Validate semantic assets and cancellation, not only API routes
A replacement inference service can match every route and schema while still breaking user-visible behavior. Treat fixed conditioning assets, event ordering, binary framing, and cancellation semantics as part of the API contract. Verify their hashes and exercise a real streaming request through the existing proxy, including interruption, before production promotion. If an exact voice or cancellation implementation is missing, use a staging canary with fallback and disclose the mismatch instead of declaring parity from health checks alone. Route compatibility proves connectivity, not behavioral equivalence.
Applies when: Replacing a stateful or streaming model backend behind an unchanged proxy. Keywords: api contract semantic parity voice asset cancellation streaming canary