Nemotron and DOTS exports alter inline LaTeX from model responses #4904

Bug

Nemotron Parse 2.0 and DOTS keep inline LaTeX in ordinary text items. Markdown export escapes underscores, and Nemotron's Markdown-to-text path can interpret math asterisks as emphasis. This changes formulas already present in the model response.

Steps to reproduce

• Parse a Nemotron Text region containing The value is $x_1 * y_2 * z_3$..
• Or parse a DOTS JSON Text element containing **Value** $x_1 * y_2 * z_3$..
• Export Markdown.

Actual: math contains x\_1, y\_2, and z\_3. Expected: inline formula items holding the original LaTeX and an export containing $x_1 * y_2 * z_3$.

The fix should share the math-aware Markdown extension and rich HTML mapping. It must preserve code spans, literal/escaped currency, surrounding content, and the native block's single coordinate owner. This is a focused VLM-response conversion issue related to #2352; it does not address formula recognition in the standard PDF pipeline.

Docling version

Current main d01c9c0991920924cf17f6b8366079ac035ffc38 (2.140.0); reproduced locally on Python 3.12.