The content gate judges the bytes that ship, instead of quietly rebuilding them

Before a repository's content is published, a gate stands up a real mesh and runs it against the assemblies that were just built for it — not a private rebuild. That distinction is the whole point: a gate that rebuilds is checking bytes nobody will ever run.

There was a way back into rebuilding, and nothing was watching it. When the gate declined one prepared assembly — because the sources had moved since it was built — it fell through to compiling that one piece itself. The environment it compiled in has never successfully compiled anything: on the run where this was found, ninety-six assemblies were accepted as prepared, two were declined, and both of the rebuilds that followed failed. The same count on every other repository using this gate: everything accepted, nothing rebuilt, ever.

What the failure looked like was hundreds of "could not be found" errors naming types the piece legitimately depends on, which reads like the content is broken. What it actually meant was one sentence printed two lines earlier: the prepared assembly was declined, and was judged compatible — safe to use — a moment before being thrown away for a rebuild that could not succeed. The publication then never completed, so roughly seven hundred commits of that repository reached no running portal for two days.

The gate now rebuilds nothing. A prepared assembly that is still compatible is used, which is what the gate exists to check. One that genuinely cannot be used — a real incompatibility, or a missing build — is reported as exactly that, in one line, at the highest severity, instead of as a wall of compiler errors about something else. Nothing that passed before starts failing: acceptance was already the path every repository's gate took.