fix(soniox): emit PREFLIGHT_TRANSCRIPT for preemptive LLM generation

[octo-patch]

Fixes #5536

Problem

The Soniox STT plugin only emits INTERIM_TRANSCRIPT events, so AgentSession's preemptive LLM generation never fires when using Soniox for STT-driven end-of-turn detection. Users pay full LLM TTFT in wall-clock time after every turn, which is too high for real-time voice use cases.

Solution

When a Soniox token batch contains only final tokens (no non-final/speculative tokens pending), speech has reached a stable state — this indicates a likely pause, analogous to Deepgram v2's EagerEndOfTurn signal. In that case, emit PREFLIGHT_TRANSCRIPT instead of INTERIM_TRANSCRIPT.

This allows AgentSession to start speculative LLM generation. If the user resumes speaking, the next batch will contain non-final tokens, triggering a regular INTERIM_TRANSCRIPT that aborts the preemptive generation — the same abort path used by Deepgram v2's TurnResumed event.

The change is minimal: one conditional in the event-type selection within _recv_messages_task.

Testing

Tested by code review against the Deepgram v2 pattern (EagerEndOfTurn → PREFLIGHT_TRANSCRIPT). The logic mirrors that behavior using the token finality heuristic native to Soniox's streaming API.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

octo-patch seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 2 additional findings.