Live API: what exactly does session resumption restore for native-audio models?

Gemini API
1

Model: gemini-3.1-flash-live-preview, audio in and out, via the
@google/genai Node SDK.

The documentation says session resumption stores “text, video, audio prompts
and model outputs” so a conversation can continue seamlessly after a brief
disconnect. In practice we see the model restart the conversation with a
fresh greeting after a resumption, as if it has no history.

Our client reconnects with the last newHandle from
sessionResumptionUpdate and re-sends the conversation as text turns. What
is unclear from the documentation:

  1. When a client reconnects with a valid handle, does the server restore the session’s own audio context, or does it only restore what the client re-sends in the setup message?
  2. If the client re-sends a voice conversation as text-only turns, is the model expected to treat that as a continuation, or as a new conversation?
  3. SessionResumptionConfig.transparent is documented in the SDK types as causing the server to send last_consumed_client_message_index, but it does not appear in the public WebSocket API reference. Is it supported, and is it the recommended way to handle resumption?

Thanks for the help Google friends !

Mario