[PierreMesure]

Chandra <h3> and Nemotron ### section headings exported as ####. Convert source depths to Docling's title-relative section levels so exports keep their original depth and coordinates.

The HTML change is in the shared Chandra mapper, including _add_html_fragment callers. The related LightOnOCR PR #4789 now uses this correction and removes its compensating pass; the shared mapper change is identical in both branches.

Resolves #4902.

Checklist:

• Documentation assessed: output correction; no API change.
• Examples are included in regression tests.
• Tests added for native envelopes, shared HTML fragments, formatted headings, and provenance.

Review evidence:

• All 10 new regression cases failed before the fix.
• .venv/bin/python -m pytest -q tests/test_vlm_heading_exports.py tests/test_chandra_vlm.py tests/test_nemotron_parse_v2.py tests/test_dots_parser.py: 93 passed on Python 3.12.
• UV_NO_SYNC=1 make validate: all applicable hooks passed.
• Existing assertions that stored HTML/Markdown depth directly now check the correct section level. No reference fixtures were regenerated.
• Full Linux CI and model inference were not run locally.

[github-actions]

✅ DCO Check Passed

Thanks @PierreMesure, 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 155d5ffd42. The analysis read the diff only. It did not run code from this pull request.

The PR corrects VLM section heading depth: both chandra_utils.py and nemotron_parse_utils.py now subtract 1 from the raw HTML/Markdown heading level so that Docling's title-relative level system exports the correct depth. A new parametrized test file verifies the fix across both parsers.

Topic labels added: bug, vlm-pipeline.

CI safety: no concern found. A maintainer must still approve CI.

Possible duplicates

PR	State	Signal	Assessment
#4789	open	referenced as #4789; changes the same code	duplicate: PR #4789 makes an identical change to chandra_utils.py and test_chandra_vlm.py — same lines removed and added, same comment — so only one of these two PRs can be merged.

2 related pull request(s) in the same code

PR	State	Signal	Assessment
#4907	open	changes the same code	related: PR #4907 modifies nemotron_parse_utils.py and test_nemotron_parse_v2.py but fixes a different problem (stripping math delimiters from formula text), not the heading-level offset.
#4908	open	changes the same code	related: PR #4908 modifies the same Section-header block in nemotron_parse_utils.py but adds inline-math preservation, not a heading-level correction — a different fix in the same area.

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

[PierreMesure]

The AI code review did not produce a verdict for head 155d5ff. The review job hit the 40-turn limit and returned no valid JSON. The publishing job then could not find ai-review-result. Both jobs show success because these steps allow errors.

The triage duplicate warning needs context: the shared Chandra correction also appears in #4789 by design. This PR also fixes Nemotron headings and adds focused export tests. Both PRs can merge after reconciling the identical shared change; they are not mutually exclusive. The four focused parser fixes were tested together again after the #4908 review fix: 129 passed, 1 skipped.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[PierreMesure]

Closing as superseded by merged PR #4964, which includes the original signed-off commit and the heading-depth correction. The fix is now on main, and issue #4902 is closed. The remaining inline-math and LightOnOCR work continues in #4908 and #4789.