved = parse_image_data_uri_strict(block.image_base64)
+                if resolved is not None:
+                    payload, extension = resolved
+                    block.image_path = _register_image(result, payload, extension, page_idx=page.page_idx, owner=owner)
+                    block.image_base64 = None
+                    block.image_url = None
+                elif image_resolver is None and block.image_path and block.image_path in result:
+                    block.image_url = None
+            if isinstance(block, (ImageBodyBlock, TableBodyBlock, ChartBodyBlock)):
+                block.content = _materialize_markup(
+                    block.content, result, page_idx=page.page_idx, owner=owner, asset_resolver=asset_resolver
+                )
+            pending.extend((child, owner) for child in reversed(list(iter_child_blocks(block))))
+    return document, result

{docvortex-0.4.18 → docvortex-0.4.20}/src/docvortex/export/files.py

@@ -2,30 +2,32 @@
 
 from __future__ import annotations
 
+from collections.abc import Callable
 from pathlib import Path
 from urllib.parse import urlsplit
 
 from ..assets import AssetStore
 from ..content.tree import iter_child_blocks
-from ..schema import MiddleJson
 from ..result import ExportResult, RenderArtifact
-from .middle import _commit_export_files, _prepare_export_copy, _resolve_export_target, _validate_export_path_relationships
+from ..schema import ImagePayloadBlock, MiddleJson
+from ._images import _materialize_images
+from .middle import _commit_export_files, _resolve_export_target, _validate_export_path_relationships
 
 
-def materialize_middle(middle_json: MiddleJson, assets: AssetStore | None = None) -> tuple[MiddleJson, AssetStore]:
-    """在文档副本上外置所有内嵌图片，保留调用者提供的外部素材。"""
-    document, embedded = _prepare_export_copy(middle_json)
-    result = assets.copy() if assets is not None else AssetStore()
-    for path, payload in embedded.items():
-        result.add(path, payload)
-    pending = [block for page in document.pages for block in page.blocks]
-    while pending:
-        block = pending.pop()
-        pending.extend(iter_child_blocks(block))
-        image_path = getattr(block, "image_path", None)
-        if image_path and image_path in result and getattr(block, "image_url", None):
-            block.image_url = None
-    return document, result
+def materialize_middle(
+    middle_json: MiddleJson,
+    assets: AssetStore | None = None,
+    *,
+    image_resolver: Callable[[ImagePayloadBlock, int], tuple[bytes, str] | None] | None = None,
+    asset_resolver: Callable[[str], bytes] | None = None,
+) -> tuple[MiddleJson, AssetStore]:
+    """在副本上按视觉父块外置图片；仅通过显式回调获取宿主文件或裁图。
+
+    image_resolver 接收载荷块和原始页索引，返回字节与扩展名；返回 None
+    表示保留当前载荷，不触发默认解析。asset_resolver 仅接收经过校验的
+    HTML 图片相对路径。不提供回调时，只解析内嵌图片并保留已有素材引用。
+    """
+    return _materialize_images(middle_json, assets, image_resolver=image_resolver, asset_resolver=asset_resolver)
 
 
 def validate_materialized_assets(middle_json: MiddleJson, assets: AssetStore) -> None:

{docvortex-0.4.18 → docvortex-0.4.20}/src/docvortex/export/middle.py

@@ -6,9 +6,9 @@ import os
 from dataclasses import dataclass
 from pathlib import Path
 from tempfile import NamedTemporaryFile
-from typing import Any, Sequence
 
-from ..schema import BlockBase, MiddleJson, _iter_child_blocks
+from ..schema import MiddleJson
+from ._images import _materialize_images
 
 
 @dataclass(frozen=True, slots=True)
@@ -20,65 +20,10 @@ class MiddleJsonExportResult:
     image_paths: tuple[Path, ...]
 
 
-def _iter_page_blocks(blocks: Sequence[BlockBase]) -> list[BlockBase]:
-    """按深度优先顺序展开页面 block 树，供图片外置统一遍历。"""
-    result: list[BlockBase] = []
-    pending: list[BlockBase] = list(reversed(blocks))
-    while pending:
-        block = pending.pop()
-        result.append(block)
-        pending.extend(reversed(_iter_child_blocks(block)))
-    return result
-
-
-def _register_export_file(files: dict[str, bytes], relative_path: str, payload: bytes) -> None:
-    """登记待写文件，并在同名内容冲突时立即终止导出。"""
-    existing = files.get(relative_path)
-    if existing is not None and existing != payload:
-        raise ValueError(f"Conflicting image payload for path: {relative_path}")
-    files[relative_path] = payload
-
-
 def _prepare_export_copy(middle_json: MiddleJson) -> tuple[MiddleJson, dict[str, bytes]]:
-    """复制对象、解析直接及 HTML 图片，并回填副本中的相对路径。"""
-    from ..foundation._image_payload import INLINE_IMAGE_DATA_URI_RE, parse_image_data_uri_strict
-
-    exported = middle_json.model_copy(deep=True)
-    image_files: dict[str, bytes] = {}
-    for page in exported.pages:
-        for block in _iter_page_blocks(page.blocks):
-            data_uri = getattr(block, "image_base64", None)
-            if data_uri is not None:
-                if block.index is None:
-                    raise ValueError(f"Image carrier requires index: page_idx={page.page_idx}, type={block.type}")
-                image_bytes, extension = parse_image_data_uri_strict(data_uri)
-                relative_path = f"images/page_{page.page_idx}_{block.type}_{block.index}.{extension}"
-                _register_export_file(image_files, relative_path, image_bytes)
-                block.image_path = relative_path  # type: ignore[attr-defined]
-                block.image_base64 = None  # type: ignore[attr-defined]
-
-            content = getattr(block, "content", None)
-            if (
-                str(block.type) not in {"table_body", "chart_body", "image_body"}
-                or not isinstance(content, str)
-                or "data:image/" not in content
-            ):
-                continue
-            if block.index is None:
-                raise ValueError(f"HTML image carrier requires index: page_idx={page.page_idx}, type={block.type}")
-            ordinal = 0
-
-            def _replace_data_uri(match: Any) -> str:
-                """将当前 HTML data URI 登记为 sidecar，并返回确定性相对路径。"""
-                nonlocal ordinal
-                ordinal += 1
-                image_bytes, extension = parse_image_data_uri_strict(match.group(0))
-                relative_path = f"images/page_{page.page_idx}_{block.type}_{block.index}_{ordinal}.{extension}"
-                _register_export_file(image_files, relative_path, image_bytes)
-                return relative_path
-
-            block.content = INLINE_IMAGE_DATA_URI_RE.sub(_replace_data_uri, content)  # type: ignore[assignment]
-    return exported, image_files
+    """复用统一物化流程，为原子文件导出提供文档副本及图片字节。"""
+    document, assets = _materialize_images(middle_json)
+    return document, dict(assets)
 
 
 def _resolve_export_target(output_root: Path, relative_path: str) -> Path:

{docvortex-0.4.18 → docvortex-0.4.20}/src/docvortex/version.py

@@ -1,3 +1,3 @@
-__version__ = "0.4.18"
+__version__ = "0.4.20"
 
 __all__ = ["__version__"]

{docvortex-0.4.18 → docvortex-0.4.20/src/docvortex.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: docvortex
-Version: 0.4.18
+Version: 0.4.20
 Summary: A fast, multi-format document parsing and conversion engine
 License-Expression: MIT
 Project-URL: Repository, https://github.com/myhloli/DocVortex

{docvortex-0.4.18 → docvortex-0.4.20}/src/docvortex.egg-info/SOURCES.txt

@@ -343,6 +343,7 @@ src/docvortex/document/pdf/text/extract.py
 src/docvortex/document/pdf/text/geometry.py
 src/docvortex/document/pdf/text/groups.py
 src/docvortex/export/__init__.py
+src/docvortex/export/_images.py
 src/docvortex/export/bundle.py
 src/docvortex/export/files.py
 src/docvortex/export/middle.py
@@ -446,6 +447,7 @@ tests/test_complete_pipeline.py
 tests/test_converter_lifecycle.py
 tests/test_document_envelope.py
 tests/test_document_metadata.py
+tests/test_export_image_names.py
 tests/test_format_matrix.py
 tests/test_geometry_api.py
 tests/test_history_geometry_tools.py
@@ -456,6 +458,7 @@ 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
+tests/test_pdf_numeric_citations.py
 tests/test_pdf_output_text.py
 tests/test_pdf_render_pool.py
 tests/test_pdf_source_row_merging.py
@@ -496,6 +499,7 @@ tests/fixtures/native_pdf_table_demo_manifest.json
 tests/fixtures/native_pdf_table_script_truth.json
 tests/fixtures/pdf_char_dedup.json
 tests/fixtures/pdf_mixed_font_script_line.json
+tests/fixtures/pdf_numeric_citations_review.json
 tests/fixtures/rtf/semantic.rtf
 tests/unittest/_docx_equationxml_test_utils.py
 tests/unittest/_epub_test_utils.py
@@ -641,6 +645,7 @@ tests/unittest/test_text_utils.py
 tests/unittest/pdfs/flash_table_annotations_synthetic.pdf
 tests/unittest/pdfs/metabolic_pathway_page_3.pdf
 tests/unittest/pdfs/native_cjk_layout_synthetic.pdf
+tests/unittest/pdfs/numeric_citations_introduction.pdf
 tests/unittest/pdfs/test.pdf
 tests/unittest/pdfs/flash_layout/bloom_form_labels.pdf
 tests/unittest/pdfs/flash_layout/journal_dual_caption.pdf

{docvortex-0.4.18 → docvortex-0.4.20}/tests/fixtures/flash_reviewed_history.json

@@ -610,7 +610,7 @@
         },
         {
           "page_index": 2,
-          "fingerprint": "223f8dc3618805a8688095824d638d00348540403923ffddb74ba95a3a1961df",
+          "fingerprint": "0dfc82cf2795f0ed939b1c77e7c53c59f0788c28c5f1bef1ae4bd4e8fff8b8c7",
           "bbox_fingerprint": "c83743f89e12f6984bd9eb20f37303a668c8df982be11f86c9ae4851e035aafe"
         },
         {

docvortex-0.4.20/tests/fixtures/pdf_numeric_citations_review.json

@@ -0,0 +1,33 @@
+{
+  "accepted": [
+    {
+      "case": "klee-numeric-footnote",
+      "source": "demo/pdfs/mixed_elements_pages_07_10.pdf",
+      "source_sha256": "89076ee9cce2168ab894ee7e280872f8a3c09053b9dcdeda9c1da76d111adc1e",
+      "page": 3,
+      "anchor": "symbolic execution tool KLEE 1",
+      "before_code_sha256": "be87727e75b972fd5c409c6d6792fb8094443daa5e8a69f8044080d435f261f5",
+      "after_code_sha256": "22c18d95daf883ed457012463d5616d209e4c4c6e632815225cc116189f395f5",
+      "old_fingerprint": "223f8dc3618805a8688095824d638d00348540403923ffddb74ba95a3a1961df",
+      "new_fingerprint": "0dfc82cf2795f0ed939b1c77e7c53c59f0788c28c5f1bef1ae4bd4e8fff8b8c7",
+      "decision": "accept_improvement",
+      "reason": "原页右栏 KLEE 后的 1 明显抬高且缩小；补为上标，字符、顺序、块型和 bbox 均不变。",
+      "baseline_update": "flash_reviewed_history.json: mixed_elements_pages_07_10 page_index=2 fingerprint",
+      "evidence": [
+        "output/pdf/intro-superscripts-20260920/klee-source.png",
+        "output/pdf/intro-superscripts-20260920/klee-public.html",
+        "output/pdf/intro-superscripts-20260920/history-final/comparison.html"
+      ]
+    }
+  ],
+  "rejected": [
+    {
+      "source": "mixed_elements_pages_03_06.pdf",
+      "pages": [
+        2,
+        3
+      ],
+      "reason": "拒绝只提升分式指数分子的候选；保留原数学过滤及原基线。"
+    }
+  ]
+}