feat(image): pass reference images to GPT image model for edits

Server: forward canvas/reference images to the image model so extract,
continue, and edit requests stay faithful to the source product.
- Add ImageRequestOptions.Images and route requests with inputs to the
  /v1/images/edits endpoint, via multipart upload (default) or JSON
  image_url payload, selectable through Agent.Image.InputImageTransport.
- Extract reference URLs from prompt directives, inline @image tokens,
  and chat message image/mask contents; fold them into the idempotency
  key and prepend a reference-fidelity instruction to the prompt.

Frontend:
- Shrink CanvasToast to a compact top pill.
- Paste an internal transparent clipboard node instead of re-uploading
  when it matches the clipboard image.
- Keep agent-project polling alive on stream error while streaming.

server/README.md

@@ -58,6 +58,7 @@ Agent:
   Image:
     BaseURL: http://154.21.87.3:8080
     APIKeyEnv: GPT_IMAGE_API_KEY
+    InputImageTransport: file
     TimeoutSeconds: 600
 ObjectStorage:
   Driver: minio
@@ -71,6 +72,8 @@ ObjectStorage:
   ExpiresIn: 900
 ```
 
+`Agent.Image.InputImageTransport` controls how input/reference images are sent to the image model for edit/image-to-image calls: `file` uploads image bytes with multipart form data, while `url` sends `images: [{"image_url": "..."}]` for gateways that can fetch public URLs directly.
+
 Start local Docker dependencies for debugging:
 
 ```bash