Don't answer @ai questions that get a badword warning

[raf-li]

Problem

@ai anjing in a channel got a badword warning by PM, while the AI still answered the question in the channel.

Change

• An @ai question that is counted as a badword is not answered (nothing is posted in the channel; the usual warning PM is sent).
• With /abt ai on, a question with only ambiguous words (@ai anjing) is judged by the AI reviewer right away (no 8 s wait for follow-ups), with the earlier messages as context, and the reviewer is told the message is a question to the assistant:
  • ok → not counted, answered
  • insult (e.g. @ai budi itu monyet ya) → counted, not answered
  • reviewer fails or its queue is full → not counted, answered (same "AI failed = not counted" rule)
• Clear badwords (@ai apa arti kontol) are counted at once and not answered.
• ai_review.py: prompt rule + two worked examples for questions to the assistant; format_message(..., to_assistant=True) adds a note line.
• tools/ai_eval.py: 4 @ai question cases (42 total).
• /help ai, docs/ai.md, CHANGELOG updated.

Testing

• New scratch suite for this case (27 checks): both switches off, aichat on/ai off, ai on (ok / insult / reviewer failure / queue full), clear and mixed badwords, admins with and without BADWORDS_IGNORE_ADMINS, @aisyah, @ai in a PM, ordinary messages still waiting for follow-ups.
• All existing scratch suites pass (abuse admin, admin badwords, AI review, AI chat, connections, fixes, prompts, shared accounts, stretched, toggles, wildcards).

🤖 Generated with Claude Code