fix: filter non-handler packets before fan-out

[chubes4]

Summary

Filter DataPackets before fan-out to only include ai_handler_complete packets. Non-handler packets (tool_result, ai_response) are dropped before child job creation since no downstream step can use them.

Root Cause

When the AI step calls both a handler tool (upsert_event) AND a non-handler tool (daily_memory), it produces 2+ DataPackets. The engine fanned ALL packets into child jobs, but only the ai_handler_complete packet carries data the UpdateStep needs. The tool_result packet created a child job guaranteed to fail.

BEFORE:
AI produces 2 packets → fan-out → 2 child jobs
  Child A (handler result) → UpdateStep succeeds ✅
  Child B (tool_result)    → UpdateStep fails ❌ (no handler data)

AFTER:
AI produces 2 packets → filter → 1 handler packet → inline continuation
  Same job continues → UpdateStep succeeds ✅
  No dead child created

Changes

In ExecuteStepAbility::routeAfterExecution(), before the fanOut() call:

1. Filter packets to only ai_handler_complete source type
2. If no handler packets found, keep originals (safety fallback)
3. If filtering reduces to ≤1 packet, use inline continuation instead of fan-out

Impact

• Eliminates ~1,300 dead grandchild jobs/day at the source
• Reduces fan-out overhead — fewer child jobs = fewer Action Scheduler entries = less DB churn
• The UpdateStep bandaid from PR fix: silence fan-out grandchild handler-miss noise #954 is kept as a safety net but should no longer trigger