Fix Nano: call @spaces.GPU function directly from Gradio handler

Disable SSR mode — fixes button not working in Gradio 6

Fix GPU task aborted: call @spaces.GPU synchronously, not via to_thread

Fix ZeroGPU: follow documented pattern — model.to('cuda') at module level

Fix Trinity Nano ZeroGPU: load on CPU, explicit CUDA move, no asyncio.to_thread

Remove judge/scoring entirely — show responses and costs only

Remove Tab 2 (Pick the Right Model) — single-page layout

feat: show each model response as it arrives, not waiting for all three

fix: eager import of local_model to prevent ZeroGPU abort

fix: import spaces before gradio to patch torch.cuda correctly

fix: load model to CUDA at module level, increase GPU duration to 120s

fix: proper ZeroGPU pattern — load CPU at startup, move to CUDA on call

fix: pin transformers<5.0 for Trinity Nano compatibility

fix: patch missing pad_token_id in AfmoeConfig for Trinity Nano

fix: import local_model at startup so ZeroGPU detects @spaces.GPU

chore: switch judge to Gemini 2.5 Flash for better scoring calibration

feat: run Trinity Nano locally via ZeroGPU, Mini/Large via OpenRouter

feat: switch all models to HF Inference API, drop OpenRouter

feat: use HF Inference API for Trinity Nano instead of OpenRouter

feat: add streaming progress indicator, fix judge model, add debug logging

fix: add css to Blocks for HF Spaces compat, improve auth error message

fix: move css/theme to launch() for Gradio 6 compatibility

feat: add sample prompts highlighting different model strengths

fix: shorten Space description to meet 60 char limit

chore: add .gitignore

feat: add Gradio UI with two-tab layout and custom dark theme

test: add end-to-end integration tests

feat: add HF Space metadata and README

feat: add comparison engine with best-value selection

feat: add judge system with anonymized shuffled scoring

feat: add IP-based rate limiter with configurable window

feat: add OpenRouter async client with retry logic and cost calculation

feat: add project config and dependencies

docs: add design spec and implementation plan