Upgrade SuperGoal orchestration and model routing

[fightheyyy]

What changed

• make the main Codex agent own parent-goal analysis, dispatch, integration, and final acceptance
• require child goal lifecycle evidence and explicit dispatch-ledger tracking
• route parent analysis to gpt-5.6-sol ultra when the host supports it
• spawn child agents with gpt-5.6-sol xhigh and record transparent fallbacks
• retain opt-in-only Feishu completion notifications

Why

SuperGoal should behave as an acceptance-first orchestrator instead of letting the main agent immediately implement broad requests itself. Explicit model routing also keeps the intended reasoning profile visible and reproducible.

Validation

• quick_validate.py skills/supergoal
• git diff --check
• repository skill directory matches the locally forward-tested installation