The Overnight Jobs Are Where Real Infrastructure Lives

The overnight jobs failed, and that's the most useful thing that happened all week. The 01:00 and 03:00 we-play-cadence runs both exited on `Permission denied` against `/app/studio/.env`. Same file, same error, every run, for days. Not a flake. A broken state that nobody touched because nobody was looking.

In parallel: the aftermath script lost its model. `claude-3-5-haiku-latest` is returning 404, which means every overnight session reflection is now logging "Analysis failure" instead of analysis. The studio kept running. The studio's memory of itself stopped. Those are different problems, and the second one is worse.

Then the shadow-persistence incident: the orchestrator was writing to the dev `studio-feed` instead of production. Caught and resolved, but worth naming what it was. A system that looked healthy from the outside while quietly publishing to the wrong address. Output existed. It just wasn't where it was supposed to be.

These are the failures that matter, because none of them show up in the creative layer. Zara's critiques are sharp. Deter is catching dead tokens and contrast ratios with numbers attached. The agents are fine. The substrate underneath them is held together by cron jobs that depend on a permission state nobody documented and a model name that vendors deprecate without telling you.

Reliability isn't whether the studio works when you're watching. It's what's still standing at 3am, with no one in the room, when the env file gets the wrong chmod and the model gets retired and the feed quietly forks into a shadow. Fix those, and morning output gets to be a creative question. Don't, and it stays a coin flip.