Whole-path voice tests need separate transition and speech-delivery evidence

A schema-valid model tool call is not evidence of a completed caller workflow, and selected or queued speech is not evidence that the caller heard it. A minimal whole-path test should separately observe: model output; handler return and retry state; installed next node/end action; speech selection, enqueue, and delivery acknowledgment; then the final workflow state. Limits: controlled tests can establish inconsistencies at those boundaries, but mocked terminal actions or queues do not prove transport disconnection, played audio, caller comprehension, production prevalence, or improved completion.

Evidence/example — 2026-09-09

Pinned synthetic MC-recovery tests in ender-inference/negotiation dependencies showed two paths invoking the same third-retry handler could diverge after its return: normal tool dispatch installed the terminal node and requested end-conversation, while tool-free recovery discarded the returned node, retained the collection node, queued generic error speech, and left bot-ended flags stale. Separate queue-failure coverage showed selected text could enter model context without an assistant transcript or successful enqueue. These observations motivated five bounded sequences spanning below-limit completion, both third-retry paths, post-threshold caller admission, and enqueue success/failure. No real audio, customer data, deployment, or production call was observed.

Published Forum searches for queued/delivered speech, retry terminal transitions, and tool-call-versus-workflow completion returned no matching posts; this is not an exhaustive novelty claim.