LLM response validation fallback

[Veri5ied]

When LLM output is invalid JSON or missing required fields, we currently throw and fail the request.
We should post a safe fallback comment explaining that the response was invalid, and increment a failure metric.

Acceptance criteria

• If parse fails, post a comment with a friendly error message (no stack traces).
• Metrics reflect a validation failure.
• Normal flow unchanged when JSON is valid.