The content bake no longer needs a mesh

Producing a NodeType's assembly used to require standing up a mesh. The CI bake — the platform's own and every content repo's — booted an in-process mesh, imported the whole checkout, let the mesh compile every NodeType behind its hub scheduler, and then collected whatever the mesh had produced. That is "compile through mesh nodes", and it is where the minutes went: mesh startup, message routing, and one hub activation per type.

mw-compiler now carries a compile verb that does it as an ordinary build step:

mw-compiler compile <checkout-root> --output <dir> [--source-sha <sha>]

It reads the node files out of the tree, resolves each NodeType's Sources/Tests queries against that in-memory set, resolves @@ includes, compiles with the MeshWeaver.Compiler toolchain and emits DLL + PDB into the same prebuilt-assembly bundles as before. No MeshBuilder, no AddGraph(), no content import, no hub anywhere in the path.

Nothing downstream changes. Same bundle format, same framework-identity keying, same per-type dependency records. The portals' shipped-bundle seeder, the plugin bundle client and PrebuiltAssemblySeeder cannot tell which producer wrote a bundle — and must not.

The gate keeps its mesh, because it earns it. Rendering a layout area and executing a Tests area are genuine runtime behaviours, so mw-plugin-test still stands up a mesh — it now just consumes a bake instead of producing one. Bake is a build step; gate is a runtime check.

And a live instance still compiles its own. There will always be code that never went through CI, so the emergency path stays exactly as it was. That is recovery, not a build lane.

Why the boring part is the interesting part

A baker that resolves sources even slightly differently from the runtime emits assemblies that are subtly not what the mesh would have built — and nothing fails. The bundle is well-formed, the framework identity matches, every portal adopts it, and the first symptom is a page rendering empty in production.

So the source-resolution rules are not re-implemented: query expansion, the @@-include walk, the deduplication, the executable-cell filter, the join order, the generated skeleton and the emit are all the runtime's own code, and the new resolver refuses any query it cannot evaluate rather than quietly matching less. On top of that, a test bakes one content set both ways and asserts the two producers agree on the resolved source set, the dependency records, the framework identity and the emitted assemblies' type-and-member surface.

See CI Content Bake for the full split and the equivalence argument.