Send a prompt to the vendor-specific AI assistant. The response is grounded in the
authenticated vendor's uploaded knowledge base and persona configuration. A session ID
can be supplied to continue an existing conversation; otherwise a new session is created.

Response semantics:

• ai_response.products: Always an empty array. The assistant no longer recommends catalogue
  products — a customer asking what to use or buy is steered into the skin scan instead, since
  the scan is what personalised recommendations are built from. The key is retained so existing
  clients continue to parse the response unchanged.
• ai_response.actions: List of non-visual action markers emitted by the assistant to drive UI behavior.
  • "SKIN_SCAN": show a call-to-action for the user to take a skin scan. This is the primary
    response to a product question.
  • "BOOK_CONSULTATION": surface an option to connect the user with a human consultant or
    aesthetician.
    Markers are not part of the visible assistant text; they are provided only for clients to
    render contextual UI affordances.
• persona: Snapshot of the vendor's active persona that shaped the reply. Includes tone,
  communication_style, brand_character, and optional custom_emphasis. Clients can surface this
  metadata (e.g., in a conversation header) or log it for analytics.