fix(goal): pause blocked autonomous continuation

[paralin]

Active goals automatically continue after every normal assistant turn until the model marks them complete. When progress depends entirely on an external actor or event, the model currently has no way to yield, so it can spend repeated turns reporting the same blocker.

Expose goal.pause(reason) and goal.resume() through the existing goal host bridge and bundled Python skill. Pausing reuses the persisted paused state and clears queued goal contexts, while resuming reactivates the same goal after new input arrives.

Teach the continuation prompt and skill guidance to pause instead of emitting unchanged holding updates. Active goals that are not paused retain the existing continue-until-complete behavior, and the change adds no new daemon protocol status.

Note

Pause blocked autonomous goal continuation via goal.pause() and goal.resume()

• Adds goal.pause(reason) and goal.resume() Python skill functions in goal/init.py that send requests over the kernel host bridge.
• Adds _pauseGoalFromHost and _resumeGoalFromHost methods to AgentSession in agent-session.ts; pausing is only allowed when the goal is active, resuming only when paused.
• Updates goal context instructions in goals.ts to tell the model to call goal.pause("waiting for …") when blocked on external input and goal.resume() after new input arrives.
• Pause reasons are validated: trimmed, must be non-empty, and capped at 1000 characters (MAX_THREAD_GOAL_PAUSE_REASON_CHARS).
• Behavioral Change: a paused goal no longer enqueues autonomous continuation turns; a subsequent goal.resume() (without external queuing) also does not automatically enqueue a continuation.

^{Macroscope summarized d170f8b.}