fix(task): rewrite main-session gate prompt to prevent self-answering

[yanyihan-xiaomi]

Summary

• Rewrite task gate's main-session reentry prompt from "complete the work then continue or respond" to a structured decision tree that prevents the model from self-answering
• Offer question tool as the primary exit path when user input is needed
• Add explicit negative constraints: "Do NOT answer your own questions or assume user intent"

Problem

When the model's turn is classified as "final" with open/in_progress tasks, the task gate injects a synthetic user message. The old prompt said:

For EACH: complete the work then `task done <id> <summary>`, or `task abandon <id> <reason>` if it is genuinely not needed.
Then continue or respond.

This caused the model to, after asking the user a question:

1. Receive the gate nudge saying "complete the work"
2. Answer its own question ("I'll just fix the core logic")
3. Start implementing, ignoring that the user never responded

Solution

Replace the main-session gate prompt with a structured decision tree:

1. Need user input → use question tool (supports choices, free-text via empty options, recommendations)
2. Work incomplete, no user input needed → continue working, then task done
3. Already complete → task done
4. Blocked externally → task block
5. No longer needed → task abandon

When the model uses the question tool, the turn is not classified as "final" (pending tool call), so the gate never re-fires.

Subagent path is unchanged.