Task: 支持 AI Design 图生图选图交互

Identity

Generalize the existing video-first-frame picker into a single-image source picker that supports both image reference generation and video first frames.
Reuse existing workspace generated-image selection, local upload and attachmentAssetIds transport without adding a parallel upload path.
Send medium-correct user intent text to the Design Agent and cover both interactions with focused page tests.

Image-to-image accepts exactly one image from the current workspace or one local upload; task and asset ownership remain Workspace-scoped.
Do not alter Workspace / Conversation / Task ownership or event routing.
Preserve the currently deployed string quick-reply trigger for compatibility, while removing video-only labels and submission semantics from the picker.
A structured picker action is preferable long-term, but is outside this server-compatible increment and is recorded as a follow-up.

The former video-only first-frame dialog is now a reusable single-image source picker. For an image Brief it presents “选择图生图参考图” and sends one selected Workspace result Asset or one uploaded Asset through the existing attachmentAssetIds Turn transport; for a video Brief it retains the original first-frame wording and behavior.
The picker closes as soon as a completed work is selected and after a local upload succeeds, while request failures continue to surface through the existing design-conversation error path.
Only an explicit brief.medium === 'image' selects image-reference semantics; video, null and missing medium values default to the historical video behavior for compatibility with delayed or older server snapshots.
The deployed exact quick reply “从作品列表选择图片” remains supported. README now documents text-to-image, single-reference image-to-image, project-work selection and local upload; no parallel upload API or client state model was introduced.

pnpm exec vitest run tests/unit/image-canvas-page.test.tsx: 25 passed.
pnpm run typecheck: passed.
pnpm exec eslint src/pages/ImageCanvas/index.tsx tests/unit/image-canvas-page.test.tsx: passed.
pnpm run build:vite: passed; only the repository's existing dynamic-import and chunk-size warnings were emitted.
git diff --check: passed.
Independent read-only Sol review: PASS, with no P0-P3 findings after inspecting both the client and matching server diffs; the reviewer reran this 25-test page suite.
No packaged desktop smoke or live Works Square provider request was performed.

Evolve DesignMessage.quickReplies from display-only strings to structured entries such as select_image_asset plus image_reference | video_first_frame. The current release preserves the deployed exact Chinese string for compatibility, but localization or wording changes remain coupled to behavior until the protocol evolves.
Add a first-class “remove reference image” control if the server introduces an explicit clear-source action; omission of an attachment must not implicitly clear a Conversation's frozen reference.

Target canonical documents: 20-architecture/system-overview.md, 20-architecture/module-map.md, 20-architecture/data-flow.md, and 40-domain/business-rules.md. Proposal: record the generalized single-image source picker, its single-Asset attachment boundary, medium-specific intent, and null-medium video compatibility fallback. Evidence: the 25 focused page tests plus typecheck, scoped ESLint and production build listed above. Future impact: future picker redesigns must preserve Workspace ownership and attach one real Asset ID rather than relying on a visual description. No semantic conflict; canonical promotion belongs to a serialized Integration task.
Target canonical ADR/protocol documentation: the accepted AI Design Conversation/Gateway contract. Proposal: replace localized quick-reply string matching with a structured picker action before adding more asset-input purposes. Evidence: one display string now intentionally opens both image and video source flows, with purpose inferred from current Brief medium. Future impact: removes localization and delayed-state coupling. This is a future protocol change and requires coordinated server/client design confirmation.