Surface Docling structure and document processing journey

[farhat-is-coding]

What changed

• Add a Structure view for parsed document elements.
• Show pages, sections, headings, tables, pictures, captions, classifications, and machine-generated visual descriptions.
• Add element-type filtering so reviewers can isolate images, tables, headings, or text.
• Replace the generated-style visual-description icon with the standard image treatment.
• Add a processing journey view covering Docling parsing, visual enrichment, embedding generation, and vector indexing.
• Show retry attempts, stage timing, stage-specific errors, parser metrics, chunk metrics, and vector-storage metrics.
• Distinguish visual operations from fresh Gemini requests and cache hits.
• Add the typed frontend contracts and API integration for structure and journey responses.

Why

Users and reviewers previously could not see what Docling extracted, what Gemini enriched, where processing time was spent, or which layer caused a failure. The original "vision calls" label also made cached work look like fresh provider traffic.

This UI makes the document's parsed structure and the complete processing path inspectable without exposing implementation noise.

Impact

The file viewer now provides a clear client-facing explanation of document structure and processing. Historical traces label the old aggregate as "Visual operations"; new traces additionally show fresh Gemini requests and cache hits.

Internal-reference resolution and the user-approved large scanned-document OCR workflow are intentionally out of scope for follow-up PRs.

Validation

• ESLint passed for the touched components and types.
• TypeScript compilation passed.
• Full Vite application build passed.
• Documentation site production build passed.
• Verified against real backend processing traces, cached visual enrichment, and a local Weaviate index.