Log generation time and token count so the GPU duration can be measured, not guessed

Raise ZeroGPU duration 180->300s: 180 was killing longer completions mid-generation

Match sft_train.ipynb eval config: max_new_tokens 768->2048, GPU duration 120->180

Increase GPU task duration budget to 120s (45s was too tight for 4-bit tensor packing + generation, causing GPU task aborted)

Fix: use sft/seed_42/best (not rl/), correct plan-then-SQL prompt, 4-bit NF4 quantization, and SQL extraction from fenced block, per MODEL_CARD.md

Fix: use PEFT torch_device=cpu (not device_map, which was silently ignored) to force CPU-side adapter load

Fix: force adapter safetensors load onto CPU (device_map=cpu) to avoid ZeroGPU emulation gap

Fix: point PeftModel.from_pretrained at the correct adapter subfolder (rl/seed_42/best)

Add app.py, requirements.txt, schema.sql, README.md to serve the fhirsql model via ZeroGPU