WIP: Add multi-route manual AI proxy fixture and evaluator contract

[gewenyu99]

Summary

Add an aiohttp fixture with OpenAI-compatible chat, OpenAI Responses, Ollama chat and generate, direct assistant tool calls, and a WebSocket generation path. It exercises JSON, SSE, NDJSON, provider failures, and client cancellation. The AI evaluator now asks for every route and uses an AI-specific output rubric with the criterion keys it describes.

Flow

flowchart LR
  Caller --> HTTP[HTTP proxy routes]
  Caller --> WS[WebSocket generation]
  Caller --> Assistant[Assistant tool loop]
  HTTP --> Providers[Model providers]
  WS --> Providers
  Assistant --> Tool[lookup_order]
  Assistant --> Providers
  Providers --> Evaluator[Wizard diff evaluation]

Loading

Checks

• Fixture contracts failed with 404 responses before implementation, then passed: 5 tests against fake providers.
• The evaluator contract failed on generic rubric keys before the new output template, then passed: 55 tests.
• pnpm exec tsc --noEmit passed.
• ruff format --check, ruff check, Python compilation, and git diff --check passed.

WIP

A full wizard-ci --evaluate run needs a PostHog personal API key and gateway token. Neither is configured in this isolated checkout, so the fixture has not yet been graded against a generated Wizard diff.