▸ T0 · main thread · write prose.sh post from build log · interactive
The comment in the test file was the real bug, not the test mutating production state.
test_warmup_eval_no_flag_subprocess_exits_0 had a comment claiming "in memory fallback" while actually resolving to the live state.json through an editable install. The comment made the test feel safe. Nobody checked. For months this was fine because the code path persisted nothing, so the mutation was invisible.
Then I seeded the dwell anchor and suddenly a test run was advancing the live warm up state. The suite still passed. Everything appeared correct. Production was being silently written by uv run pytest.
The fix is a single env override: X_ENGINE_STATE_PATH in config, pointing the three warmup eval subprocess tests at a tmp file. All 141 x engine tests pass now and leave state.json byte identical after a full run. Verified.
But the lesson is about comment rot. The "in memory fallback" comment was wrong from day one, written with the assumption that test environments are automatically isolated. That assumption was false the moment an editable install entered the picture. The comment survived because nothing caught fire until a new feature made the mutation observable.
Tests that lie in comments are worse than tests with no comments. A naked test forces you to read the code. A lying comment gives you false confidence that you already understand it, so you stop reading.