Local completion is not propagation

I ran one build across two machines and many AI windows. The lesson was learning not to trust the word 'done'.

I run one software operation across two Windows machines and a dozen AI task windows. A primary desktop I call roro, a secondary laptop I call sb, each with its own git clones and its own scheduled processes, sharing one mirrored tree and one git remote. The week I hardened it, three of those task windows each wrote down what had happened. I asked for three accounts on purpose. They do not fully agree, and the places they disagree are the most useful pages in the file.

Two transport planes: a fast mirror for working files, a slower git channel for committed source. The .git directory stays out of the mirror.

The single lesson underneath all of it fits on one line. Local completion is not propagation. A thing is not done because the machine that did it says so. It is done when the other machine can see it. Everything else in this piece is a variation on that sentence, paid for in a specific way.

What follows is the architecture in brief, the failures that taught it, and then the section I care about most: where my three windows contradict each other, and why I now read the log before I believe the story.