Instrument GenAI image generation (OpenAI, Google GenAI, Vercel AI)

Image generation is not instrumented by any Sentry JS AI integration. This is a platform-wide gap (not OpenAI- or Vercel-specific) and ultimately stems from the OpenTelemetry GenAI semantic conventions not defining an image-generation operation yet. Text, agent, and embedding calls are instrumented fine.

*Investigated against* `@sentry/node` */* `@sentry/core` *10.62.0,* `ai` *7.0.4,* `@ai-sdk/openai` *4.0.2.*

## Current state

| Integration | Instrumented methods | Image gen? |
| -- | -- | -- |
| OpenAI | `responses.create`, `chat.completions.create`, `embeddings.create`, `conversations.create` | ❌ |
| Anthropic | `messages.create`, `messages.stream`, `completions.create`, `models.*`, `beta.messages.create` | ❌ (no API) |
| Google GenAI | `models.generateContent(Stream)`, `models.embedContent`, `chats.create`, `chat.sendMessage(Stream)` | ❌ |
| LangChain | chat / LLM / embeddings callbacks | ❌ |
| LangGraph | agent / graph invocation | ❌ |
| Vercel AI | `generateText`, `streamText`, `generateObject`, `streamObject`, `embed`, `embedMany`, `rerank` | ❌ |

### Why it's missing (two layers)

1. **Sentry side:** No integration lists an image method. An exhaustive grep for `generateImage` / `images.generate` / `imagen` / `dall` / `gen_ai.image` across every AI tracing module returns nothing.
2. **Provider side:** Even where Sentry relies on the SDK's own telemetry (Vercel AI), the SDK emits none for images. `ai`'s `generateImage` opens no OTel span and does not accept `experimental_telemetry` (confirmed: its compiled body has zero telemetry/tracer references, unlike `generateText`).

## What would need to happen

### Provider methods to instrument

* **OpenAI:** add `images.generate`, `images.edit`, `images.createVariation` to `OPENAI_METHOD_REGISTRY`
* **Google GenAI:** add `models.generateImages` (Imagen)
* **Vercel AI:** add `generateImage` to `INSTRUMENTED_METHODS`
* **Anthropic / LangChain / LangGraph:** N/A (no first-class image-gen primitive)

### Required SDK work beyond the method list

1. **New span op + semantics:** define `gen_ai.generate_image` and map image-specific fields (`size`, `quality`, image count `n`, model) onto `gen_ai.request.*`; handle image-based usage/billing (per-image/resolution, not just tokens) for `gen_ai.usage.*`.
2. **Vercel AI needs a self-built span:** unlike text methods, `generateImage` emits no telemetry, so Sentry's Proxy wrapper must create the span itself (as the OpenAI integration already does manually) rather than toggling `experimental_telemetry`.
3. **Dashboard support:** confirm Insights → AI Agents renders a `generate_image` op (or treat it as a trace-only span initially).
4. **Upstream alignment:** ideally land an image-generation operation in the OTel GenAI semconv so attribute names are standard rather than Sentry-specific.

Start by locating OPENAI_METHOD_REGISTRY and INSTRUMENTED_METHODS, then compare their existing AI tracing entry points with Vercel AI's generateImage telemetry gap. Review the proposed gen_ai.generate_image fields and image-based usage requirements. Done means the listed provider methods are instrumented, dashboard handling is confirmed, and upstream semantic-convention alignment is addressed.

Rédigé par le modèle d'indexation à partir du texte de l'issue.