Tighten group chat engagement classifier prompt

The shouldEngageInGroup classifier prompt (lines 1600-1619 in Omni/Agent/Telegram.hs) is too permissive. It currently says engage if "The message is a direct question Ava could answer" but this gets triggered by statements that aren't actually questions.

Example: Kate sent "We are thinking about a little Florida vacation in mid April... Anna Maria island, sanibel, or siesta key are all recommended" - this is a statement sharing plans, not a question, but the classifier said yes because it could be interpreted as an implicit request for advice.

Fix: Tighten the classifier prompt: 1. Change "direct question Ava could answer" to require an actual question mark or explicit request 2. Add to the "Say 'no' if" section:

• "The message is a statement sharing plans or information, not asking for input"
• "The message is addressed to other humans in the group (discussing plans together)"

3. Consider requiring more explicit signals of wanting Ava's input

Location: shouldEngageInGroup function around line 1589 in Omni/Agent/Telegram.hs