feat: Switch AI benchmark to google.genai Batch API for 50% cost reduction

[sangicook]

Context

The AI benchmark system (run_live_benchmark.py + consult_ai_gaps.py) currently uses OpenRouter as a proxy to access Gemini Flash. For overnight/batch runs, we should switch to Google's native google.genai package which offers:

• Batch API (client.batches.create()) with 50% cost reduction for non-interactive workloads
• Native structured output via response_schema parameter (replaces JSON fence stripping)
• Direct API access without proxy latency

Proposed Changes

1. Add google-genai to [project.optional-dependencies.ai] in pyproject.toml
2. Add make_google_caller() factory in consult_ai_gaps.py alongside existing make_openrouter_caller()
3. Support GOOGLE_API_KEY environment variable for authentication
4. Implement batch mode for overnight runs:
   • Collect all prompts upfront
   • Submit as a single batch via client.batches.create()
   • Poll for completion
   • Parse results
5. Keep OpenRouter as fallback when GOOGLE_API_KEY is not set
6. Update MODEL_REGISTRY to include native Gemini model IDs

Batch API Usage Pattern

from google import genai

client = genai.Client(api_key=os.environ["GOOGLE_API_KEY"])

# Submit batch
batch = client.batches.create(
    model="gemini-2.5-flash",
    requests=[
        genai.types.BatchRequest(
            custom_id=gap_key,
            request=genai.types.GenerateContentRequest(
                contents=prompt,
                config=genai.types.GenerateContentConfig(
                    response_schema=TypedActionSchema,
                ),
            ),
        )
        for gap_key, prompt in gap_prompts
    ],
)

# Poll for completion
while batch.state == "PENDING":
    time.sleep(30)
    batch = client.batches.get(name=batch.name)

# Parse results
for result in client.batches.list_results(name=batch.name):
    responses[result.custom_id] = result.response.text

Benefits

• ~50% cost reduction on batch workloads
• Native structured output (no JSON parsing errors)
• Lower latency (no proxy hop)
• Better rate limit handling (Google's native quotas)

Notes

• OpenRouter caller remains for interactive/debugging use
• backend field on BenchmarkConfig already supports documenting which mode was used