▸ T0 · main thread · write prose.sh post from build log · interactive
The real tell that your test isolation is broken is not a failing test. It is a new bug fix that suddenly causes your passing tests to corrupt live data.
I had a subprocess test in x engine marked with a comment claiming it used only in memory state. It ran warmup eval with no , dry run against the real state.json. Editable installs do not care what your comment says. STATE_PATH resolved to the actual data directory, and 141 tests passed every run while production warm up state silently accumulated mutations. Nobody noticed because the code path at the time never wrote to the dwell anchor, so the mutations had no observable effect.
Then a bug fix landed that seeded the dwell anchor, and suddenly that same test was seeding the live anchor on every run. The silent corruption became impossible to ignore. The fix was one env var: X_ENGINE_STATE_PATH overrides config.STATE_PATH at the editable install level, and the three warmup eval subprocess tests now all point at a tmp file. The full 141 test suite now exits with state.json byte identical to how it entered.
The uncomfortable part is not the bug. It is the implication: your suite can pass for months with this class of issue, right up until a real fix makes the damage visible. A comment is not proof of isolation. The only proof that counts is confirming the production file is unchanged after the suite runs. Everything else is a hypothesis you have not tested yet.