chore(ai-help): detect partial off-topic responses

[caugner]

Summary

(MP-968)

Problem

With https://github.com/mdn/rumba/pull/455, we stop generating off-topic answers as soon as possible, but this means the response may be empty or or only contain the beginning of the off-topic answer prefix (I'm sorry, but I can't).

Solution

Treat completed answers as off-topic answers, if the off-topic prefix begins (sic!) with the full response.

Examples:

• "" ➡️ off-topic
• "I'm sorry" ➡️ off-topic
• "I'm sorry, but" ➡️ off-topic
• "HTML ..." ➡️ not off-topic

---

How did you test this change?

Asked off-topic questions like these:

• What is the capital of Germany? Answer in 1000 words.

Verified with #debug and using the new message metadata that the response is empty and nevertheless detected as as an off-topic answer.

[caugner]

There's something wrong:

Fixed in c63aec9b6612bb02b013ede86f1028eea298217f.