Structured Output Parser — JSON from LLM

## Task
Reliably extract structured data from LLM responses with validation and retry.

## Requirements
- Language: TypeScript
- Schema validation: Zod
- LLM: Any (OpenAI, Anthropic, etc.)

## Specifications
```typescript
// Define expected output schema
const ProductReview = z.object({
  sentiment: z.enum(["positive", "negative", "neutral"]),
  score: z.number().min(1).max(10),
  pros: z.array(z.string()).min(1),
  cons: z.array(z.string()),
  summary: z.string().max(200),
});

// Parser should:
// 1. Try native JSON mode (if provider supports it)
// 2. Fallback: extract JSON from markdown code blocks
// 3. Fallback: regex extract { ... } from response
// 4. Validate with Zod schema
// 5. On validation fail: retry with error feedback to LLM
// 6. Max 3 retries before throwing
```

## Implementation Notes
1. Use provider-native JSON mode when available (OpenAI response_format, Anthropic tool_use)
2. Include the Zod schema description in the system prompt
3. On retry, include the validation error: "Previous output failed: field 'score' must be 1-10, got 15"
4. Log all attempts for debugging
5. Support nested objects and arrays
6. Handle edge cases: empty arrays, null vs undefined, number-as-string