[FEATURE]: Vertex AI route for the Google Interactions protocol #54162

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

The Google Interactions protocol from #52981 can only reach the Gemini API. Could you add a Vertex AI route for it, in
the same way the google-vertex provider already has Gemini, Messages and Responses routes? Teams that use Gemini only
through Vertex AI cannot use the new protocol today.

Current state on v2 (5cfe3b4)

• packages/ai/src/protocols/google-interactions.ts has one base URL:
  https://generativelanguage.googleapis.com/v1beta.
• packages/ai/src/providers/google/interactions.ts accepts apiKey and baseURL, and providers/google.ts sends the
  key as x-goog-api-key. There is no project, location or OAuth option, so a service account or Application
  Default Credentials cannot be used. A static Authorization header is not practical because the token expires.
• packages/ai/src/providers/google-vertex.ts already builds the project and location URL and the OAuth token for the
  Gemini route, so the pieces exist.

Vertex endpoint (Preview; the documented project endpoint uses locations/global):
POST https://aiplatform.googleapis.com/v1beta1/projects/{project}/locations/global/interactions, with ADC or an
express-mode API key.
Developer guide
and API reference.

What we checked on Vertex (2026-10-09, gemini-3.8-flash, service account through ADC, called directly):

• Non-streaming and ?alt=sse streaming both return 200. The stream sends interaction.created,
  interaction.status_update, step.start, step.delta, step.stop, interaction.completed and done.
• service_tier: "flex" in the body and the x-vertex-ai-llm-shared-request-type: flex header (which the Vertex Gemini
  route sends) are both accepted. The response does not say which tier ran, so it is worth confirming which one Vertex
  uses.

Not requested: automatic continuation. On the same endpoint we compared full-history replay with
previous_interaction_id, over 5 turns with a 12.7K-token system instruction:

Continuation billed the same input tokens and got no implicit cache hits in this test. The system instruction must
also be sent on every turn: when we sent it on turn 1 only, later answers were wrong. So the current store: false
default looks right for Vertex, and we are not asking for automatic continuation.