Add LiteParse engine adapter and comprehensive benchmarking suite

[Mihailorama]

Summary

This PR introduces a new LiteParse engine adapter for fast local document parsing via CLI, along with a comprehensive benchmarking framework to compare document processing engines.

Key Changes

LiteParse Engine Adapter (src/docfold/engines/liteparse_engine.py)

• New LiteParseEngine class that wraps the LiteParse CLI (lit parse) as a subprocess
• Supports multiple document formats: PDF, DOCX, PPTX, XLSX, images (PNG, JPG, etc.)
• Configurable options: OCR enable/disable, OCR language, DPI, worker threads, max pages
• Extracts bounding boxes from LiteParse JSON output with page coordinates
• Handles multiple output formats: TEXT, MARKDOWN, JSON, HTML
• Robust JSON parsing that handles log lines prefixed before JSON output

Benchmarking Suite (benchmark.py)

• Generates synthetic test PDFs with known ground truth content (invoices, reports, financial docs)
• Runs both PyMuPDF and LiteParse engines on identical documents
• Measures performance metrics:
  • Processing time (milliseconds)
  • Character Error Rate (CER) via Levenshtein distance
  • Word Error Rate (WER)
  • Bounding box coverage
• Produces detailed JSON report (docs/benchmark_results.json) with per-document and aggregate statistics
• Includes sample benchmark results showing PyMuPDF (~4.5ms avg) vs LiteParse (~382ms avg)

Test Coverage (tests/engines/test_liteparse_engine.py)

• 300+ lines of comprehensive unit tests
• Tests metadata (name, capabilities, availability detection)
• Tests all output formats (TEXT, MARKDOWN, JSON)
• Tests CLI flag generation (OCR, DPI, workers)
• Tests bounding box extraction from textItems
• Tests error handling and JSON parsing with log prefixes
• Uses mocking to avoid Node.js dependency in test environment

Integration Updates

• Registered LiteParseEngine in CLI router (src/docfold/cli.py)
• Added LiteParse to extension priority routing (src/docfold/engines/router.py)
• Updated benchmarks documentation with LiteParse comparison table
• Added optional dependency group in pyproject.toml (no Python deps, requires Node.js 18+)

Notable Implementation Details

• LiteParse runs as a subprocess with async I/O for non-blocking execution
• Bounding boxes are computed from textItems: [x, y, x+width, y+height]
• JSON extraction handles CLI log output by finding the first { character
• Benchmark uses text normalization (whitespace collapsing) for fair CER/WER comparison
• All async operations properly measure elapsed time using time.perf_counter()

https://claude.ai/code/session_017D1jcELbNYGtk9QoUX6FA7