feat: Add DocumentProfile for document-type-specific extraction

[rookslog]

Summary

• Implements document profiles system for configuring CascadingExtractor based on document type
• Adds 5 standard profiles: BOOK, ARTICLE, ESSAY, REPORT, DEFAULT (generic)
• Adds auto-detection via get_profile() using existing estimate_document_type()
• CascadingExtractor now accepts optional profile parameter
• Backward compatible: existing code without profiles continues to work

Changes

• Add DocumentProfile frozen dataclass with extraction settings
• Add for_profile() and for_document() classmethods to CascadingExtractor
• Add profile_used field to StructureResult for tracking
• Update module exports in extractors/__init__.py
• Update ROADMAP.md to mark milestones 1.4/1.5 complete

Test plan

• 19 unit tests for DocumentProfile dataclass and get_profile()
• 9 tests for CascadingExtractor profile integration
• 3 integration tests with real philosophy PDFs
• All 288 tests pass

[rookslog]

Code Review

Status: ✅ Approved

Strengths

• Clean Design: DocumentProfile as a frozen dataclass is the right choice - immutable, type-safe, simple to use
• Backward Compatible: Existing code without profiles continues to work unchanged
• Good Test Coverage: 31 new tests covering all standard profiles, auto-detection, immutability, and real PDF integration
• Follows Project Patterns: Uses existing estimate_document_type() instead of reinventing
• Proper Exports: All new symbols exported in extractors/__init__.py with __all__

Minor Observations (Non-blocking)

1. pytest.mark.slow warning: Two tests use unregistered mark at tests/integration/test_convert.py:283, 301. Can fix later by adding to pyproject.toml:

   [tool.pytest.ini_options]
   markers = ["slow: marks tests as slow"]

2. min_content validator: ESSAY_PROFILE references it but MinimumContentValidator may not be fully implemented yet. Code handles this gracefully (skips unknown validators).

Verification

• ✅ All 288 tests pass
• ✅ Lint passes
• ✅ CI passes (Python 3.11/3.12)

LGTM - ready to merge.