Per PDF spec, CS/cs also reset the current color to the color space's initial value: black for the device spaces and the resolved resource spaces we a

Set the log level using a string label. Parameters: level (str): Logging level as a string. One of ['fatal', 'er

Decides whether a text cell should use glyph-bbox fitting when glyph bbox metadata is available. ASCII alphanumeric/punctuation/space text usually ren

Get all static timing key names.

Every regression PDF must parse without exception under French locale.

Test that annotations match parser groundtruth.

Requesting False then True for the same page returns the correct result each time.

Requesting True then False for the same page returns the correct result each time.

include_bitmap_bytes=False produces placeholders with image=None.

bitmap_resources count and rects match regardless of bitmap bytes.

Test loading PDF with different boundary types.

Loading from BytesIO must work under comma locale.

Text content extraction must not depend on locale. While the primary bug is coordinate corruption, we verify that the text itself is

Single-character cells expose their decoded PDF character code.

Opening without word cells, then requesting them, re-decodes the page.

Parse same PDF under alternating locales — results must be identical.

The C++ decode config exposes extraction, defaulting to off. Parse-only pipelines must not pay for font stream decoding; the threaded render

Two documents with identical fonts share one embedded-face cache. The cache inside one parser is shared by all workers and documents and is k

All instructions of one PDF font reference the same blob. The content-hash cache key is the observable identity of the shared blob: one font

Smoke test: every embedded-font-bearing page renders successfully. Covers the .notdef fallback (subset fonts missing glyphs must not produce

An embedded TrueType face must actually be used for drawing. With resolve_fonts=False the only alternatives are the hardcoded fallback font o

Type 1 embedded fonts render through the FreeType outline path. Blend2D rejects raw /FontFile programs, so the renderer decomposes their glyp

Text instructions carry embedded font metadata (never the bytes).

Test accessing document annotations.

Parse-only results must fail loudly when image access is requested.

Test accessing individual pages without iterating all pages.

Cells of fonts with /Encoding /Differences expose their glyph name. The page-1 NimbusRomNo9L fonts carry a /Differences array, so at least so

The guard must not change anything when LC_NUMERIC is already safe.

Importing docling_parse must ensure LC_NUMERIC uses '.' separator.

Verify that lazy and eager loading produce identical pages.

Test document key management through load/unload lifecycle.

Test loading PDF from BytesIO with lazy=False.

Test loading PDF from BytesIO with lazy=True.

All pages of a multi-page PDF must parse correctly.

Page dimensions must retain their fractional part. The pre-fix bug would truncate 612.0 → 612 (benign) but 595.276 → 595 (wrong by 0.

Page count must not depend on locale.

Page dimensions must not be truncated under French locale. This is the core bug: a page width of 612.0 points would be parsed as 612

Verify rendering completes correctly with max_concurrent_results=1.

RenderConfig exposes the opt-in glyph bbox fit flag.

RenderConfig exposes the embedded-font opt-out, defaulting to on.

Parser accepts a non-default RenderConfig without error.

Render a document loaded from a BytesIO object.

Verify rendered image dimensions are positive and stable.

Load multiple PDFs and verify all pages are rendered.

Render all regression PDFs and verify parse output against groundtruth.

Render all pages of one document and verify each result is a valid RGBA image.

Render with a single thread as a sequential baseline.

Compare Blend2D threaded-render output with renderer groundtruth artifacts.