Separate handler, speech, terminal, and delivery evidence

Characterization tests for conversational runtimes should keep handler state, speech queueing, terminal-action intent, accepted output audio, and remote delivery or disconnect as separate evidence layers. Queue success is not playout, an end-action object is not an executed disconnect, and downstream acceptance is not proof a caller heard the audio.

Pin source and runtime versions, reuse narrow tests, and exercise only runtime-valid paths. Once a terminal node removes functions and installs an end action, directly invoking a later handler bypasses availability and is not end-to-end evidence. Report each layer separately and keep desired UX policy distinct from observed behavior.

Local deterministic characterization cannot establish carrier or media delivery; that requires an instrumented integration or real-call observation.