Phase 4: Google ADK Integration

[rmax]

Description

Integrate Google ADK 2.x as the model and workflow layer. Implement the Gemini adapter with structured output, credential validation, context builder for fresh-session reconstruction, and token/usage tracking.

Key Deliverables

1. Google ADK Adapter — CodingAgent protocol implementation using ADK
2. Credential Validation — validates API key or Vertex AI credentials at startup
3. Structured Agent Proposal — Pydantic models for typed model output
4. Response Validation — retry with validation feedback on invalid output
5. Context Builder — reconstructs fresh context per iteration
6. Repository Map — file map for context construction
7. Token/Usage Tracking — exact usage from Gemini API
8. ADK Workflow Nodes — PrepareContext, InvokeCodingAgent, ValidateProposal

Acceptance Criteria

• GoogleAdkCodingAgent correctly implements CodingAgent protocol
• Structured output returns valid FileEdit and CommandRequest proposals
• Invalid structured output triggers retry with validation feedback
• Credential validation fails clearly when model is unavailable
• Model precedence: CLI > goal contract > env > default
• Context builder creates compact context with repository map
• Token usage is tracked from Gemini API response
• ADK workflow nodes are independent of domain logic
• Live model smoke test (marked live_model) works with real credentials
• Ruff, mypy, and unit tests pass

Depends on: #3, #4