fix: AI privacy — connection policy, Gemini key, file protection

[datlechin]

Summary

Fixes 3 privacy/security issues found during AI system audit:

• H2 — Inline suggestions ignore connection AI policy: InlineSuggestionManager never checked AIConnectionPolicy. A connection set to "Never" still sent SQL to the AI provider for inline completions. Fix: added connectionPolicy property, passed through the view chain (MainEditorContentView → QueryEditorView → SQLEditorView → SQLEditorCoordinator → InlineSuggestionManager), and checked in isEnabled().

• H6 — Gemini API key exposed in URL query string: All 3 Gemini endpoints (streamGenerateContent, fetchAvailableModels, testConnection) embedded the API key as ?key= in the URL, making it visible in proxy logs, network captures, and OS-level logging. Fix: moved to x-goog-api-key HTTP header (supported by Gemini API), matching how Anthropic uses x-api-key and OpenAI uses Authorization: Bearer.

• H8 — Chat history stored as unencrypted plain-text JSON: AI conversations in ~/Library/Application Support/TablePro/ai_chats/ had no file protection beyond atomic writes. Fix: added .completeFileProtectionUntilFirstUserAuthentication to both file writes and the directory itself, providing FileVault-level encryption at rest.

Test plan

• Build succeeds with no new warnings
• Set connection AI policy to "Never" → inline suggestions should not appear in the SQL editor
• Set connection AI policy to "Always Allow" → inline suggestions work normally
• Gemini provider: test connection, fetch models, and chat all work (key now in header)
• AI chat conversations are saved with file protection attributes