Working notes on the engineering problems behind MBE3Dstudio, written from what we have
actually built and hit, including the parts that do not work yet.
We repaired a specification, recorded four numeric conflicts as resolved, and were wrong about three of them. The requirements file was correct. The file the rule engine actually reads still held the old numbers.
A converter that correctly refuses to invent provenance becomes a data-degradation engine the moment it runs over data that already has some. We shipped that, and measured it before we found it.
A configuration flag named after a behaviour is a claim about that behaviour, not the behaviour itself. When isolation is a filter rather than a wall, the only evidence worth having is a write one tenant cannot see.
A container told us it was running one commit. Its image label said another. The difference was ten commits of work, and the only thing that caught it was a script that refused to believe either of us.
Exact distance between two oriented boxes is not free, and the cheap approximation is wrong in one specific direction. That direction turns out to be the one you want.
The drop on a feed is decided by where the wire actually goes, not by how far apart its endpoints are. One routing choice failed two separate requirements in the worked example.