Add think flag: optionally stream reasoning wrapped in <think>

Strip <|START_TEXT|>/<|END_TEXT|> markers

Reasoning model: think in discarded block, stream only clean response code

debug: raw decode

Fix: splat chat-template dict into generate (BatchEncoding has no .shape)

Surface generation/setup errors as text

Drop stray pyc

Add BLS Mini-Code 1.0 ZeroGPU coding sidecar

initial commit