Unicode \u202f breaks the next pages of template

[fathurwithyou]

Describe the bug

I recently develop auto document generation using gpt-oss:20b model. Unexpectedly, the model gives output like 5\u202fseconds and this results a bug to the template where all next pages disappeared.

Expected behavior

I think we should normalize the problematic unicode before rendering.

Screenshots

[CAOShurong]

I could not reproduce this yet against a wheel built from current master (1f143fbe86c19ecb28c3205d5f4b1547c7e2d7ad).

I created a two-page DOCX template with the value on page 1 and a fixed SECOND_PAGE_MARKER after an explicit page break. I tested both forms that the report might mean:

• an actual U+202F character: "5\u202fseconds"
• the six literal characters: r"5\u202fseconds"

After rendering, both output packages reopened successfully with python-docx; both retained all three paragraphs (value, page-break paragraph, second-page marker), and the marker remained in word/document.xml. The outputs were 36,683 and 36,681 bytes respectively. I did not render them through Microsoft Word, so this does not rule out a layout interaction in the original template.

Could you attach a minimal source template plus the broken rendered output, and show repr(the_value_passed_to_render)? That would distinguish an actual U+202F code point from an escaped string and let us test the specific page/section XML that disappears.

OpenAI Codex assisted with the reproduction script and package inspection; I verified the run and results before posting.