[HenryZhu123]

Description

Native Word equations are stored as structured Office Math (OMML), but the
current DOCX walker collects only their text leaves. This causes expressions
such as 1/2 + 1/3, 2^3, and sqrt(9) to be extracted as 12 + 13, 23,
and 9.

This change makes the shared DOCX extraction path namespace-aware and
recursively renders common OMML structures into unambiguous, model-readable
text:

• fractions: a/b
• superscripts and subscripts: a^b, a_b, and a_b^c
• square roots and indexed roots: sqrt(x) and root(n, x)
• nested expressions: for example, x^2/sqrt(y)

Unsupported structural elements are represented explicitly as
[unsupported equation: ...] instead of being silently flattened into a
potentially incorrect expression.

Because the fix is implemented in the shared DOCX extractor, the preserved
equations reach:

• chat attachments
• the text_only parsing path used for knowledge ingestion
• Immersive Reading

The tests build valid DOCX files containing real OMML nodes rather than
mocking the extracted text.

Related Issues

• Closes [Bug]:Native Word equations lose their structure during DOCX extraction #2144
• Related to fix(docx): preserve native Word (OMML) equation structure in text extraction #2154

Module(s) Affected

• agents
• api
• config
• core
• knowledge
• logging
• services
• tools
• utils
• web (Frontend)
• docs (Documentation)
• scripts
• tests
• Other: ...

Checklist

• This PR targets dev (not main).
• I have read and followed the contribution guidelines.
• My code follows the project's coding standards.
• I have run pre-commit run --all-files and fixed any issues.
• I have added relevant tests for my changes.
• Documentation is not required for this internal extraction fix.
• My changes do not introduce any new security vulnerabilities.

Validation

Focused DOCX, parsing, attachment, and reading tests:

585 passed

Additional Notes

I noticed #2154 after completing this implementation. The two PRs overlap on
the core fraction, script, and radical handling.

This implementation additionally focuses on:

• explicit handling of unsupported OMML structures instead of silently
  flattening them;
• nested equation regression coverage;
• integration coverage for chat attachments, text_only parsing, and
  Immersive Reading.