feat(ocr): release reviewed multimodal knowledge v0.3.1

[Rcloudso]

Product impact

• Adds reviewed OCR ingestion for PNG, JPEG, WebP, and scan-only PDF sources.
• Keeps unreviewed visual content outside retrieval until an administrator publishes the complete draft.
• Preserves page, Block, extraction-job, engine, and version provenance in retrieval and customer source cards.
• Supports optional DeepSeek-OCR-2 shadow comparison without allowing it to overwrite authoritative PaddleOCR output.

Implementation

• Adds durable SQLite extraction jobs, interrupted-job recovery, retries, immutable results, and revisioned review drafts.
• Adds bilingual responsive Block review, optimistic revision checks, low-quality fail-closed gates, and atomic whole-document publication.
• Adds the bounded authenticated PaddleOCR PP-StructureV3 HTTP Worker and pins PaddleOCR/PaddleX to the verified 3.0.3 pair.
• Adds scan-PDF routing, table structure preservation, extraction history, release evidence, and OCR Worker tests in CI.

Verification

• Worker mapper/request tests: 6/6
• Real PaddleOCR CPU HTTP extraction: HTTP 200, 16 Blocks, one page, no warnings
• Full regression and TypeScript suite: passed
• FAQ: Top1/Top3/no-match 100%
• Document: Top3 100%, MRR 0.958
• Mixed: 6/6 Top1
• Quality: decision/recall/MRR 100%, zero unsafe answers or failures
• OCR: CER 1.33%, structure/page/table/low-quality checks 100%
• Playwright API/Web: 46/46
• Production build and diff checks: passed

Remaining risk

Docker is unavailable in the local verification environment, so the container image was not built. The equivalent Python 3.10 dependency stack, real Paddle model initialization, and authenticated HTTP extraction were verified natively.