feat(vad): set default vad in agent session

[chenghao-mou]

AgentSession now provisions a bundled silero VAD automatically. Existing call-site behaviour is preserved via an is_default marker on the VAD instance — the framework-supplied default is transparent to every code path that previously cared about "user configured a VAD."

Before

from livekit.agents import AgentSession, inference

session = AgentSession(
    stt=inference.STT("deepgram/nova-3"),
    llm=inference.LLM("openai/gpt-4.1-mini"),
    tts=inference.TTS("cartesia/sonic-3"),
    vad=inference.VAD(model="silero"),  # required boilerplate
)

…or, in examples using the prewarm pattern:

def prewarm(proc: JobProcess) -> None:
    proc.userdata["vad"] = inference.VAD(model="silero")

server.setup_fnc = prewarm

@server.rtc_session()
async def entrypoint(ctx: JobContext) -> None:
    session = AgentSession(
        # ...
        vad=ctx.proc.userdata["vad"],
    )

After

from livekit.agents import AgentSession, inference

session = AgentSession(
    stt=inference.STT("deepgram/nova-3"),
    llm=inference.LLM("openai/gpt-4.1-mini"),
    tts=inference.TTS("cartesia/sonic-3"),
    # vad= omitted — bundled silero is provided automatically
)

# Opt out explicitly if needed:
# AgentSession(..., vad=None)

# Customise by passing your own:
# AgentSession(..., vad=inference.VAD(model="silero", min_silence_duration=0.3))

What changes (and what doesn't)

Concern	Behaviour
AgentSession() with no vad=	Now gets inference.VAD(model="silero") with is_default=True
AgentSession(vad=None)	Unchanged — no VAD
AgentSession(vad=my_vad)	Unchanged — user-supplied VAD, is_default=False
RealtimeModel with server-side turn detection	Default VAD is constructed but not wired into the audio pipeline (saves CPU); server-side VAD remains canonical
turn_detection="vad" / AudioTurnDetector opt-ins	Now satisfied by the default — no more "VAD model not provided" warnings
Realtime fallback mode = "vad" when server-side turn detection is off	No change — only flips when the user explicitly supplied a VAD
Adaptive interruption detection eligibility	No change — opt-in stays opt-in; default VAD doesn't auto-enable it
STT hook speaking= payload type and _last_speaking_time source	No change for sessions that didn't configure a VAD — default VAD is treated as absent at these sites

The marker is exposed as a read-only vad.VAD.is_default property on the base class, so plugin-supplied VADs inherit False automatically.

[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 5 additional findings.

[theomonnom]

Should we create a separate field? Like one inside the session itself? self._using_default_vad = True

[chenghao-mou]

I thought about that, but then we need to pass extra flag to activity for resolution or audio recognition like in update_vad.