PDF paragraph split into two texts items mid-sentence (leftover last line)

### Bug
A single PDF paragraph is exported as two texts items, cut mid-sentence. The last line of the paragraph becomes its own label: "text" item.

This is a layout grouping bug, not a request for sentence splitting. I know texts is layout boxes. Here Docling draws two boxes for one paragraph: a large block plus a leftover last line. Downstream tools then see a broken sentence.

Example 1 (page 1)

texts[5] ends with no period:
… Ein bedeutender Meilenstein war die Mondlandung im Jahr 1969. Die Apollo Mission brachte

texts[6] is only the leftover line (starts lowercase):
erstmals Menschen auf die Oberfläche des Mondes.

Same page. Boxes sit one under the other (BOTTOMLEFT):
texts[5] bbox t≈441.43, b≈296.20 (tall)
texts[6] bbox t≈293.63, b≈280.35 (thin strip).
orig equals text on both (born-digital PDF, not OCR).

Example 2 (same document, page 1)

texts[8] ends: … ausgestoßen wird entsteht eine Kraft die die
texts[9]: Rakete nach vorne beschleunigt.

Expected: one texts item for the whole paragraph, ending at Mondes. / beschleunigt.

Actual: two items; the second is a last-line fragment that continues the first.

Document JSON: schema_name: DoclingDocument, version: 1.10.0. Conversion via docling-serve (POST /v1/convert/file/async, from_formats=pdf, to_formats=json only; no custom pdf_backend). Image: quay.io/docling-project/docling-serve-cpu:latest.
...

### Steps to reproduce
1. Convert the PDF with default PDF backend (json output), e.g. docling-serve:

curl -X POST "http://127.0.0.1:5001/v1/convert/file" \
-F "files=@pleasewritesomelineslikeabove.pdf;type=application/pdf" \
-F "from_formats=pdf" \
-F "to_formats=json"

(Async is the same: /v1/convert/file/async, then poll /v1/status/poll/{task_id}, then GET /v1/result/{task_id}.)

2. Open document.json_content.texts.

3. Compare item 5 and 6 (and 8 / 9). First item does not end with .!?; next item starts with a lowercase letter and is the last line of the same paragraph.

### Docling version

Docling version: 2.115.0
Docling Core version: 2.87.1
Docling IBM Models version: 3.13.3
Docling Parse version: 7.8.1
Platform: Linux-5.15.167.4-microsoft-standard-WSL2-x86_64-with-glibc2.34

### Python version
Python: cpython-312 (3.12.13)

Start by reproducing the report with the supplied PDF and default PDF conversion, then inspect document.json_content.texts around items 5/6 and 8/9. Trace the layout grouping path that produces these text boxes; done means each leftover lowercase final line is grouped with its preceding paragraph without breaking unrelated text items.

Written by the indexing model from the issue text.