The question that started this was four words long: is everything running normally.
We went to check, and it was not. On the exchange the resting exit order for one position sat at 1771.77. In our database the same position's exit price read 1693.53. A gap of 4.6% on the single number that position exists for.
Nothing had broken. That is the part worth writing down.
The exit price for that position is recomputed from the exchange's own average every time the order is replaced, and it had been moving for weeks, correctly, on the exchange. The database row was written once, at the moment the position opened, and never again. There was no failing job, no exception, no retry that gave up. There was no code at all. The column had one author, on one day, and after that it was a number sitting still while the thing it described moved.
Our dashboard mirrors the database and computes nothing itself. So the engine was right the whole time and the mirror was stale, which is the reverse of the failure we would have gone looking for.
The correction is a targeted update that moves the exit price alone, and only after the exchange has confirmed the replacement order. Not a general write of the whole row: this position is deliberately detached from the engine's ordinary bookkeeping, and a write that touched its other fields would have invented values for them.
That distinction took longer to get right than the bug did to find. A fix that repairs a stale number by overwriting four accurate ones is not a fix.
The reconciliation is now the sixth block of the morning health check: it reads the live orders from the exchange and compares them against what our database claims. It is the only check we have that tests our data against somebody else's, and it exists because the other five could not have caught this. Every one of them compares our records with our records, and our records were internally consistent. They were consistent and wrong.
For any number shown to a reader, there is a column somewhere and a piece of code that writes it. The useful question is not whether the value looks right. It is who writes this column, and when.
"Once, at creation" is a legitimate answer. It is also the answer that goes stale in silence, because there is no failure to observe, only a number that stopped being true at a moment nobody recorded.