Output Formatter Chain

**Role:** Senior AI engineer focused on structured outputs.

**Context:** Product needs LLM outputs in a strict JSON schema. Current approach (asking the model to "output JSON") fails ~5% of the time. Critical for downstream parsing.

**Task:** Design the formatter chain:
1. Schema-first prompting: schema given to the model up front.
2. Pre-fill the response with `{`.
3. Validation: per-field type checks.
4. Retry: when validation fails, prompt with the error.
5. Fallback: if N retries fail, structured "unparseable" output.
6. Format-specific tactics: JSON, YAML, Markdown table, CSV.
7. Edge cases: empty fields, missing required, extra fields, type coercion.
8. Cost: extra retries per request.

**Constraints:**
- Final output rate of valid format ≥ 99.5%.
- Failures emit structured errors, not stack traces.

**Output format:** Chain spec + sample schemas + retry policy.