Text-only model image rejection can poison managed-agent channel queue

[LordMelkor]

🤖 A managed agent using a text-only endpoint can get stuck when its conversation history contains an image returned by a tool.

Reproduction

1. Configure buzz-agent with a text-only OpenAI-compatible model endpoint.
2. Run a tool that returns ToolResultContent::Image so the image is retained in the session history.
3. Send the next model request.
4. The endpoint rejects the request with HTTP 400, for example: Image input is not supported for this endpoint.

Current behavior

• Provider serializers replay the tool-result image as binary/base64 input.
• The failed turn remains associated with the ACP session.
• buzz-acp treats the application error as retryable, requeues the batch at the front of the channel queue, and applies channel-wide backoff.
• Later text-only events queue behind the deterministic failure and can share its retry/dead-letter fate.

Expected behavior

• Preserve the image and accompanying text/metadata in session history.
• When an endpoint proves it cannot accept images, retry the request once with a textual image placeholder instead of binary image data.
• Remember that negative capability for the exact effective model for the process lifetime, without persisting it or inferring it from model names.
• Do not let the deterministic unsupported-image error consume the queue retry budget or block later text events.

Proposed scope

• Add a narrow provider-error classifier for the observed HTTP 400 signature.
• Degrade images only while serializing requests for a model that has returned that signature.
• Keep multimodal behavior unchanged for unknown and image-capable models.
• Add an ACP-side non-retryable fallback with session invalidation and a user-visible notice for defense in depth.

Related work: #3300, #3223, #4477, and #4127.