fix: upgrade torch to 2.5.1+cpu for surya-ocr>=2.5.1 compatibility

Merge remote-tracking branch 'origin/main'

fix: upgrade opencv to >=4.11.0.86 for surya-ocr compatibility

Update main.py

Merge remote-tracking branch 'origin/main'

chore: force rebuild — Surya OCR engine (2026-03-21)

feat: Surya OCR engine + /predict_surya endpoint (Surya -> LayoutXLM NER)

chore: force Docker rebuild — fix (2) filenames loading from HF Hub

fix: download HF Hub model files with (2) names to local cache with standard names

Update src/ocr/doctr_engine.py

chore: bust Docker cache — force full rebuild with OCR merge pipeline

chore: force Docker rebuild — OCR merge pipeline v2

chore: trigger Space rebuild

feat: merge OCR words into line blocks before LayoutXLM

fix: update type annotation AutoModelForTokenClassification -> LayoutLMv2ForTokenClassification

fix: detectron2 build with --no-build-isolation (needs torch in scope)

fix: build detectron2 from source, use LayoutLMv2ForTokenClassification, transformers 4.44.0

fix: install detectron2 CPU wheel for LayoutXLM/LayoutLMv2

fix: remove LanguageDetector import from preprocessing __init__ (not needed in backend)

fix: use AutoModelForTokenClassification instead of LayoutXLMForTokenClassification (import issue)

fix: move --extra-index-url to top, remove pyvips (not needed in Docker)

fix: remove docTR preload from build (lazy load at startup instead)

fix: libgl1 (Bookworm compat), libvips-dev only, fix docTR preload env

fix: remove COPY model/ (model loaded from HF Hub), fix docTR cache path

chore: add HF Spaces README config

feat: backend FastAPI OCR+LayoutXLM — model loaded from HF Hub