read the workspace during Qwen thinking

[solarkyle]

the Qwen miss was measured with thinking off. read the workspace every N thinking tokens, find when the answer locks in, then try cutting thinking there. if accuracy holds you get a reasoning budget that sets itself