[CaffeinatedCoder]

Closes #113
Closes #141

Problem

Two wrong wirings were refused late, by a message that did not say what to do:

• UseNpgsql(cs).UseSchemaKit() is the kernel's method. It stamps the model and registers no contributor, so DeclarePostgres passed and the first diff was refused with "no contributor with that owner is registered", naming a package the consumer already references.
• A direct reference to CodoMetis.EFCore.SchemaKit at another version than the PostgreSQL package's breaks the package's exact pin. NuGet answers with the warning NU1608 and builds; the pair failed at run time with a MissingMethodException at whichever member differed first.

Point 4 of the issue (the consumer's model told by reference) is #98, merged. Point 2 (the name of the kernel's UseSchemaKit()) was a decision about the public surface (#141): the maintainer decided on 2026-11-14 that the name stays, and the second commit records it in the plan and the guide.

Change

• A capability beside the stamp. AddSchemaContributor also adds a SchemaContributorRegistration (owner and type), a plain value, since no contributor may be resolved while a model is built. SchemaKitWiringConvention stamps the model and puts the registered owners into a ConditionalWeakTable keyed by the model. Declare and Redeclare require the declaration's owner there and name the declaration's assembly as the package to register. It is not an annotation: that would be written into the snapshot and change with the packages a context registers.
• It is the kernel's rule, for every provider and every contributor package, not PostgreSQL's.
• SchemaKitPackages.RequireKernel(Assembly), part of the contract (SK9001): compares the informational versions without the commit. The PostgreSQL package calls it first in UseSchemaKit() and AddEntityFrameworkSchemaKitNpgsql(), from a method that calls nothing else of the kernel; the rest moved into a NoInlining method, since the runtime binds a method's calls when it compiles the method.
• A snapshot that holds a declaration without a contributor is still refused by the planner at diff.
• A contributor put into the collection without AddSchemaContributor can no longer be declared for; the method's documentation says so.
• The guide's table of missed steps, the changelog, the plan's section 3.12 and CLAUDE.md say what is now true.

Tests

• A_declaration_whose_contributor_the_options_do_not_register_is_refused_where_it_is_written on Npgsql, SQL Server and SQLite, and Only_the_owners_the_options_register_may_be_declared with its allowed twin; The_kernels_call_alone_is_refused_where_a_PostgreSQL_object_is_declared in the PostgreSQL suite. A_declaration_with_no_contributor_is_refused_on_either_side now diffs another context's model, which is what a snapshot is.
• PackageVersionTests: the same version with and without a commit passes, another is refused with both names and versions, the assembly version alone does not pass.
• Consumer smoke test: the two kernel rows expect the new refusal, and a new consumer Mixed/ references a kernel packed as <version>.mixed beside the package. The row requires NU1608 in the build's log, so that it tests a pair NuGet really lets through.
• Shown failing: with the owner check off, the five tests above fail and the smoke test stops at "the kernel's UseSchemaKit() alone: migrations add". With RequireKernel taken out of the package's UseSchemaKit(), the mixed consumer prints REGISTERED and the smoke test stops at that row. SMOKE_BREAK=kernel-only still fails the run.
• Locally on PostgreSQL 18.4: kernel 658/658, PostgreSQL package 238/238 with the integration tests, the smoke test green in 48 s.

🤖 Generated with Claude Code