Render one bounded managed voice-selector preview

An explicit Listen action for one selected TTS voice. It accepts a bounded caller sentence (defaulting to a short product sample only when no audio_environment is supplied), keeps no audio or telemetry, and returns a short in-memory WAV only. A selected outgoing audio_environment requires explicit non-blank text and makes a deterministic, headroom-safe local audition; it never substitutes the text, language, voice, or claims a receiver-side final mix. A small per-workspace fixed-window limit protects provider spend. Every language a provider advertises can be previewed. On a platform credential the voice must be one the workspace could browse and select, except an existing saved hosted agent may re-test its exact current platform-managed selection by supplying agent_id when it has disappeared from a listing; arbitrary typed IDs never create shared runtime evidence. With a selected BYOK credential the credential is validated at the provider and then used to render, so any voice id that provider accepts works, and a platform credential is never a fallback. A GPT-Live-1 voice (provider openai, model gpt-live-1) previews differently: it is a short real live session on the platform credential that speaks Skysay's own sample sentence for the requested language, rendered once per voice, language and sample and then served from a tenant-agnostic cache (the response carries text_source: platform_sample and cached). Typed text is refused with error code voice_preview_fixed_sample rather than replaced, credential_id and delivery do not apply, a platform-wide concurrency cap answers 429 when every preview slot is busy, and the preview is never metered as a call.