End-to-end validation of v2 extract/install across multiple AI tools

[troylar]

Description

Validate the v2 extract → install pipeline end-to-end with real projects and real AI tools. Create integration tests and a manual test matrix that exercises extraction from real-world project configs and installation into multiple tool formats.

Context

v2 (0.12.0) shipped with AI-powered extraction and installation, but testing has been primarily unit-level with mocked LLM responses. We need to verify:

1. Extraction produces valid practice declarations from real project configs
2. Installation correctly translates practices to each tool's native format
3. AI adaptation produces sensible merges with existing rules
4. File-copy fallback (--no-ai) works correctly for all tools
5. Round-trip fidelity: extract from tool A, install to tool B, content is meaningful

Acceptance Criteria

• Integration test suite that runs extract → install for at least 3 tool pairs (e.g., Claude→Cursor, Cursor→Kiro, Claude→Copilot)
• Test with real project configs (use devsync's own .claude/rules/ as source)
• Verify file placement, extensions, and content format for each target tool
• Test --no-ai mode produces valid output for all tools
• Test v1 backward compatibility (ai-config-kit-package.yaml packages still install)
• Document a manual test matrix for pre-release validation
• Tests pass: invoke test-unit && invoke test-integration

Test Matrix

Source	Target	Mode	Status
Claude Code	Cursor	AI
Claude Code	Kiro	AI
Claude Code	Copilot	AI
Claude Code	Cursor	--no-ai
Cursor	Claude Code	AI
v1 package	Claude Code	file-copy
v1 package	Cursor	file-copy