A release says where its assemblies are

Installing a plugin should not cost a compile. It only avoids one when the installation can find assemblies that were built for exactly the framework build it runs — and until now the only way to look was an index describing whatever the serving instance happened to have installed. A release itself said nothing about where its bytes were.

It does now. A Release node carries the link: for each build lane, the framework identity those assemblies were compiled against, the architecture that produced them, and where they live. Resolution becomes "read the release, follow its link" — a property of the release, not a list to poll — and the download route takes the asking installation's own lane, so it is served the build it can run rather than the one the registry happens to prefer.

Why the architecture is written down. The framework identity is the compatibility proof, and it already differs between the x64 and arm64 halves of one image — but it is an opaque hash. An arm64 installation reading an x64-baked registry could therefore only ever be told "not adoptable", which reads identically to an incompatible framework and hid the real answer: no arm64 build had been published. Naming the architecture beside the identity turns that into a sentence you can act on.

Nothing became more permissive. The identity is still matched exactly, there is still no "near enough" fallback, and a lane that cannot be proven simply compiles — which always works. What changed is that the misses are now counted and named: a bundle that arrives with fewer assemblies than the package has types says so, at both ends, instead of quietly looking complete.