[dolfim-ibm]

Follow-up to #4897. The mapper reads model output, which document content can steer, and which a remote server sends when the API engine is used. Three inputs used too much time or memory, or changed the table structure.

Changes

• Span size (nemotron_parse_v3_5_utils._span): \multicolumn and \multirow spans are capped at _MAX_TABLE_GRID_CELLS + 1 before the span grid is built. Before this fix, \multicolumn{10000000}{c}{x} (about 100 characters) took 13 s and 1.2 GB, and larger values exhaust memory. A table above the limit becomes an empty table, as before.
• Backtick runs (dots_utils._MARKDOWN_ESCAPE_PATTERN): the regex now reads the opening backtick run of a code span atomically ((?=(+))\1`, Python 3.10-compatible) and only at the start of a run. Before this fix, an unclosed run took cubic time: 2000 backticks took 15 s in both the Nemotron 3.5 and dots mappers. Now 20000 backticks take about 1 ms.
• Cell markup (latex_tabular_to_html): cell text is HTML-escaped before the table HTML is built. Only inline formatting tags (b, br, em, i, strong, sub, sup, u) stay as markup. Before this fix, a cell such as a</td></tr><tr><td colspan="9">b added a row and a span to the table.

Checklist:

• Documentation has been updated, if necessary.
• Examples have been added, if necessary.
• Tests have been added, if necessary.

Review evidence:

• New tests: test_table_with_oversized_span_is_rejected_without_building_it, test_table_cell_markup_stays_cell_text, and two test_literal_markers_stay_literal cases (a nested code span and a 5000-backtick unclosed run).
• test_table_cell_markup_stays_cell_text fails on main ((2, 10) != (1, 2)). I did not run the span test on main, because it exhausts memory there.
• tests/test_nemotron_parse_v3_5.py, tests/test_dots_parser.py, tests/test_chandra_vlm.py: 107 passed. make validate passes.

Limits (not changed here):

• _LITERAL_INLINE_PATTERN is still quadratic on many unclosed \(: 20000 of them take about 3 s. This pattern was there before feat: add nvidia/Nemotron-3.5-Parse as supported VLM model #4897.
• The shared inline HTML builder keeps any <a href> scheme, including javascript:, and export_to_html() writes it unchanged. This was there before feat: add nvidia/Nemotron-3.5-Parse as supported VLM model #4897; a scheme allowlist belongs in a separate fix.

[github-actions]

✅ DCO Check Passed

Thanks @dolfim-ibm, all your commits are properly signed off. 🎉

[mergify]

Merge Protections

🟢 Merge protection satisfied — ready to merge.

Show 1 satisfied protection

🟢 Enforce conventional commit

Make sure that we follow https://www.conventionalcommits.org/en/v1.0.0/

• title ~= ^(fix|feat|docs|style|refactor|perf|test|build|ci|chore|revert)(?:\(.+\))?(!)?:

[github-actions]

AI triage (advisory)

Checked commit ab7c550e73. The analysis read the diff only. It did not run code from this pull request.

Warning

The AI analysis did not finish. Only the deterministic checks are shown.

CI safety: not assessed.

Possibly related pull requests

PR	State	Signal	Assessment
#4908	open	changes the same code	not assessed
#4897	merged	referenced as #4897	not assessed

_{The ai:* labels are advisory. They do not approve CI and they do not add tests:full. A new push runs this check again.}

[github-actions]

AI first review (advisory): no blocker found

This review covers the full diff at 47d058e974. It read the code and did not run it.

I found no defect that should stop the merge. I read the span cap, the cell escaping, and the atomic backtick regex against the base helpers _add_html_table and _markdown_inline_html, and they hold. The span cap of 1001 always trips the existing grid check in docling/utils/chandra_utils.py. A raw & always splits the cell first, so the new escaping cannot double-escape an entity.

_{A maintainer makes the merge decision. This review never approves the pull request.}

[codecov]

Codecov Report

❌ Patch coverage is 90.90909% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
docling/utils/nemotron_parse_v3_5_utils.py	90.90%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

[cau-git]

I pushed ab7c550 with three fixes on top of this PR:

• Escaped angle brackets in table cells: \textless{}b\textgreater{}x became bold text x. The < and > were changed before the escape step, so they became a real <b> tag. Now the cell text is <b>x</b>, as in the source.
• <code> in table cells: this tag now stays markup. Before this fix, the cell showed the literal text <code>z</code>.
• Many wide spans in one row: the span limit applied to each cell, but not to the full row. 2000 cells with a span of 1000 (54 KB of input) used 511 MB. Now a row above 1000 columns stops the table conversion early, and the table becomes empty, as before.

One limit remains: in a cell that also has *emphasis*, < in a code span shows as <. I did not change this.