▸ T0 · main thread · write prose.sh post from build log · interactive
Parallel CI clusters aren't complexity overhead. They're the thing that makes your test suite tell the truth.
Here's what happened. P1 and P2 in my x engine build are in the same parallel cluster. Each phase gate runs the full suite in a worktree branched from the integration branch. That means P2's gate worktree has its own conftest.py but does not yet have the WARMUP_ENABLED constant that P1 was supposed to add. In a serial pipeline, you never see this failure: P1 always finishes before P2 runs, so the constant exists by the time the fixture reaches for it.
In parallel, the truth surfaces. The fixture raises AttributeError because monkeypatch.setattr, by default, requires the attribute to already exist on the object. Fix: raising=False. A single argument. Now monkeypatch creates the attribute if absent instead of blowing up, and the fixture is correct whether it runs in P2's isolated worktree or after integration.
The broader point is that "order dependent tests are a code smell" is backwards. Order dependent tests are fine if the order is guaranteed. The problem is when you think the order is guaranteed and it isn't. Parallel gate execution is the forcing function that proves whether your guarantee is real. If your test suite only passes when run in a specific sequence on a single branch, you don't have a test suite. You have a ceremony.
raising=False is the fix. The parallel cluster is what forced you to write it.