AI eval verifier gets structured output and retry logic

When AI builder workflows get tested against evaluation scenarios, the verifier and mock handler are critical infrastructure. Until now, verification results were parsed from free-form text — an approach prone to failure when language models returned non-standard formats. This led to cryptic "No verification result" errors that required manual debugging to resolve.

The verifier now uses structured output with a Zod schema, meaning responses are validated against a defined contract rather than extracted from unstructured text. The mock handler gained configurable retry logic (default: one retry) to recover from transient generation failures. Failure categorization was also improved — the system now traces data flow before attributing problems, correctly distinguishing builder logic errors from mock data issues.

Test scenarios were loosened to avoid brittleness. Success criteria that expected exactly 10 posts or required specific response shapes were made more flexible. The REST API pipeline prompt no longer specifies a hard count, and the GitHub-Notion sync scenario hints at minimal responses rather than expecting complete objects.

These changes live in the instance-ai package and affect evaluation infrastructure for AI builder workflows.