fix(livekit-agents): forward prewarm to the primary LLM in the fallback adapter

[fxhxdxd]

Fixes #6572.

Problem

LLM.prewarm() short-circuits unless the concrete class overrides _prewarm_impl (livekit-agents/livekit/agents/llm/llm.py:162-163). llm.FallbackAdapter doesn't override it, so every prewarm() call it receives — from AgentSession.__init__ (voice/agent_session.py:490) and from AgentActivity (voice/agent_activity.py:636,721) — returns immediately without ever reaching the wrapped instances. Providers that do implement prewarming are therefore never warmed when used behind the adapter, so the first reply pays the full DNS + TLS setup cost.

tts.FallbackAdapter already handles this correctly (tts/fallback_adapter.py:126-128); the LLM adapter is the gap.

stt.FallbackAdapter has the same gap — happy to cover it in this PR or a follow-up, whichever you prefer.

Fix

Override prewarm() to delegate to the first instance, mirroring the TTS adapter. Only the primary is prewarmed: the remaining instances are not expected to serve traffic unless it fails, and warming them would open connections to providers that may never be used.

The loop argument is forwarded, which is the one place this differs from the TTS version — LLM.prewarm() accepts loop (the TTS and STT variants don't) and AgentSession passes the session loop explicitly, so dropping it would schedule the prewarm task on the wrong loop.

Tests

New tests/test_llm_fallback.py:

• test_prewarm_forwarded_to_primary_llm — the primary's _prewarm_impl runs and the fallback instance stays cold.
• test_prewarm_forwards_event_loop — the prewarm task is scheduled on the loop supplied by the caller.

Both fail on main and pass with this change. ruff check, ruff format --check and mypy --strict are clean on the touched files, and test_llm_fallback.py, test_tts_fallback.py and test_stt_fallback.py pass together (15 passed).

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

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

[chenghao-mou]

lgtm. One small test nit.

[chenghao-mou]

nit: even if we drop loop=loop, it will pass as it defaults to the running loop. We can test it with a new loop instead:

class RecordingLLM(FakeLLM):
    def __init__(self) -> None:
        super().__init__()
        self.prewarm_loop: asyncio.AbstractEventLoop | None = None

    def prewarm(self, *, loop: asyncio.AbstractEventLoop | None = None) -> None:
        self.prewarm_loop = loop

...
supplied_loop = asyncio.new_event_loop()
try:
    fallback_adapter.prewarm(loop=supplied_loop)
    assert primary.prewarm_loop is supplied_loop
finally:
    supplied_loop.close()

[fxhxdxd]

Good catch, you're right that the original assertion held either way, since the wrapped LLM defaults to the running loop when loop is omitted.

Switched to recording the loop the primary is asked to prewarm on and supplying one distinct from the running loop, per your suggestion. Verified it now fails if loop=loop is dropped from the adapter, which the previous version did not catch. Pushed in aeb0b79.