feat(documents): add Section.markdown() accessor

[HonzaCuhel]

Summary

Section.text() flattens tables and bullet lists; Filing.markdown() preserves structure but is whole-document. This adds Section.markdown() — the missing per-section markdown accessor — so per-item chunkers / RAG pipelines can get item-aware and structure-preserving markdown in one call.

Why

Real-world example: tenq[\"Part II, Item 2\"] (Issuer Purchases of Equity Securities table) returns vertically-stacked cell text. filing.markdown() would render the same content as pipe-delimited markdown — but you lose the per-item slice. Today downstream consumers have to choose between item-awareness and structure preservation.

section.markdown() removes the trade-off: one call returns markdown scoped to one section.

Implementation

Heading/pattern-based sections — render the cached node tree via MarkdownRenderer().render_node(self.node). Tables emit pipe syntax, bullet lists keep markers. Same boundary-artifact cleanup that text() applies runs on the rendered output, extended to handle markdown-decorated bleed-in artifacts (# Item 5, **Item 5\\.**, # PART II\\n\\n# Item 5).

TOC-based sections (where node.children is empty and content is fetched lazily from _html_source) — fall back to Section.text(). Iterative review surfaced multiple correctness landmines in TOC-section HTML extraction: next-section heading leaks via nested anchors, shared-wrapper LCA computation, same-anchor start/end boundaries, inline anchor wrappers, last-section-in-wrapper leaks, and table-row-bounded anchors losing <table>/<tbody> wrappers. Rather than ship a partial fix for any one, the TOC path is conservative: no regression vs text(), no new landmines. Tracked as a follow-up.

Verification

• 13 new unit tests in tests/test_section_markdown.py cover: table-pipe preservation, bullet-list preservation, TOC fall-back contract, boundary-artifact cleanup across plain/escaped/heading-decorated/bold-decorated/combined-PART+Item variants, idempotency, and a negative control asserting text() does NOT preserve pipes (so the contrast is meaningful).
• 100/100 broader section-detection regression tests still pass with EDGAR_IDENTITY set (test_documents, test_10q_section_detection, test_sections_membership, test_section_detection_edge_cases, test_section_tables_toc_fix).
• 11 rounds of codex review — each round caught a real edge case, all addressed except the documented TOC fall-back (which would re-open the whole TOC extraction class of issues).

Test plan

• CI runs the full repo test suite
• Maintainer confirms the API name + scope
• If maintainer wants TOC structural rendering shipped together, can split into a follow-up PR with the LCA-walk approach (incomplete vs. table-row anchors but covers the more common cases)

🤖 Generated with Claude Code