Issue 7: Validator's structured-output call on Ollama is fragile

[sanjaigridsandguides]

Problem

• validation_chain = VALIDATION_PROMPT | ollama_llm.with_structured_output(ValidationResult) uses Ollama with structured output.
• with_structured_output for Ollama relies on the model supporting JSON mode / function calling — many Ollama models silently fall back to free-form output, which then fails Pydantic validation and throws an exception.

Impact

• Random validation failures depending on which model the user happens to have pulled.
• The catch block then treats them all as "invalid" and burns the retry budget on infrastructure errors.

Required Fix

• Document the requirement that the validator model must support JSON output.
• Pin a known-good model (e.g. llama3.1:8b-instruct-q4_K_M) in llm.py.
• Or, fall back to manual JSON parsing with regex if structured-output fails, and only count "validator returned a parseable invalid verdict" as an iteration.