Metadata tagging & validation silently mislabel all newly-processed questions (gpt-5-mini reasoning starves the token budget)

[drozado]

• After the “rewrite LLM forecaster” refactor (6566923, 2026-06-26), metadata jobs call gpt-5-mini with low max_output_tokens and no reasoning.effort. Since the model defaults to medium reasoning, every request exhausts the budget and fails with status=incomplete, reason=max_output_tokens. After retries, tag_questions / validate_questions fall back to category="Other" and valid_question=True, silently mislabeling new questions.

• A Kalshi tagging run made 120 requests; 0 succeeded, then hit the 30-minute timeout.

• Questions processed through this path since 2026-06-26 may be contaminated.

• Fix: Set reasoning={"effort": "minimal"} and raise the tagging/validation output budgets.