Words lose their separating spaces in 7.21.0 ("Related Work" -> "RelatedWork"), bisected to #351

[YHC66]

What happens

Since 7.21.0, some word boundaries that 7.20.0 recovered are lost, most visibly in bold section headings of LaTeX papers. Using 2305.03393v1-pg9.pdf from docling's test data (file):

import sys
from docling_core.types.doc.page import TextCellUnit
from docling_parse.pdf_parser import DoclingPdfParser

doc = DoclingPdfParser(loglevel="fatal").load(path_or_stream=sys.argv[1])
for page_no, page in doc.iterate_pages():
    for c in page.iterate_cells(unit_type=TextCellUnit.LINE):
        if "Optimization" in c.text:
            print(repr(c.text))

docling-parse	output
7.20.0	'Hyper Parameter Optimization'
7.21.0	'5.1 HyperParameterOptimization'

The same thing shows up across docling's test PDFs (line cells, 7.20.0 -> 7.21.0):

• 2305.03393v1.pdf: Related Work -> 2 RelatedWork, Problem Statement -> 3 ProblemStatement, Fig. 1. -> Fig.1. Comparison between ..., Fig. 3. -> Fig.3. ..., Language Definition -> LanguageDefinition
• 2305.03393v1-pg9.pdf: Quantitative Results -> QuantitativeResults
• 2203.01017v2.pdf: Implementation Details -> ImplementationDetails, custom OCR -> customOCR, problem where -> problemwhere
• redp5110_sampled.pdf: dotted TOC leaders get attached to the preceding word (Roles., duties.)

In docling this makes tests/test_interfaces.py::test_convert_path, ::test_convert_stream and tests/test_heading_hierarchy_pdf.py::test_pdf_pipeline_assigns_heading_levels_from_existing_fixture fail against their ground truth. docling's CI doesn't see it because its uv.lock still pins 7.20.0, but pip install docling resolves docling-parse>=7.20.0,<8.0.0 to 7.21.0.

Bisect

Built from source on macOS arm64 (Python 3.12):

• 669616c (fix: fall back to real metrics when a visible glyph declares zero width #328): 'Hyper Parameter Optimization' (correct)
• e64d838 (feat: bbox of brackets, summations, missing font-mappings etc #351): '5.1 HyperParameterOptimization'

So the change comes from the cell-contraction rewrite in #351. Most of the affected lines are short, and the lost gaps are the ones between words of a single bold run. That fits the new line-local threshold (two-cluster k-means over normalized gaps) putting real word gaps in the "intra-word" cluster when a line has few gaps and no explicit space glyphs. That's a guess from reading the PR description, though; I haven't traced it in the code.

Environment

• docling-parse 7.21.0 (PyPI wheel) and a source build of e64d838
• macOS 26 arm64, Python 3.12