[Gal-Gilor]

What changed

• MarkdownSplitter is a frozen, keyword-only dataclass with one field, max_level (1-6, default 6). MarkdownSplitter(3) raises TypeError, and setting max_level after creation raises FrozenInstanceError.
• split_text and split_file are methods. split_text builds the sections in one pass with a stack of the enclosing sections.
• Removed get_document_outline, OutlineNode and _create_sections_from_outline. The package exports MarkdownSplitter, Section and SectionMetadata.
• Tests use a module-level splitter = MarkdownSplitter(). Removed the outline-only tests; the section tests cover the same behavior.
• Updated the README, the CHANGELOG 0.3.0 entry (compared with 0.2.0) and the CI wheel check.

Why

• One class with immutable options. One instance can be reused safely.
• One representation of the tree: sections with parent_id, instead of both sections and an outline.
• One pass over the headers, instead of building a tree and then flattening it.

Verification

• Equivalence with main: 508 distinct inputs (every text main's test run passed to the splitter, every string literal in main's tests/test_splitter.py and conftest.py, README.md and CHANGELOG.md, the README examples, and 194 generated documents with preamble, BOM, CRLF, front matter and code fences). Each was compared with the default and with max_level 1-6, 3556 comparisons in all. [s.model_dump() for s in ...] was equal in every case.
• ruff check and ruff format --check pass. 94 tests pass, splitter.py coverage is 98%, the same as main.
• Doctests pass. The built wheel passes the CI import check. The tests pass with pydantic 2.0.2 on Python 3.12.
• The README Python examples run and give the outputs shown.