feat(agent): support Vertex MaaS benchmark runs

[amitdeshmukh]

Summary

• add configurable json_schema and json_object response formats for OpenAI-compatible providers while preserving the generated schema as a system-prompt contract
• classify Vertex AI authentication and capacity/rate-limit errors consistently
• allow explicitly audited publication of incomplete provider usage as an unpriced, disclosed subtotal
• publish the Gemma 4 26B A4B IT Vertex AI MaaS DeepORG run in the current 2027.1 cohort

DeepORG result

• full-pass recall: 64.0%
• pass@3: 75.0%
• pass³: 27.0%
• method recall: 62.2%
• safety precision: 100.0%
• unsafe effects: 0

The run page discloses that Vertex AI Model as a Service capacity pressure affected observed latency. It also records two unmetered interrupted attempts, labels the token count as a known subtotal, and omits a cost estimate.

Validation

• go test ./agent ./agent/eval ./serv ./cmd
• Hugo Extended 0.162.1 static build
• npm run index
• npm run check
• git diff --check