GPU reservation 60 s (anonymous ZeroGPU quota)

canonical oracles only, k=9, no layer field, system font, planning examples with last token pre-selected

cards: model first, oracle second, no text row, default font

UI: result cards with token strips, short oracle labels, settings bar, re-run on setting change

probe API endpoint

load adapter weights on CPU (torch_device); log attach tracebacks

assemble on CPU, move to CUDA once (ZeroGPU); serve -Base checkpoints only

Future Oracle explorer (ZeroGPU)

initial commit