HTMLTextSerializer adds the <p> or <hN> tag of a text item first and applies the item's formatting and hyperlink afterwards, so a bold paragraph is exported as <strong><p>Bold paragraph</p></strong>. <strong>, <em>, <u>, <sub> and <sup> only accept phrasing content, so a paragraph or a heading inside them is invalid HTML.

Resolves #874

Root cause

In HTMLTextSerializer.serialize, the heading branch and the regular text branch wrap the text in <hN> or <p>, and only after that the shared "Apply formatting and hyperlinks to the parent's own text+tag" step calls doc_serializer.post_process. The list item branch already calls post_process before it adds <li>, so list items were not affected. A bold-only table cell is a RichTableCell that is serialized through the same text path, which gave <th><strong><p>Bold cell</p></strong></th>.

Change

• The formatting and the hyperlink are applied once, right after the text is escaped, for every text item except code, formulas and list items, so they end up inside the <p> or <hN> tag. List items already did this in their own branch (after review, eb6487f).
• The hyperlink and strikethrough are applied in the same step, so they move inside too: <p><a href="..."><strong>...</strong></a></p>, the same order as for list items.
• Formulas, code and text inside an inline group keep the current behavior. A text item whose content is an inline group is unchanged as well, since its runs carry their own formatting.

Behavior on a docx like the one in the issue (one bold paragraph, and a table whose first cell is bold), converted with docling 2.138.0:

• Before: <strong><p>Bold paragraph</p></strong> and <th><strong><p>Bold cell</p></strong></th>
• After: <p><strong>Bold paragraph</strong></p> and <th><p><strong>Bold cell</strong></p></th>

Ground truth

Regenerated with DOCLING_GEN_TEST_DATA=1. Each file changes in one line, and only in the order of the tags (same text, same tags):

• tests/data/doc/polymers.gt.html: <strong><p>Common migration testing methods</p></strong> becomes <p><strong>Common migration testing methods</strong></p>.
• tests/data/doc/rich_table.gt.html: the italic rich cell, <em><p>text in italic</p></em> becomes <p><em>text in italic</em></p>.
• tests/data/doc/inline_and_formatting.gt.html: <em><h1>Whole heading is italic</h1></em> becomes <h1><em>Whole heading is italic</em></h1>.
• tests/data/chunker/0c_out_chunks.json (test_chunk_rich_table_custom_serializer, which uses the HTML serializer): the same italic rich cell.

Downstream, docling's tests/data/docx/groundtruth/docx_rich_cells.docx.html has the old tag order and will need regenerating when docling picks up this change.

Tests

New tests in tests/test_serialization.py:

• test_html_text_formatting_inside_paragraph: bold, italic, underline, strikethrough, subscript, superscript, a hyperlink, and bold with italic and a hyperlink.
• test_html_heading_formatting_inside_heading_tag: an italic title and a bold section header with a hyperlink.
• test_html_rich_table_cell_formatting_inside_paragraph: bold-only rich cells in a header row (<th>) and in a body row (<td>).

All 10 test cases fail on main and pass with this change.

Results (Windows 11, Python 3.10.20, uv sync --frozen --all-extras --all-packages):

• uv run pytest tests/test_serialization.py tests/test_hierarchical_chunker.py tests/test_hybrid_chunker.py tests/test_plain_text_serialization.py tests/test_latex_serialization.py: 255 passed, 1 xfailed.
• uv run pytest tests: 701 passed, 287 failed (on eb6487f). The same 287 tests fail on main on this machine (DocLang Schematron validation through Java, /tmp paths and Windows path separators); none of them is an HTML serializer test.
• uv run pytest packages/dclq/tests: 10 passed.
• uv run pre-commit run <hook> --files <changed files> for ruff-check, ruff-format and mypy: passed. The docs hook leaves docs/DoclingDocument.json unchanged apart from line endings, and uv lock --check passes.