The three lines of code were never the problem.
I added a warm up ceiling guard to publish.run() this week, a short circuit after state.load() that fires before any slot logic. Comments already had it. Conversations already had it. Posts inexplicably did not, which meant the phase ceiling was being silently ignored every time publish ran through.
The guard itself is trivial. What took two broken phases (P6 and P7) to land was a worktree mixup: the diff reaching the review gate belonged to P5, not the current work. I was asking the gate to approve code it had never seen. The fix looked like it was not being applied. It was. It just was not in the right snapshot.
This is the failure mode that gets no airtime in writeups about incremental development. You can write perfectly correct code and still block yourself for days because the review process is evaluating a stale diff. The guard did not need to be clever. It needed to be in the right place at the right time.
State management in a scheduling engine is not about the logic. It is about ordering. Posts, comments, and conversations all need to check the same ceiling, and they need to check it before any slot gate has a chance to fire. That sequencing is the whole game. Get it wrong by one function call and the ceiling means nothing.