fix: prevent UnboundLocalError crash when delegate_tasks gets an unknown agent_id

[Captain-Tripps]

Summary

• _run_single_task() returns an error message early (before children_token is ever assigned) when the caller passes an agent_id that doesn't match any available sub-agent name.
• The finally block unconditionally calls available_sub_agents_ctx.reset(children_token) on the way out, which raises UnboundLocalError on that early-return path — replacing the intended graceful "Sub-agent named 'X' does not exist" response with an unhandled exception.
• This is easy to trigger in any agent with 2+ sibling sub-agents at the same delegation level, where the orchestrator LLM has to pick a target by exact name — a slightly wrong name (typo, case mismatch, hallucinated variant) is enough to hit it. Single-child delegation chains rarely exercise this path, which is likely why it hasn't surfaced before.

Fix

Initialize children_token = None before the try block and only reset it in finally if it was actually set.

Test plan

• Confirmed the affected code path (skills/local-agent-builder/examples/basic-tui-agent/src/engine/orchestrator.py) now returns the graceful error message instead of raising on an unknown agent_id.