fix(docx): disable numeric coercion of Word text runs

[eliziff]

Summary

Prevent numeric-looking Word text runs from being altered when Mike parses or rewrites DOCX files.

Why / Motivation

fast-xml-parser converts element text to JavaScript numbers and booleans by
default. Word stores visible document text in w:t, so a run containing
12.10 becomes 12.1.

This affects text returned by read_document and can rewrite an untouched run
when Mike saves an unrelated tracked edit.

Changes

• Disable tag-value coercion in the shared DOCX parser.
• Add regression coverage for extracted text and the rewritten
  word/document.xml.

Tradeoffs & risks

Element text in word/document.xml is now retained as strings, matching how
the DOCX helper consumes it. Attribute parsing, tracked-change IDs, entity
handling, and whitespace behavior are unchanged.

How verified

• Confirmed the regression test fails on current main, where the rewritten
  w:t contains 12.1.
• Confirmed the patched output retains 12.10 through an unrelated tracked
  edit.
• npm test --prefix backend — 551 passed, 23 skipped.
• npm run build --prefix backend.

Checklist

• Ran the relevant build and tests.
• Reviewed the diff and removed unrelated changes.
• Docs and environment examples are not applicable.
• No secrets, API keys, real documents, or .env files committed.