feat(privacy): 增加应用黑名单与敏感信息脱敏

[v833]

Suggested labels: enhancement

Background

Current behavior:

• AI actions send the selected text and active conversation to the configured model endpoint.
• API keys use Electron safe storage, but all applications are eligible for selection listening.

Problem:

• Password managers, banking software, remote desktop, and internal tools can expose sensitive text accidentally.
• Selected content may include phone numbers, email addresses, identity numbers, secrets, or tokens.

Needed change:

• Add application exclusions and optional sensitive-data redaction before network requests.

Goal

Prevent accidental data transmission while keeping local copy behavior available.

User Stories

• As a privacy-conscious user, I want the assistant disabled in selected applications.
• As a user handling sensitive documents, I want a redaction preview before sending.
• As an administrator, I want predictable local-only behavior for blocked applications.

Requirements

1. Application blacklist

• Match by executable name and optionally window title.
• Provide recommended exclusions without enabling unverifiable guesses automatically.
• Suppress AI toolbar actions in blocked applications while preserving configurable local copy behavior.

2. Sensitive-data detection

• Detect configurable patterns for phone numbers, email addresses, identity numbers, API keys, and tokens.
• Show a preview and require explicit confirmation when redaction changes content.
• Allow per-pattern enablement and replacement style.

3. Privacy controls

• Keep telemetry disabled.
• Keep conversation history disabled by default.
• Provide one command to clear local settings and optional history.

Acceptance Criteria

• Blocked applications never trigger a model request.
• Blacklist matching works across executable names and configured title patterns.
• Redaction never modifies clipboard or source application content.
• Users can preview, accept, or cancel redacted requests.
• API keys and detected secrets never appear in logs or error details.
• Privacy behavior is documented in Chinese README and settings.
• Automated tests cover matching, redaction, cancellation, and no-network guarantees.

Test Plan

• Automated tests: process matching and redaction fixtures including false positives.
• Manual verification: password manager test window, remote desktop title, and selected sensitive samples.
• Regression coverage: non-blocked apps and local copy action.

Out of Scope

• Enterprise DLP integration.
• Central policy management.
• Guaranteeing detection of every sensitive value.

Dependencies / Risks

• Pattern detection has false positives and must never silently alter sent content.
• Window-title matching can be unstable and should be secondary to executable matching.

Effort Estimate

• Implementation: 5-7 days.
• Tests + docs: 3 days.
• Total: 8-10 days.