The pre-push check now compiles NodeType source the way the portal does

A NodeType's C# lives in the mesh, not in a project, and the portal compiles it when it loads the node. Several files can belong to one NodeType — the type's own Source, its Test folder, a snippet shared from another module — and the portal joins them into a single piece of code before handing it to the compiler, lifting every using line to the top.

The check that runs before that source reaches a portal did not. It handed the compiler each file separately, which is a perfectly valid way to compile them and not the one that ships.

Why that gap was invisible rather than merely different

Several things in C# apply to a file, from where they are written to the end of it. The important one here is the line that switches on null-safety checking. In the joined form, a file that turns it on turns it on for every file that follows — including files whose author never asked for it, and including generated code. Compiled one file at a time, it reaches nothing past its own file.

So the check saw a different program from the one the portal runs, and the difference went only one way: the check was clean, and the later build was not. Worse, the build reports those in aggregate — a code and a count, no file and no line — and they were recorded against the NodeType's name. Nobody who owned that NodeType could reproduce them, and a note explaining that the debt was genuine and belonged to that content was, twice in one file, simply wrong: both times the fault was in generated code, and once it was an invisible stray character a generator had written into the middle of a file.

What changed

The check now builds one piece of code per NodeType, joined in the same order, with the using lines lifted and de-duplicated exactly as the portal does it — and the two are now compared against each other automatically, so they cannot quietly drift apart again.

Two things follow. Diagnostics the check could not see before are now reported before the push, which is the point: they were always real, and they were always going to be reported — just later, to somebody else, about the wrong thing. And every diagnostic now names the file and the line it came from, instead of a count attributed to the NodeType as a whole.

One class of finding is fixed rather than reported, because it was never the content's to fix: the byte-order mark some generators write into the middle of a file. It is invisible in an editor, it is harmless in a file compiled on its own, and in the joined form it hides the using line behind it and breaks the compile. The check now strips it on reading, the same way the portal's own file reader does.