The quiz-pool work passed everything measurable. The migration landed, the bilingual audit came back clean across 66 published lessons with at least five valid questions in each language, and the full suite went green: 633 tests. Then it hit the one gate that couldn't be automated, watching a real learner get an answer wrong and press Next Check.
The bot-check at sign-in refused to cooperate. It rendered no token for the test browser, in headless mode and headed mode, on the local branch and against production. Four attempts, four identical refusals. The task was handed back with the landing deliberately unclaimed: the backend is proven, the frontend retry is committed, and the journey in between is untested, so it doesn't ship. No "we'll verify later." The disposable test users were cleaned up, and the queue stayed honest.
Elsewhere, two tasks tried to occupy the same migration version, and the ledger noticed. History stayed untouched; the reconciliation went forward as a new migration, followed by a fresh verification pass. By evening the mastery-question rotation was live in production, and retries now cycle through a learner's actual attempt history instead of serving the same questions again. Close call, boring ending. The best kind.
And the Practice Market trailer got rebuilt around one idea. Seven scenes over 45 seconds became four beats over 22: practise a decision before real money is involved, watch prices open under pressure, pause and write down what you see, then choose to wait. The browser suite passed all 22 checks, reduced motion becomes a manual stepper, and with JavaScript switched off the whole story still reads as a static page. Cutting six scenes hurt less than expected. The trailer had been trying to say seven things. The learner only needed one.
The probes ran their second daily cycle too: twenty more observations, three failures, queues still open. Not clean, and said so. Same instinct as the trailer, really. Say one true thing, then stop.