feat(server): add imageQuality/imageCount and output_format to API contract

Extend CreateProject/Generate/AgentChat request types with imageQuality
and imageCount, and add output_format to GeneratorTaskInputArgs for
layered (PSD) generation output.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

server/API.md

@@ -167,6 +167,8 @@ Returns all project summaries, sorted by `updatedAt` descending.
   "prompt": "为 still day 品牌服装独立站设计官网视觉方向",
   "imageModel": "gpt-image-2",
   "imageSize": "1024x1024",
+  "imageQuality": "auto",
+  "imageCount": 1,
   "enableWebSearch": true
 }
 ```
@@ -217,6 +219,8 @@ Deletes the project record. PostgreSQL cascades canvas nodes, chat messages, his
   "enableWebSearch": false,
   "imageModel": "gpt-image-2",
   "imageSize": "1024x1024",
+  "imageQuality": "auto",
+  "imageCount": 1,
   "mode": "design"
 }
 ```
@@ -253,7 +257,7 @@ The generation path uses the configured design agent runner. The default is the
 }
 ```
 
-`layer-separation` returns a task id that can be polled with `GET /api/generator/tasks?task_id=<id>&project_id=<project-id>`. Completed tasks return a clean background image, foreground image artifacts with `bbox`, and a `text_render_data` artifact for editable text reconstruction.
+`layer-separation` returns a task id that can be polled with `GET /api/generator/tasks?task_id=<id>&project_id=<project-id>`. When layered output is available, the service asks GPT Image 2 for a PSD, converts PSD layers into canvas image/text nodes, and falls back to the local separation path if PSD generation or decoding fails. Completed tasks return a clean background image, foreground image artifacts with `bbox`, and a `text_render_data` artifact for editable text reconstruction.
 
 The canvas `edit-elements` node action uses the same separation path but writes the separated result beside the original image. The original image remains unchanged; the clean background is a bottom image layer, and a transparent frame above it holds the separated foreground images and editable text layers.