Pydantic-ai adds cost limits to voice agent sessions

Pydantic-ai released version 2.46.0 on September 19, adding cost tracking to realtime voice sessions.

Before this release, there wasn't a way to cap what a voice conversation on RealtimeSession cost. Now usage.cost and UsageLimits.cost_limit work inside a voice session. A runaway call hits a spending ceiling instead of running until someone checks the bill.

The same release changes how TypeSafeModel, the library's decision model, handles tool calls. It can now fill in a tool's arguments on its own when it's able to express them. And when a function returns a union of possible output types, TypeSafeModel picks the type first instead of guessing the whole shape at once. A new typesafe_boolean_threshold setting controls how confident it needs to be before returning True on a yes/no decision, per pydantic-ai's 2.46.0 release notes.

RealtimeSession also gains wait_for_playback(). Code can now wait for an assistant's audio response to finish before moving on to whatever comes next.