refactor(extract): Modularize extract package for testability

[prosdev]

Context

The packages/extract/src/index.ts is a 386-line monolith containing:

• Zod schemas
• MIME type detection
• PDF to image conversion
• OCR processing
• Gemini provider
• Ollama provider
• Streaming logic

This makes unit testing difficult - we cannot test OCR, PDF conversion, or providers in isolation.

Architecture Decisions

• Package: packages/extract
• Pattern: Module decomposition with barrel export
• Goal: Each module testable independently

Proposed Structure

packages/extract/src/
├── index.ts              # Barrel export only
├── schemas.ts            # Zod schemas
├── mime.ts               # getMimeType()
├── pdf.ts                # pdfToImages()
├── ocr.ts                # ocrImages() + types
├── extract.ts            # extractDocument() orchestrator
├── providers/
│   ├── gemini.ts         # extractWithGemini()
│   └── ollama.ts         # extractWithOllama()
└── types.ts              # StreamChunk, StreamCallback, ExtractOptions

Requirements

• Extract Zod schemas to schemas.ts
• Extract MIME detection to mime.ts
• Extract PDF conversion to pdf.ts
• Extract OCR logic to ocr.ts
• Extract Gemini provider to providers/gemini.ts
• Extract Ollama provider to providers/ollama.ts
• Create shared types in types.ts
• Create barrel export in index.ts
• Update tests to use new module structure
• Maintain 100% backward compatibility

Success Criteria

• All existing tests pass
• Each module can be imported/tested independently
• No breaking changes to public API
• Coverage maintained or improved

References

• Related: Epic Epic: Core Ingestion Pipeline + Local Storage #1 (completed)

[prosdev]

Closed by #13