Update vision GGUF variant test to v1 fine-tune defaults

Shrink QA context to fit the T4 alongside other models

Raise Q&A max output tokens to 2048

Delete the legacy trip/payment/document model and instructor

Extract reasoning-strip helpers out of the llama.cpp adapter

Force llama.cpp to CPU whenever SPACES_ZERO_GPU is set

Enable flash attention on GPU to fix NaN logits on the T4

Auto-offload llama.cpp layers when a CUDA build is present

Serialize model inference; cut Q&A cost to shorten lock holds

Drop ZeroGPU: llama.cpp on CPU is the only runtime

Tune MiniCPM5 Q&A: low temperature, 32k context, filtered thinking

Apply ruff format across the repository

Harden OCR parsing with JSON grammar, payer rule, and doc detection

Improve extraction quality with higher resolution, better ordering, and text enrichment

Revert hallucination reduction layers 2-6

Add logit-based confidence flag for vision extraction observability

Add two-pass verification for low-confidence extractions

Re-enable Instructor retries for vision extraction

Fix vision thinking blocks with raw JSON fallback (T9)

Skip local GGUF integration tests unless explicitly enabled

Add Instructor JSON schema vision extraction

Recover trips with ticket-first vision extraction

Ship CUDA runtime libs for llama-cpp on ZeroGPU

Enable CUDA llama-cpp in ZeroGPU vision workers

Remove MiniCPM-V 4.5 backend entirely

Fix ZeroGPU vision timeout with per-image GPU slots

Drop progress callback from ZeroGPU vision batch

Fix ZeroGPU batch duration arg order on upload

Add dual-variant llama.cpp adapter for MiniCPM-V 4.6 and 4.5

Pivot to CPU-only: transformers_cpu adapter, streaming UI, remove legacy backends

Switch to MiniCPM-V-4.6-Thinking GGUF across all adapters (A6 / T11)

Bind batch_process to LlamaCppVisionModel (A1)

Lazy-load LlamaCppBackend: defer model download to first use

Add llama.cpp adapter tests: unit helpers + guarded integration