fix(poll-loop): apply pre-task scripts to follow-up injections too

[robbyczgw-cla]

Summary

Pre-task scripts were applied to the initial message batch but not to messages arriving as follow-ups during an active query. A scheduled task with a script gate that returned wakeAgent: false would therefore wake the agent every time it landed mid-stream — silently bypassing the gate.

Reproduction

1. Schedule a task with a script that returns { wakeAgent: false } and an interval that fires while the agent is mid-turn (e.g. */10 * * * * monitoring cron).
2. Send a chat message that triggers a multi-step turn (anything taking >10s).
3. Observe: the cron task wakes the agent every 10 min anyway. [task-script] log lines never appear for those follow-up firings.

Expected: the script gate suppresses the tick exactly as it does for the initial batch.

Fix

container/agent-runner/src/poll-loop.ts — the setInterval follow-up poller now runs applyPreTaskScripts on the new-message batch, mirroring the initial-batch path at line ~141. Behavior:

• wakeAgent: false (or script error) → task row marked completed and dropped from the follow-up push
• wakeAgent: true → scriptOutput enrichment carried through into the prompt, same as the initial path
• A pollInFlight guard serializes async runs so an interval tick won't overlap with a still-running script execution

The work is wrapped in a MODULE-HOOK:scheduling-pre-task-followup marker block to match the convention used by the initial-batch hook.

Why now

The follow-up branch existed before scheduling/script-gating was added. When applyPreTaskScripts was introduced for the initial batch it was never extended to the concurrent-poll path — the bug has been latent since the script-gate feature landed.

Test plan

• TypeScript compile (pnpm exec tsc -p container/agent-runner/tsconfig.json --noEmit) clean
• Reproduced on a real install with a */10 monitoring cron + active query → cron fired every 10 min before fix; gated correctly after fix (no [task-script] log lines = drop, with output = inject)
• Reviewer: confirm pollInFlight guard semantics — interval-fires-while-script-runs is the only realistic overlap path, and the guard makes that a no-op rather than queueing.

Risk

Low. The follow-up branch already issued an async query.push; the only behavior change is gating that push behind the same script-result rules the initial batch already uses. If the scheduling module is absent (no task-script.js), the dynamic import throws, the try/finally clears pollInFlight, and the next tick retries — same failure-recovery posture as the rest of the poller.

[gavrielc]

@robbyczgw-cla Thank you for the contribution!!