feat(selection): 增加剪贴板回退和截图 OCR

[v833]

Suggested labels: enhancement

Background

Current behavior:

• The app depends on the Windows text selection service exposed through selection-hook.

Problem:

• Some PDFs, canvases, remote desktops, images, video subtitles, and custom controls do not expose selected text.
• Users receive no guided fallback when the native selection API cannot return content.

Needed change:

• Add a safe clipboard fallback and an optional screenshot OCR workflow.

Goal

Extend useful selection coverage without silently reading or replacing unrelated clipboard content.

User Stories

• As a user in an unsupported application, I want a clear fallback using copied text.
• As a user viewing an image or scanned PDF, I want to select a screen region and run OCR locally.
• As a privacy-conscious user, I want recognized text reviewed before it is sent to a model.

Requirements

1. Clipboard fallback

• Detect native selection failure and present a Copy then retry flow.
• Preserve and restore the previous clipboard when automatic capture is enabled.
• Never send clipboard content until the user chooses an AI action.

2. Screenshot OCR

• Add an explicit region-capture command with clear cancellation.
• Prefer a proven local OCR engine and support Chinese and English initially.
• Show recognized text for correction before invoking actions.

3. Source awareness

• Label results as native selection, clipboard, or OCR source.
• Keep OCR disabled in blacklisted applications.

Acceptance Criteria

• Unsupported native selection produces a clear fallback instead of failing silently.
• Clipboard fallback preserves previous clipboard contents when configured.
• OCR region selection supports cancel and never captures outside the chosen region.
• Recognized text can be reviewed and edited before sending.
• OCR runs locally by default and no image is uploaded implicitly.
• Result source is visible to the user.
• Tests cover clipboard restoration, cancellation, OCR errors, and privacy exclusions.

Test Plan

• Automated tests: clipboard state helper and OCR workflow state machine.
• Manual verification: browser canvas, scanned PDF, image, video subtitle, and remote desktop samples.
• Regression coverage: supported native text selection remains the fastest path.

Out of Scope

• Full-screen continuous OCR.
• Cloud OCR without explicit opt-in.
• Document layout reconstruction.

Dependencies / Risks

• OCR increases package size and should be evaluated against an optional-download design.
• Clipboard restoration can race with other applications and requires conservative timing.
• Region capture must respect multi-monitor DPI and application blacklist rules.

Effort Estimate

• Implementation: 8-12 days.
• Tests + docs: 3-4 days.
• Total: 11-16 days.