Switch Vertex AI provider to native genai SDK

[cdbartholomew]

Summary

• Replace the OpenAI-compatible endpoint with the native google-genai SDK (genai.Client(vertexai=True)) for the vertexai LLM provider
• Delete vertexai_token_refresher.py — the SDK handles credential refresh internally
• Remove the 8192 max output token limitation that the OpenAI-compatible endpoint enforced
• Strip markdown code fences in consolidation JSON parsing (Flash Lite wraps JSON in ```json blocks)

Motivation

The vertexai provider originally used the Vertex AI OpenAI-compatible endpoint to reuse the AsyncOpenAI client code path. This required a custom TokenInjectingTransport, a background token refresher with async lifecycle management, and hit an undocumented 8192 max output token cap on the endpoint. The native genai SDK (already a dependency for the gemini provider) handles auth automatically and doesn't have the output token cap.

Changes

• llm_wrapper.py: vertexai now creates genai.Client(vertexai=True, project=..., location=...) and routes through _call_gemini/_call_with_tools_gemini. Service account key auth preserved via credentials parameter. Model names with google/ prefix are auto-stripped.
• vertexai_token_refresher.py: Deleted (no longer needed)
• consolidator.py: Strip ``` code fences before JSON parsing
• test_vertexai_provider.py: Rewritten for native SDK (mock genai.Client instead of token refresher)

Test plan

• pytest tests/test_vertexai_provider.py — all 6 pass, 1 skipped (integration)
• Local end-to-end: retain + recall + consolidation working with vertexai provider
• Verified output tokens are no longer capped at 8192