Feature: scriptable citations / annotations on text responses #547

Problem

Responses API output_text parts are always emitted with annotations: [], and Chat Completions / Anthropic / Gemini have no citation equivalent. There is no way to script a model turn that returns sources, so clients that surface citations (e.g. the AI SDK source stream part) can't be tested end-to-end against aimock.

Proposal

An optional citations (or annotations) field on TextResponse / ContentWithToolCallsResponse, e.g.

llm.on({ endpoint: 'responses' }, {
  content: 'Paris is the capital of France.',
  citations: [{ type: 'url', url: 'https://example.com/paris', title: 'Paris', startIndex: 0, endIndex: 5 }],
});

Emitted per provider wire format:

• OpenAI Responses: url_citation (and file_citation) entries in output_text.annotations, plus response.output_text.annotation.added events when streaming
• Chat Completions: message.annotations / delta.annotations (url_citation)
• Anthropic: citations on text blocks (citations_delta when streaming)
• Gemini: groundingMetadata.groundingChunks / groundingSupports

Use case

Mastra is moving its agent-loop regression tests from hand-written mock models to aimock scenarios; source/citation propagation is one of the few behaviours we can't currently cover.