[magic-peach]

Each emitter tests itself with its own harness, so nothing checks that they agree. Review of the Java emitter (#10) found bugs that other emitters had already fixed or avoided, which a shared fixture set would have caught once.

• Two new cases in model/conformance: 14-scalar-definition-references (properties that reference scalar, unconstrained, enum, and chained $defs) and 15-keyword-property-names (reserved words in common languages). Their expected models are generated by the normalizer conformance test, which now lists them.
• expected/<case>/manifest-neutral.yaml for every positive case: the part of the binding manifest that comes from the model, not the language (declarations, imported marker contracts, and root bindings).
• ci/emitter_conformance.py generates every positive case with each emitter and requires the neutral part of its manifest to equal the fixture. It has adapters for the Python, TypeScript and Rust emitters, an --update mode, and unit tests that fail if a case has no fixture. A new verify-emitter-conformance job runs it for Python and TypeScript.
• model/conformance/README.md explains the fixtures, how to check an emitter, and how to add a case.

I ran the Python, TypeScript and Go emitter test suites against the two new cases (all pass), and the checker for Python and TypeScript: all 12 positive cases agree.

Two things it showed about the Go emitter, which is why it is not in the checker:

• Its manifest uses vocabulary coordinates for model references and leaves out the schema and vocabulary coordinates, so it does not match the shape the other emitters share.
• It rejects an unconstrained definition (payload: {} in case 14) with RCG1024, where the other emitters produce an unconstrained type.

I did not change the Go emitter, since that would change published manifests. The Rust emitter in #11 has an adapter here and agrees on all twelve cases when run locally. The CI job runs Python and TypeScript for now; adding Rust is one flag once #11 lands.