feat(markdown): include table image in Markdown export when available

[maxmilian]

Closes #474.

What

MarkdownTableSerializer only ever emitted the text/HTML table grid — TableItem.image was ignored entirely, unlike MarkdownPictureSerializer, which already honors image_mode (PLACEHOLDER / EMBEDDED / REFERENCED) for pictures.

Change

• Extracted the shared image-emission logic (previously MarkdownPictureSerializer._serialize_image_part) into a module-level function reused by both the picture and table serializers — no behavior change to the existing picture path (verified byte-for-byte identical branches).
• Moved _image_to_base64 from PictureItem up to the shared FloatingItem base so TableItem (which is also a FloatingItem) can use it.
• MarkdownTableSerializer.serialize now appends the table image alongside the existing text table when item.image is set and image_mode is not PLACEHOLDER — the text table is preserved so this is purely additive.

Design choice (open in the issue, unconfirmed by a maintainer)

I'd asked in the issue whether the image should replace or be emitted alongside the text table, and whether it should reuse image_mode or a new dedicated option. I went with alongside + reuse existing image_mode, since:

• it requires no new public API surface, and
• replacing the text table by default would be a breaking output change for existing consumers who already rely on the Markdown table text (e.g. for RAG chunking).

Happy to adjust if a maintainer prefers different behavior.

Backward compatibility

Tables without a generated image (item.image is None — the common case, since table images require the generate_table_images pipeline option) see zero output change, even if image_mode is explicitly set to something other than PLACEHOLDER. Covered by a dedicated regression test.

Testing

Added 3 new tests in test/test_serialization.py:

• test_md_table_with_image_embedded — EMBEDDED mode, image present
• test_md_table_with_image_referenced — REFERENCED mode, image present (asserts exact output)
• test_md_table_without_image_unaffected_by_image_mode — no image present, asserts byte-exact no-op regardless of image_mode

Full suite: uv run pytest -q → 545 passed, 6 skipped (unrelated to this change).
ruff check / ruff format --check / mypy docling_core test all clean.

[github-actions]

✅ DCO Check Passed

Thanks @maxmilian, all your commits are properly signed off. 🎉

[mergify]

Merge Protections

🔴 1 of 2 protections blocking · waiting on 👀 reviews

	Protection	Waiting on
🔴	Require two reviewer for test updates	👀 reviews
🟢	Enforce conventional commit	—

🔴 Require two reviewer for test updates

Waiting for

• #approved-reviews-by >= 2

This rule is failing.

When test data is updated, we require two reviewers

• #approved-reviews-by >= 2

---

Show 1 satisfied protection

🟢 Enforce conventional commit

Make sure that we follow https://www.conventionalcommits.org/en/v1.0.0/

• title ~= ^(fix|feat|docs|style|refactor|perf|test|build|ci|chore|revert)(?:\(.+\))?(!)?:

[maxmilian]

Rebased and force-pushed to resolve the conflict. It turned out to be more than a textual conflict: #664 split the monolithic document.py into focused modules after this branch was opened, so the original diff no longer applied. I've reworked the change onto the new layout:

• _image_to_base64 moved from PictureItem up to the shared FloatingItem base — now in docling_core/types/doc/items/node.py (dropping the now-unused base64/BytesIO imports from items/picture/picture.py).
• _serialize_image_part extracted from MarkdownPictureSerializer into a module-level helper in transforms/serializer/markdown.py, called by both the picture and table serializers.
• MarkdownTableSerializer now emits the table image alongside the text grid when item.image is set and image_mode is not PLACEHOLDER; tables without an image stay a strict no-op.

The behavior and public surface are unchanged from the original PR. Verified on the new base: test_serialization.py (50) plus the doclang/doc/latex serialization suites (220 total) pass, and the three new table-image tests (embedded / referenced / no-op) go red without the change. ruff check, ruff format, and mypy are clean on the touched files.