Enforce strict GPU startup + runtime checks

Fix startup indentation and keep strict CUDA policy in sync

Enforce strict CUDA path for local OCR inference

Harden image/model input parsing for stable Space runs

Fix comparison runtime robustness for zero-gpu and API route

Hardened precheck and runtime recovery

Fix local transformer bootstrap and dependency baseline

Patch: fix dependency/runtime fallbacks for OCR models

Add PaddleOCR + DeepSeek OCR models and raise default output token budget

Scope GPU lease to active inference runs

Remove persistent GPU warmup; rely on sleep timeout for idle GPU control

Treat tesseract as low VRAM workload in scheduler

ZeroGPU compute-mode scheduling + safety caps