Layout model produces different SectionHeaderItem vs TextItem classification for same PDF on macOS ARM vs Linux x86_64 (same CPU device, same versions)

Bug
The same PDF document produces different layout classification results depending on the CPU architecture. On macOS ARM (Apple Silicon), a section heading is correctly classified as SectionHeaderItem with label=section_header. On Linux x86_64 (Google Colab), the exact same block is classified as TextItem with label=text and merged with the preceding paragraph, causing the heading to disappear entirely from the output.
Specifically, the heading ADVERSE REACTIONS: (see also WARNINGS and PRECAUTIONS) from an FDA drug label PDF is correctly detected on macOS ARM but completely missed on Linux x86_64.
macOS ARM output:
[DEBUG ITEM] type=SectionHeaderItem label=section_header level=1
[DEBUG ITEM] text='ADVERSE REACTIONS: (see also WARNINGS and PRECAUTIONS)'
Linux x86_64 output:
[DEBUG ITEM] type=TextItem label=text level=1
[DEBUG ITEM] text='Labor and Delivery: ... Nursing Mothers: ... serious adve'
# ADVERSE REACTIONS heading is completely absent - merged into previous block