The test test_warmup_eval_no_flag_subprocess_exits_0 was running the real warmup eval command, no , dry run , against the live production state file. It had a comment claiming "in memory fallback." That comment was wrong. The editable install ( pip install e ) resolves STATE_PATH at import time to wherever the real state.json lives. There is no separate test context, no isolated state, no in memory anything. Every time that test ran, it loaded the live warm up anchor, evaluated it, a...